notaz.gp2x.de / pcsx_rearmed.git / blame
| Commit | Line | Data |
|---|---|---|
| 75e28f62 E |
1 | /* |
| 2 | * Copyright (C) 2011 Gilead Kutnick "Exophase" <exophase@gmail.com> | |
| 3 | * | |
| 4 | * This program is free software; you can redistribute it and/or | |
| 5 | * modify it under the terms of the GNU General Public License as | |
| 6 | * published by the Free Software Foundation; either version 2 of | |
| 7 | * the License, or (at your option) any later version. | |
| 8 | * | |
| 9 | * This program is distributed in the hope that it will be useful, | |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
| 12 | * General Public License for more details. | |
| 13 | */ | |
| 14 | ||
| 15 | #define MAX_SPANS 512 | |
| 16 | #define MAX_BLOCKS 64 | |
| 17 | #define MAX_BLOCKS_PER_ROW 128 | |
| 18 | ||
| 19 | #define psx_gpu_test_mask_offset 0 | |
| 20 | #define psx_gpu_uvrg_offset 16 | |
| 21 | #define psx_gpu_uvrg_dx_offset 32 | |
| 22 | #define psx_gpu_uvrg_dy_offset 48 | |
| 23 | #define psx_gpu_u_block_span_offset 64 | |
| 24 | #define psx_gpu_v_block_span_offset 80 | |
| 25 | #define psx_gpu_r_block_span_offset 96 | |
| 26 | #define psx_gpu_g_block_span_offset 112 | |
| 27 | #define psx_gpu_b_block_span_offset 128 | |
| 28 | ||
| 29 | #define psx_gpu_b_dx_offset 132 | |
| 30 | ||
| 31 | #define psx_gpu_b_offset 144 | |
| 32 | #define psx_gpu_b_dy_offset 148 | |
| 33 | #define psx_gpu_triangle_area_offset 152 | |
| 34 | #define psx_gpu_texture_window_settings_offset 156 | |
| 35 | #define psx_gpu_current_texture_mask_offset 160 | |
| 36 | #define psx_gpu_viewport_mask_offset 164 | |
| 37 | #define psx_gpu_dirty_textures_4bpp_mask_offset 168 | |
| 38 | #define psx_gpu_dirty_textures_8bpp_mask_offset 172 | |
| 39 | #define psx_gpu_dirty_textures_8bpp_alternate_mask_offset 176 | |
| 40 | #define psx_gpu_triangle_color_offset 180 | |
| 41 | #define psx_gpu_primitive_color_offset 184 | |
| 42 | #define psx_gpu_dither_table_offset 188 | |
| 43 | #define psx_gpu_render_block_handler_offset 204 | |
| 44 | #define psx_gpu_texture_page_ptr_offset 208 | |
| 45 | #define psx_gpu_clut_ptr_offset 212 | |
| 46 | #define psx_gpu_vram_ptr_offset 216 | |
| 47 | ||
| 48 | #define psx_gpu_render_state_base_offset 220 | |
| 49 | #define psx_gpu_render_state_offset 222 | |
| 50 | #define psx_gpu_num_spans_offset 224 | |
| 51 | #define psx_gpu_num_blocks_offset 226 | |
| 52 | #define psx_gpu_offset_x_offset 228 | |
| 53 | #define psx_gpu_offset_y_offset 230 | |
| 54 | #define psx_gpu_clut_settings_offset 232 | |
| 55 | #define psx_gpu_texture_settings_offset 234 | |
| 56 | #define psx_gpu_viewport_start_x_offset 236 | |
| 57 | #define psx_gpu_viewport_start_y_offset 238 | |
| 58 | #define psx_gpu_viewport_end_x_offset 240 | |
| 59 | #define psx_gpu_viewport_end_y_offset 242 | |
| 60 | #define psx_gpu_mask_msb_offset 244 | |
| 61 | ||
| 62 | #define psx_gpu_triangle_winding_offset 246 | |
| 63 | #define psx_gpu_display_area_draw_enable_offset 247 | |
| 64 | #define psx_gpu_current_texture_page_offset 248 | |
| 65 | #define psx_gpu_last_8bpp_texture_page_offset 249 | |
| 66 | #define psx_gpu_texture_mask_width_offset 250 | |
| 67 | #define psx_gpu_texture_mask_height_offset 251 | |
| 68 | #define psx_gpu_texture_window_x_offset 252 | |
| 69 | #define psx_gpu_texture_window_y_offset 253 | |
| 70 | #define psx_gpu_primitive_type_offset 254 | |
| 71 | ||
| 72 | #define psx_gpu_reserved_a_offset 255 | |
| 73 | ||
| 74 | #define psx_gpu_blocks_offset 0x0100 | |
| 75 | #define psx_gpu_span_uvrg_offset_offset 0x2100 | |
| 76 | #define psx_gpu_span_edge_data_offset 0x4100 | |
| 77 | #define psx_gpu_span_b_offset_offset 0x5100 | |
| 78 | ||
| 79 | #define psx_gpu__vram_offset 0x005900 | |
| 80 | ||
| 81 | #define edge_data_left_x_offset 0 | |
| 82 | #define edge_data_num_blocks_offset 2 | |
| 83 | #define edge_data_right_mask_offset 4 | |
| 84 | #define edge_data_y_offset 6 | |
| 85 | ||
| 86 | ||
| 87 | #define psx_gpu r0 | |
| 88 | #define v_a r1 | |
| 89 | #define v_b r2 | |
| 90 | #define v_c r3 | |
| 91 | ||
| 92 | #define x0 r4 | |
| 93 | #define x1 r5 | |
| 94 | #define x2 r6 | |
| 95 | #define x0_x1 r5 | |
| 96 | #define x1_x2 r6 | |
| 97 | #define y0 r7 | |
| 98 | #define y1 r8 | |
| 99 | #define y2 r9 | |
| 100 | #define y0_y1 r7 | |
| 101 | #define y1_y2 r8 | |
| 102 | #define b0 r9 | |
| 103 | #define b1 r10 | |
| 104 | #define b2 r11 | |
| 105 | #define b0_b1 r10 | |
| 106 | #define b1_b2 r11 | |
| 107 | ||
| 108 | ||
| 109 | #define area_r_s r5 | |
| 110 | ||
| 111 | #define g_bx0 r2 | |
| 112 | #define g_bx r3 | |
| 113 | #define g_bx2 r4 | |
| 114 | #define g_bx3 r5 | |
| 115 | #define b_base r6 | |
| 116 | #define g_by r8 | |
| 117 | ||
| 118 | #define gs_bx r7 | |
| 119 | #define gs_by r10 | |
| 120 | ||
| 121 | #define ga_bx g_bx | |
| 122 | #define ga_by g_by | |
| 123 | ||
| 124 | #define gw_bx_h g_bx | |
| 125 | #define gw_by_h g_by | |
| 126 | ||
| 127 | #define gw_bx_l r11 | |
| 128 | #define gw_by_l gw_bx_l | |
| 129 | ||
| 130 | #define store_a r0 | |
| 131 | #define store_b r1 | |
| 132 | #define store_inc r5 | |
| 133 | ||
| 134 | ||
| 135 | #define v0 q0 | |
| 136 | #define uvrgb0 d0 | |
| 137 | #define x0_y0 d1 | |
| 138 | ||
| 139 | #define v1 q1 | |
| 140 | #define uvrgb1 d2 | |
| 141 | #define x1_y1 d3 | |
| 142 | ||
| 143 | #define v2 q2 | |
| 144 | #define uvrgb2 d4 | |
| 145 | #define x2_y2 d5 | |
| 146 | ||
| 147 | #define x0_ab q3 | |
| 148 | #define uvrg_xxxx0 q3 | |
| 149 | #define uvrg0 d6 | |
| 150 | #define xxxx0 d7 | |
| 151 | ||
| 152 | #define x1_ab q4 | |
| 153 | #define uvrg_xxxx1 q4 | |
| 154 | #define uvrg1 d8 | |
| 155 | #define xxxx1 d9 | |
| 156 | ||
| 157 | #define x2_ab q5 | |
| 158 | #define uvrg_xxxx2 q5 | |
| 159 | #define uvrg2 d10 | |
| 160 | #define xxxx2 d11 | |
| 161 | ||
| 162 | #define y0_ab q6 | |
| 163 | #define yyyy_uvrg0 q6 | |
| 164 | #define yyyy0 d12 | |
| 165 | #define uvrg0b d13 | |
| 166 | ||
| 167 | #define y1_ab q7 | |
| 168 | #define yyyy_uvrg1 q7 | |
| 169 | #define yyyy1 d14 | |
| 170 | #define uvrg1b d15 | |
| 171 | ||
| 172 | #define y2_ab q8 | |
| 173 | #define yyyy_uvrg2 q8 | |
| 174 | #define yyyy2 d16 | |
| 175 | #define uvrg2b d17 | |
| 176 | ||
| 177 | #define d0_ab q9 | |
| 178 | #define d0_a d18 | |
| 179 | #define d0_b d19 | |
| 180 | ||
| 181 | #define d1_ab q10 | |
| 182 | #define d1_a d20 | |
| 183 | #define d1_b d21 | |
| 184 | ||
| 185 | #define d2_ab q11 | |
| 186 | #define d2_a d22 | |
| 187 | #define d2_b d23 | |
| 188 | ||
| 189 | #define d3_ab q12 | |
| 190 | #define d3_a d24 | |
| 191 | #define d3_b d25 | |
| 192 | ||
| 193 | #define ga_uvrg_x q1 | |
| 194 | #define ga_uvrg_y q4 | |
| 195 | ||
| 196 | #define dx x0_x1 | |
| 197 | #define dy y0_y1 | |
| 198 | #define db b0_b1 | |
| 199 | ||
| 200 | #define uvrg_base q11 | |
| 201 | ||
| 202 | #define gs_uvrg_x q5 | |
| 203 | #define gs_uvrg_y q6 | |
| 204 | ||
| 205 | #define g_uvrg_x q1 | |
| 206 | #define ga_uv_x d2 | |
| 207 | #define g_uv_x d2 | |
| 208 | #define ga_rg_x d3 | |
| 209 | #define g_rg_x d3 | |
| 210 | ||
| 211 | #define g_uvrg_y q4 | |
| 212 | #define ga_uv_y d8 | |
| 213 | #define g_uv_y d8 | |
| 214 | #define ga_rg_y d9 | |
| 215 | #define g_rg_y d9 | |
| 216 | ||
| 217 | #define gw_uv_x q1 | |
| 218 | #define gw_rg_x q2 | |
| 219 | #define gw_uv_y q4 | |
| 220 | #define gw_rg_y q3 | |
| 221 | ||
| 222 | #define w_mask q9 | |
| 223 | #define w_mask_l d18 | |
| 224 | ||
| 225 | #define r_shift q10 | |
| 226 | ||
| 227 | #define uvrg_dx0 q0 | |
| 228 | #define uvrg_dx0l d0 | |
| 229 | #define uvrg_dx0h d1 | |
| 230 | ||
| 231 | #define uvrg_dx1 q1 | |
| 232 | #define uvrg_dx1l d2 | |
| 233 | #define uvrg_dx1h d3 | |
| 234 | ||
| 235 | #define uvrg_dx2 q2 | |
| 236 | #define uvrg_dx2l d4 | |
| 237 | #define uvrg_dx2h d5 | |
| 238 | ||
| 239 | #define uvrg_dx3 q3 | |
| 240 | #define uvrg_dx3l d6 | |
| 241 | #define uvrg_dx3h d7 | |
| 242 | ||
| 243 | ||
| 244 | .align 4 | |
| 245 | ||
| 246 | #define function(name) \ | |
| 247 | .global name; \ | |
| 248 | name: \ | |
| 249 | ||
| 250 | @ r0: psx_gpu | |
| 251 | @ r1: v_a | |
| 252 | @ r2: v_b | |
| 253 | @ r3: v_c | |
| 254 | ||
| 255 | function(compute_all_gradients) | |
| 256 | // First compute the triangle area reciprocal and shift. The division will | |
| 257 | // happen concurrently with much of the work which follows. | |
| 258 | @ r12 = psx_gpu->triangle_area | |
| 259 | ldr r12, [ psx_gpu, #psx_gpu_triangle_area_offset ] | |
| 260 | stmdb sp!, { r4 - r11, lr } | |
| 261 | ||
| 262 | @ load exponent of 62 into upper half of double | |
| 263 | movw r4, #0 | |
| 264 | clz r14, r12 @ r14 = shift | |
| 265 | ||
| 266 | movt r4, #((62 + 1023) << 4) | |
| 267 | mov r12, r12, lsl r14 @ r12 = triangle_area_normalized | |
| 268 | ||
| 269 | @ load area normalized into lower half of double | |
| 270 | mov r5, r12, lsr #10 | |
| 271 | vmov.f64 d30, r5, r4 @ d30 = (1 << 62) + ta_n | |
| 272 | ||
| 273 | movt r4, #((1022 + 31) << 4) | |
| 274 | mov r5, r12, lsl #20 | |
| 275 | ||
| 276 | add r4, r4, r12, lsr #11 | |
| 277 | vmov.f64 d31, r5, r4 | |
| 278 | ||
| 279 | vdiv.f64 d30, d30, d31 @ d30 = ((1 << 62) + ta_n) / ta_n | |
| 280 | ||
| 281 | // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) = | |
| 282 | // ( d0 * d1 ) - ( d2 * d3 ) = | |
| 283 | // ( m0 ) - ( m1 ) = gradient | |
| 284 | ||
| 285 | // This is split to do 12 elements at a time over three sets: a, b, and c. | |
| 286 | // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so | |
| 287 | // two of the slots are unused. | |
| 288 | ||
| 289 | // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as | |
| 290 | // is g. | |
| 291 | ||
| 292 | // First type is: uvrg bxxx xxxx | |
| 293 | // Second type is: yyyy ybyy uvrg | |
| 294 | // Since x_a and y_c are the same the same variable is used for both. | |
| 295 | ||
| 296 | vld1.u32 { v0 }, [ v_a, : 128 ] @ v0 = { uvrg0, b0, x0, y0 } | |
| 297 | ldrsh x0, [ v_a, #8 ] @ load x0 | |
| 298 | ||
| 299 | vld1.u32 { v1 }, [ v_b, : 128 ] @ v1 = { uvrg1, b1, x1, y1} | |
| 300 | ldrh x1, [ v_b, #8 ] @ load x1 | |
| 301 | ||
| 302 | vld1.u32 { v2 }, [ v_c, : 128 ] @ v2 = { uvrg2, b2, x2, y2 } | |
| 303 | ldrh x2, [ v_c, #8 ] @ load x2 | |
| 304 | ||
| 305 | vmovl.u8 uvrg_xxxx0, uvrgb0 @ uvrg_xxxx0 = { uv0, rg0, b0-, -- } | |
| 306 | ldrh y0, [ v_a, #10 ] @ load y0 | |
| 307 | ||
| 308 | vmovl.u8 uvrg_xxxx1, uvrgb1 @ uvrg_xxxx1 = { uv1, rg1, b1-, -- } | |
| 309 | ldrh y1, [ v_b, #10 ] @ load y1 | |
| 310 | ||
| 311 | vmovl.u8 uvrg_xxxx2, uvrgb2 @ uvrg_xxxx2 = { uv2, rg2, b2-, -- } | |
| 312 | ldrh y2, [ v_c, #10 ] @ load y2 | |
| 313 | ||
| 314 | vmov.u8 uvrg0b, uvrg0 @ uvrg0b = { uv0, rg0 } | |
| 315 | vdup.u16 xxxx0, x0_y0[0] @ xxxx0 = { xx0, xx0 } | |
| 316 | ||
| 317 | orr x1_x2, x1, x2, lsl #16 @ x1_x2 = { x1, x2 } | |
| 318 | pkhbt x0_x1, x0, x1, lsl #16 @ x0_x1 = { x0, x1 } | |
| 319 | ||
| 320 | vmov.u8 uvrg1b, uvrg1 @ uvrg1b = { uv1, rg1 } | |
| 321 | vdup.u16 xxxx1, x1_y1[0] @ xxxx1 = { xx1, xx1 } | |
| 322 | ||
| 323 | vmov.u8 uvrg2b, uvrg2 @ uvrg2b = { uv2, rg2 } | |
| 324 | vdup.u16 xxxx2, x2_y2[0] @ xxxx2 = { xx2, xx2 } | |
| 325 | ||
| 326 | ldrb b2, [ v_c, #4 ] @ load b2 | |
| 327 | orr y0_y1, y0, y1, lsl #16 @ y0_y1 = { y0, y1 } | |
| 328 | ||
| 329 | ldrb b1, [ v_b, #4 ] @ load b1 | |
| 330 | orr y1_y2, y1, y2, lsl #16 @ y1_y2 = { y1, y2 } | |
| 331 | ||
| 332 | vdup.u16 yyyy0, x0_y0[1] @ yyyy0 = { yy0, yy0 } | |
| 333 | vsub.s16 d0_ab, x1_ab, x0_ab | |
| 334 | ||
| 335 | ldrb b0, [ v_a, #4 ] @ load b0 | |
| 336 | orr b1_b2, b1, b2, lsl #16 @ b1_b2 = { b1, b2 } | |
| 337 | ||
| 338 | vdup.u16 yyyy1, x1_y1[1] @ yyyy1 = { yy1, yy1 } | |
| 339 | vsub.s16 d2_ab, x2_ab, x1_ab | |
| 340 | ||
| 341 | vdup.u16 yyyy2, x2_y2[1] @ yyyy2 = { yy2, yy2 } | |
| 342 | vsub.s16 d1_ab, y2_ab, y1_ab | |
| 343 | ||
| 344 | orr b0_b1, b0, b1, lsl #16 @ b1_b2 = { b1, b2 } | |
| 345 | ssub16 dx, x1_x2, x0_x1 @ dx = { x1 - x0, x2 - x1 } | |
| 346 | ||
| 347 | ssub16 dy, y1_y2, y0_y1 @ dy = { y1 - y0, y2 - y1 } | |
| 348 | ssub16 db, b1_b2, b0_b1 @ db = { b1 - b0, b2 - b1 } | |
| 349 | ||
| 350 | vsub.s16 d3_ab, y1_ab, y0_ab | |
| 351 | smusdx ga_by, dx, db @ ga_by = ((x1 - x0) * (b2 - b1)) - | |
| 352 | @ ((x2 - X1) * (b1 - b0)) | |
| 353 | vmull.s16 ga_uvrg_x, d0_a, d1_a | |
| 354 | smusdx ga_bx, db, dy @ ga_bx = ((b1 - b0) * (y2 - y1)) - | |
| 355 | @ ((b2 - b1) * (y1 - y0)) | |
| 356 | vmlsl.s16 ga_uvrg_x, d2_a, d3_a | |
| 357 | movs gs_bx, ga_bx, asr #31 | |
| 358 | ||
| 359 | vmull.s16 ga_uvrg_y, d0_b, d1_b | |
| 360 | rsbmi ga_bx, ga_bx, #0 | |
| 361 | ||
| 362 | vmlsl.s16 ga_uvrg_y, d2_b, d3_b | |
| 363 | movs gs_by, ga_by, asr #31 | |
| 364 | ||
| 365 | vshr.u64 d0, d30, #22 | |
| 366 | mov b_base, b0, lsl #16 | |
| 367 | ||
| 368 | rsbmi ga_by, ga_by, #0 | |
| 369 | vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 | |
| 370 | ||
| 371 | @ r12 = psx_gpu->triangle_winding_offset | |
| 372 | ldrb r12, [ psx_gpu, #psx_gpu_triangle_winding_offset ] | |
| 373 | vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 | |
| 374 | ||
| 375 | add b_base, b_base, #0x8000 | |
| 376 | rsb r12, r12, #0 @ r12 = -(triangle->winding) | |
| 377 | ||
| 378 | vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } | |
| 379 | sub r14, r14, #(62 - 12) @ r14 = shift - (62 - FIXED_BITS) | |
| 380 | ||
| 381 | vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 | |
| 382 | vdup.u32 r_shift, r14 @ r_shift = { shift, shift, shift, shift } | |
| 383 | ||
| 384 | vorr.u32 uvrg_base, #0x8000 | |
| 385 | vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) | |
| 386 | ||
| 387 | vmov area_r_s, s0 @ area_r_s = triangle_reciprocal | |
| 388 | vabs.s32 ga_uvrg_y, ga_uvrg_y @ ga_uvrg_y = abs(ga_uvrg_y) | |
| 389 | ||
| 390 | vmull.u32 gw_rg_x, ga_rg_x, d0[0] | |
| 391 | vmull.u32 gw_uv_x, ga_uv_x, d0[0] | |
| 392 | vmull.u32 gw_rg_y, ga_rg_y, d0[0] | |
| 393 | vmull.u32 gw_uv_y, ga_uv_y, d0[0] | |
| 394 | ||
| 395 | vshl.u64 gw_rg_x, gw_rg_x, r_shift | |
| 396 | vshl.u64 gw_uv_x, gw_uv_x, r_shift | |
| 397 | vshl.u64 gw_rg_y, gw_rg_y, r_shift | |
| 398 | vshl.u64 gw_uv_y, gw_uv_y, r_shift | |
| 399 | ||
| 400 | veor.u32 gs_uvrg_x, gs_uvrg_x, w_mask | |
| 401 | vmovn.u64 g_uv_x, gw_uv_x | |
| 402 | ||
| 403 | veor.u32 gs_uvrg_y, gs_uvrg_y, w_mask | |
| 404 | vmovn.u64 g_rg_x, gw_rg_x | |
| 405 | ||
| 406 | veor.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
| 407 | vmovn.u64 g_uv_y, gw_uv_y | |
| 408 | ||
| 409 | vsub.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
| 410 | vmovn.u64 g_rg_y, gw_rg_y | |
| 411 | ||
| 412 | veor.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
| 413 | mov ga_bx, ga_bx, lsl #13 | |
| 414 | ||
| 415 | vsub.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
| 416 | mov ga_by, ga_by, lsl #13 | |
| 417 | ||
| 418 | vdup.u32 x0_y0, x0 | |
| 419 | umull gw_bx_l, gw_bx_h, ga_bx, area_r_s | |
| 420 | ||
| 421 | vshl.u32 g_uvrg_x, g_uvrg_x, #4 | |
| 422 | vshl.u32 g_uvrg_y, g_uvrg_y, #4 | |
| 423 | ||
| 424 | umull gw_by_l, gw_by_h, ga_by, area_r_s | |
| 425 | vmls.s32 uvrg_base, ga_uvrg_x, x0_y0[0] | |
| 426 | ||
| 427 | eor gs_bx, gs_bx, r12 | |
| 428 | vadd.u32 uvrg_dx2, uvrg_dx1, uvrg_dx1 | |
| 429 | ||
| 430 | veor.u32 uvrg_dx0, uvrg_dx0, uvrg_dx0 | |
| 431 | eor gs_by, gs_by, r12 | |
| 432 | ||
| 433 | rsb r11, r14, #0 @ r11 = negative shift for scalar lsr | |
| 434 | add store_a, psx_gpu, #psx_gpu_uvrg_offset | |
| 435 | ||
| 436 | sub r11, r11, #(32 - 13) | |
| 437 | ||
| 438 | add store_b, store_a, #16 | |
| 439 | mov store_inc, #32 | |
| 440 | ||
| 441 | vadd.u32 uvrg_dx3, uvrg_dx2, uvrg_dx1 | |
| 442 | vst1.u32 { uvrg_base }, [ store_a, : 128 ], store_inc | |
| 443 | ||
| 444 | vst1.u32 { uvrg_dx1 }, [ store_b, : 128 ], store_inc | |
| 445 | mov g_bx, gw_bx_h, lsr r11 | |
| 446 | ||
| 447 | vst1.u32 { g_uvrg_y }, [ store_a, : 128 ], store_inc | |
| 448 | mov g_by, gw_by_h, lsr r11 | |
| 449 | ||
| 450 | vst4.u32 { uvrg_dx0l, uvrg_dx1l, uvrg_dx2l, uvrg_dx3l }, \ | |
| 451 | [ store_b, : 128 ], store_inc | |
| 452 | eor g_bx, g_bx, gs_bx | |
| 453 | ||
| 454 | vst4.u32 { uvrg_dx0h, uvrg_dx1h, uvrg_dx2h, uvrg_dx3h }, \ | |
| 455 | [ store_b, : 128 ], store_inc | |
| 456 | sub g_bx, g_bx, gs_bx | |
| 457 | ||
| 458 | lsl g_bx, g_bx, #4 | |
| 459 | eor g_by, g_by, gs_by | |
| 460 | ||
| 461 | mls b_base, g_bx, x0, b_base | |
| 462 | sub g_by, g_by, gs_by | |
| 463 | ||
| 464 | lsl g_by, g_by, #4 | |
| 465 | mov g_bx0, #0 | |
| 466 | ||
| 467 | add g_bx2, g_bx, g_bx | |
| 468 | add g_bx3, g_bx, g_bx2 | |
| 469 | ||
| 470 | stmia store_b, { g_bx0, g_bx, g_bx2, g_bx3, b_base, g_by } | |
| 471 | ||
| 472 | ldmia sp!, { r4 - r11, pc } | |
| 473 | ||
| 474 | ||
| 475 | #define psx_gpu r0 | |
| 476 | #define v_a r1 | |
| 477 | #define v_b r2 | |
| 478 | #define v_c r3 | |
| 479 | ||
| 480 | #define temp r14 | |
| 481 | ||
| 482 | #define x_a r4 | |
| 483 | #define x_b r5 | |
| 484 | #define x_c r6 | |
| 485 | #define y_a r1 | |
| 486 | #define y_b r2 | |
| 487 | #define y_c r3 | |
| 488 | ||
| 489 | #define height_minor_a r7 | |
| 490 | #define height_minor_b r8 | |
| 491 | #define height_major r9 | |
| 492 | #define height r9 | |
| 493 | ||
| 494 | #define reciprocal_table_ptr r10 | |
| 495 | ||
| 496 | #define edge_alt_low r4 | |
| 497 | #define edge_alt_high r5 | |
| 498 | #define edge_dx_dy_alt r6 | |
| 499 | #define edge_shift_alt r10 | |
| 500 | ||
| 501 | #define edge_dx_dy_alt_low r4 | |
| 502 | #define edge_dx_dy_alt_high r5 | |
| 503 | ||
| 504 | #define span_edge_data r4 | |
| 505 | #define span_uvrg_offset r5 | |
| 506 | #define span_b_offset r6 | |
| 507 | ||
| 508 | #define clip r14 | |
| 509 | ||
| 510 | #define b r11 | |
| 511 | #define b_dy r12 | |
| 512 | ||
| 513 | ||
| 514 | #define alternate_x q0 | |
| 515 | #define alternate_dx_dy q1 | |
| 516 | #define alternate_x_32 q2 | |
| 517 | ||
| 518 | #define alternate_x_low d0 | |
| 519 | #define alternate_x_high d1 | |
| 520 | #define alternate_dx_dy_low d2 | |
| 521 | #define alternate_dx_dy_high d3 | |
| 522 | #define alternate_x_32_low d4 | |
| 523 | #define alternate_x_32_high d5 | |
| 524 | ||
| 525 | #define left_x q3 | |
| 526 | #define right_x q4 | |
| 527 | #define left_dx_dy q5 | |
| 528 | #define right_dx_dy q6 | |
| 529 | #define left_edge q7 | |
| 530 | #define right_edge q8 | |
| 531 | ||
| 532 | #define left_x_low d6 | |
| 533 | #define left_x_high d7 | |
| 534 | #define right_x_low d8 | |
| 535 | #define right_x_high d9 | |
| 536 | #define left_dx_dy_low d10 | |
| 537 | #define left_dx_dy_high d11 | |
| 538 | #define right_dx_dy_low d12 | |
| 539 | #define right_dx_dy_high d13 | |
| 540 | #define left_edge_low d14 | |
| 541 | #define left_edge_high d15 | |
| 542 | #define right_edge_low d16 | |
| 543 | #define right_edge_high d17 | |
| 544 | ||
| 545 | #define y_mid_point d18 | |
| 546 | #define c_0x0004 d19 | |
| 547 | ||
| 548 | #define left_right_x_16 q11 | |
| 549 | #define span_shifts_y q12 | |
| 550 | #define c_0x0001 q13 | |
| 551 | ||
| 552 | #define span_shifts d24 | |
| 553 | #define y_x4 d25 | |
| 554 | #define c_0xFFFE d26 | |
| 555 | #define c_0x0007 d27 | |
| 556 | ||
| 557 | #define left_right_x_16_low d22 | |
| 558 | #define left_right_x_16_high d23 | |
| 559 | ||
| 560 | #define uvrg q14 | |
| 561 | #define uvrg_dy q15 | |
| 562 | ||
| 563 | #define alternate_x_16 d4 | |
| 564 | ||
| 565 | #define v_clip q3 | |
| 566 | #define v_clip_low d6 | |
| 567 | ||
| 568 | #define right_x_32 q10 | |
| 569 | #define left_x_32 q11 | |
| 570 | #define alternate_select d24 | |
| 571 | ||
| 572 | #define right_x_32_low d20 | |
| 573 | #define right_x_32_high d21 | |
| 574 | #define left_x_32_low d22 | |
| 575 | #define left_x_32_high d23 | |
| 576 | ||
| 577 | #define edges_xy q0 | |
| 578 | #define edges_dx_dy d2 | |
| 579 | #define edge_shifts d3 | |
| 580 | #define edge_shifts_64 q2 | |
| 581 | ||
| 582 | #define edges_xy_left d0 | |
| 583 | #define edges_xy_right d1 | |
| 584 | ||
| 585 | #define height_reciprocals d6 | |
| 586 | #define heights d7 | |
| 587 | ||
| 588 | #define widths d8 | |
| 589 | #define c_0x01 d9 | |
| 590 | #define x_starts d10 | |
| 591 | #define x_ends d11 | |
| 592 | ||
| 593 | #define heights_b d12 | |
| 594 | #define edges_dx_dy_64 q10 | |
| 595 | ||
| 596 | #define edges_dx_dy_64_left d20 | |
| 597 | #define edges_dx_dy_64_right d21 | |
| 598 | ||
| 599 | ||
| 600 | #define setup_spans_prologue() \ | |
| 601 | stmdb sp!, { r4 - r11, lr }; \ | |
| 602 | \ | |
| 603 | ldrsh x_a, [ v_a, #8 ]; \ | |
| 604 | ldrsh x_b, [ v_b, #8 ]; \ | |
| 605 | ldrsh x_c, [ v_c, #8 ]; \ | |
| 606 | ldrsh y_a, [ v_a, #10 ]; \ | |
| 607 | ldrsh y_b, [ v_b, #10 ]; \ | |
| 608 | ldrsh y_c, [ v_c, #10 ]; \ | |
| 609 | \ | |
| 610 | add temp, psx_gpu, #psx_gpu_uvrg_offset; \ | |
| 611 | vld1.32 { uvrg }, [ temp ]; \ | |
| 612 | add temp, psx_gpu, #psx_gpu_uvrg_dy_offset; \ | |
| 613 | vld1.32 { uvrg_dy }, [ temp ]; \ | |
| 614 | movw reciprocal_table_ptr, :lower16:reciprocal_table; \ | |
| 615 | movt reciprocal_table_ptr, :upper16:reciprocal_table; \ | |
| 616 | \ | |
| 617 | vmov.u32 c_0x01, #0x01 \ | |
| 618 | ||
| 619 | #define setup_spans_load_b() \ | |
| 620 | ldr b, [ psx_gpu, #psx_gpu_b_offset ]; \ | |
| 621 | ldr b_dy, [ psx_gpu, #psx_gpu_b_dy_offset ] \ | |
| 622 | ||
| 623 | #define setup_spans_prologue_b() \ | |
| 624 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 625 | add temp, psx_gpu, #psx_gpu_viewport_start_x_offset; \ | |
| 626 | \ | |
| 627 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 628 | vmov.u16 c_0x0004, #0x0004; \ | |
| 629 | \ | |
| 630 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 631 | vmov.u16 c_0x0001, #0x0001; \ | |
| 632 | \ | |
| 633 | vld1.u16 { left_edge_low[], left_edge_high[] }, [ temp ]; \ | |
| 634 | add temp, psx_gpu, #psx_gpu_viewport_end_x_offset; \ | |
| 635 | \ | |
| 636 | vld1.u16 { right_edge_low[], right_edge_high[] }, [ temp ]; \ | |
| 637 | vadd.u16 right_edge, right_edge, c_0x0001; \ | |
| 638 | \ | |
| 639 | vmov.u16 c_0x0007, #0x0007; \ | |
| 640 | vmvn.u16 c_0xFFFE, #0x0001 \ | |
| 641 | ||
| 642 | ||
| 643 | #define compute_edge_delta_x2() \ | |
| 644 | ldr temp, [ reciprocal_table_ptr, height, lsl #2 ]; \ | |
| 645 | \ | |
| 646 | vdup.u32 heights, height; \ | |
| 647 | vsub.u32 widths, x_ends, x_starts; \ | |
| 648 | \ | |
| 649 | vdup.u32 edge_shifts, temp; \ | |
| 650 | vsub.u32 heights_b, heights, c_0x01; \ | |
| 651 | vshr.u32 height_reciprocals, edge_shifts, #12; \ | |
| 652 | \ | |
| 653 | vmla.s32 heights_b, x_starts, heights; \ | |
| 654 | vbic.u16 edge_shifts, #0xE0; \ | |
| 655 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
| 656 | vmull.s32 edges_xy, heights_b, height_reciprocals \ | |
| 657 | ||
| 658 | #define width_alt r6 | |
| 659 | #define height_reciprocal_alt r11 | |
| 660 | #define height_b_alt r12 | |
| 661 | ||
| 662 | #define compute_edge_delta_x3(start_c, height_a, height_b) \ | |
| 663 | vmov.u32 heights, height_a, height_b; \ | |
| 664 | ldr temp, [ reciprocal_table_ptr, height_a, lsl #2 ]; \ | |
| 665 | vmov.u32 edge_shifts[0], temp; \ | |
| 666 | ldr temp, [ reciprocal_table_ptr, height_b, lsl #2 ]; \ | |
| 667 | vmov.u32 edge_shifts[1], temp; \ | |
| 668 | ldr edge_shift_alt, [ reciprocal_table_ptr, height_minor_b, lsl #2 ]; \ | |
| 669 | \ | |
| 670 | vsub.u32 widths, x_ends, x_starts; \ | |
| 671 | sub width_alt, x_c, start_c; \ | |
| 672 | \ | |
| 673 | vsub.u32 heights_b, heights, c_0x01; \ | |
| 674 | sub height_b_alt, height_minor_b, #1; \ | |
| 675 | \ | |
| 676 | vshr.u32 height_reciprocals, edge_shifts, #12; \ | |
| 677 | lsr height_reciprocal_alt, edge_shift_alt, #12; \ | |
| 678 | \ | |
| 679 | vmla.s32 heights_b, x_starts, heights; \ | |
| 680 | mla height_b_alt, height_minor_b, start_c, height_b_alt; \ | |
| 681 | \ | |
| 682 | vbic.u16 edge_shifts, #0xE0; \ | |
| 683 | and edge_shift_alt, edge_shift_alt, #0x1F; \ | |
| 684 | \ | |
| 685 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
| 686 | mul edge_dx_dy_alt, width_alt, height_reciprocal_alt; \ | |
| 687 | \ | |
| 688 | vmull.s32 edges_xy, heights_b, height_reciprocals; \ | |
| 689 | smull edge_alt_low, edge_alt_high, height_b_alt, height_reciprocal_alt \ | |
| 690 | ||
| 691 | ||
| 692 | #define setup_spans_adjust_y_up() \ | |
| 693 | vsub.u32 y_x4, y_x4, c_0x0004 \ | |
| 694 | ||
| 695 | #define setup_spans_adjust_y_down() \ | |
| 696 | vadd.u32 y_x4, y_x4, c_0x0004 \ | |
| 697 | ||
| 698 | #define setup_spans_adjust_interpolants_up() \ | |
| 699 | vsub.u32 uvrg, uvrg, uvrg_dy; \ | |
| 700 | sub b, b, b_dy \ | |
| 701 | ||
| 702 | #define setup_spans_adjust_interpolants_down() \ | |
| 703 | vadd.u32 uvrg, uvrg, uvrg_dy; \ | |
| 704 | add b, b, b_dy \ | |
| 705 | ||
| 706 | ||
| 707 | #define setup_spans_clip_interpolants_increment() \ | |
| 708 | mla b, b_dy, clip, b; \ | |
| 709 | vmla.s32 uvrg, uvrg_dy, v_clip \ | |
| 710 | ||
| 711 | #define setup_spans_clip_interpolants_decrement() \ | |
| 712 | mls b, b_dy, clip, b; \ | |
| 713 | vmls.s32 uvrg, uvrg_dy, v_clip \ | |
| 714 | ||
| 715 | #define setup_spans_clip_alternate_yes() \ | |
| 716 | smlal edge_alt_low, edge_alt_high, edge_dx_dy_alt, clip \ | |
| 717 | ||
| 718 | #define setup_spans_clip_alternate_no() \ | |
| 719 | ||
| 720 | #define setup_spans_clip(direction, alternate_active) \ | |
| 721 | vdup.u32 v_clip, clip; \ | |
| 722 | setup_spans_clip_alternate_##alternate_active(); \ | |
| 723 | setup_spans_clip_interpolants_##direction(); \ | |
| 724 | vmlal.s32 edges_xy, edges_dx_dy, v_clip_low \ | |
| 725 | ||
| 726 | ||
| 727 | #define setup_spans_adjust_edges_alternate_no(left_index, right_index) \ | |
| 728 | vmovl.s32 edge_shifts_64, edge_shifts; \ | |
| 729 | vmovl.s32 edges_dx_dy_64, edges_dx_dy; \ | |
| 730 | \ | |
| 731 | vshl.s64 edges_xy, edges_xy, edge_shifts_64; \ | |
| 732 | vshl.s64 edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64; \ | |
| 733 | \ | |
| 734 | vmov left_x_low, edges_xy_##left_index; \ | |
| 735 | vmov right_x_low, edges_xy_##right_index; \ | |
| 736 | \ | |
| 737 | vmov left_dx_dy_low, edges_dx_dy_64_##left_index; \ | |
| 738 | vmov left_dx_dy_high, edges_dx_dy_64_##left_index; \ | |
| 739 | vmov right_dx_dy_low, edges_dx_dy_64_##right_index; \ | |
| 740 | vmov right_dx_dy_high, edges_dx_dy_64_##right_index; \ | |
| 741 | \ | |
| 742 | vadd.u64 left_x_high, left_x_low, left_dx_dy_low; \ | |
| 743 | vadd.u64 right_x_high, right_x_low, right_dx_dy_low; \ | |
| 744 | \ | |
| 745 | vadd.u64 left_dx_dy, left_dx_dy, left_dx_dy; \ | |
| 746 | vadd.u64 right_dx_dy, right_dx_dy, right_dx_dy \ | |
| 747 | ||
| 748 | ||
| 749 | #define setup_spans_adjust_edges_alternate_yes(left_index, right_index) \ | |
| 750 | setup_spans_adjust_edges_alternate_no(left_index, right_index); \ | |
| 751 | \ | |
| 752 | vdup.u16 y_mid_point, y_b; \ | |
| 753 | rsb temp, edge_shift_alt, #32; \ | |
| 754 | \ | |
| 755 | lsl edge_alt_high, edge_alt_high, edge_shift_alt; \ | |
| 756 | orr edge_alt_high, edge_alt_high, edge_alt_low, lsr temp; \ | |
| 757 | lsl edge_alt_low, edge_alt_low, edge_shift_alt; \ | |
| 758 | vmov alternate_x_low, edge_alt_low, edge_alt_high; \ | |
| 759 | \ | |
| 760 | asr edge_dx_dy_alt_high, edge_dx_dy_alt, temp; \ | |
| 761 | lsl edge_dx_dy_alt_low, edge_dx_dy_alt, edge_shift_alt; \ | |
| 762 | vmov alternate_dx_dy_low, edge_dx_dy_alt_low, edge_dx_dy_alt_high; \ | |
| 763 | vmov alternate_dx_dy_high, alternate_dx_dy_low; \ | |
| 764 | \ | |
| 765 | vadd.u64 alternate_x_high, alternate_x_low, alternate_dx_dy_low; \ | |
| 766 | vadd.u64 alternate_dx_dy, alternate_dx_dy, alternate_dx_dy \ | |
| 767 | ||
| 768 | ||
| 769 | #define setup_spans_y_select_up() \ | |
| 770 | vclt.s16 alternate_select, y_x4, y_mid_point \ | |
| 771 | ||
| 772 | #define setup_spans_y_select_down() \ | |
| 773 | vcgt.s16 alternate_select, y_x4, y_mid_point \ | |
| 774 | ||
| 775 | ||
| 776 | #define setup_spans_alternate_select_left() \ | |
| 777 | vbit.u16 left_right_x_16_low, alternate_x_16, alternate_select \ | |
| 778 | ||
| 779 | #define setup_spans_alternate_select_right() \ | |
| 780 | vbit.u16 left_right_x_16_high, alternate_x_16, alternate_select \ | |
| 781 | ||
| 782 | ||
| 783 | #define setup_spans_set_x4_alternate_yes(alternate, direction) \ | |
| 784 | vshrn.s64 alternate_x_32_low, alternate_x, #32; \ | |
| 785 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
| 786 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
| 787 | \ | |
| 788 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
| 789 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 790 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 791 | \ | |
| 792 | vshrn.s64 alternate_x_32_high, alternate_x, #32; \ | |
| 793 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
| 794 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
| 795 | \ | |
| 796 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
| 797 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 798 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 799 | \ | |
| 800 | vmovn.u32 alternate_x_16, alternate_x_32; \ | |
| 801 | setup_spans_y_select_##direction(); \ | |
| 802 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
| 803 | \ | |
| 804 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
| 805 | setup_spans_alternate_select_##alternate(); \ | |
| 806 | \ | |
| 807 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 808 | str b, [ span_b_offset ], #4; \ | |
| 809 | setup_spans_adjust_interpolants_##direction(); \ | |
| 810 | \ | |
| 811 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
| 812 | \ | |
| 813 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 814 | str b, [ span_b_offset ], #4; \ | |
| 815 | setup_spans_adjust_interpolants_##direction(); \ | |
| 816 | \ | |
| 817 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
| 818 | \ | |
| 819 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 820 | str b, [ span_b_offset ], #4; \ | |
| 821 | setup_spans_adjust_interpolants_##direction(); \ | |
| 822 | \ | |
| 823 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
| 824 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
| 825 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
| 826 | \ | |
| 827 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 828 | str b, [ span_b_offset ], #4; \ | |
| 829 | setup_spans_adjust_interpolants_##direction(); \ | |
| 830 | \ | |
| 831 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
| 832 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
| 833 | \ | |
| 834 | vst4.u16 { left_right_x_16, span_shifts_y }, [ span_edge_data ]!; \ | |
| 835 | \ | |
| 836 | setup_spans_adjust_y_##direction() \ | |
| 837 | ||
| 838 | ||
| 839 | #define setup_spans_set_x4_alternate_no(alternate, direction) \ | |
| 840 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
| 841 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
| 842 | \ | |
| 843 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 844 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 845 | \ | |
| 846 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
| 847 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
| 848 | \ | |
| 849 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 850 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 851 | \ | |
| 852 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
| 853 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
| 854 | \ | |
| 855 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 856 | str b, [ span_b_offset ], #4; \ | |
| 857 | setup_spans_adjust_interpolants_##direction(); \ | |
| 858 | \ | |
| 859 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
| 860 | \ | |
| 861 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 862 | str b, [ span_b_offset ], #4; \ | |
| 863 | setup_spans_adjust_interpolants_##direction(); \ | |
| 864 | \ | |
| 865 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
| 866 | \ | |
| 867 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 868 | str b, [ span_b_offset ], #4; \ | |
| 869 | setup_spans_adjust_interpolants_##direction(); \ | |
| 870 | \ | |
| 871 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
| 872 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
| 873 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
| 874 | \ | |
| 875 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
| 876 | str b, [ span_b_offset ], #4; \ | |
| 877 | setup_spans_adjust_interpolants_##direction(); \ | |
| 878 | \ | |
| 879 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
| 880 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
| 881 | \ | |
| 882 | vst4.u16 { left_right_x_16, span_shifts_y }, [ span_edge_data ]!; \ | |
| 883 | \ | |
| 884 | setup_spans_adjust_y_##direction() \ | |
| 885 | ||
| 886 | ||
| 887 | #define edge_adjust_low r11 | |
| 888 | #define edge_adjust_high r12 | |
| 889 | ||
| 890 | #define setup_spans_alternate_adjust_yes() \ | |
| 891 | smull edge_adjust_low, edge_adjust_high, edge_dx_dy_alt, height_minor_a; \ | |
| 892 | subs edge_alt_low, edge_alt_low, edge_adjust_low; \ | |
| 893 | sbc edge_alt_high, edge_alt_high, edge_adjust_high \ | |
| 894 | ||
| 895 | #define setup_spans_alternate_adjust_no() \ | |
| 896 | ||
| 897 | ||
| 898 | #define setup_spans_down(left_index, right_index, alternate, alternate_active) \ | |
| 899 | setup_spans_alternate_adjust_##alternate_active(); \ | |
| 900 | setup_spans_load_b(); \ | |
| 901 | \ | |
| 902 | ldrsh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ]; \ | |
| 903 | subs y_c, y_c, temp; \ | |
| 904 | subgt height, height, y_c; \ | |
| 905 | addgt height, height, #1; \ | |
| 906 | \ | |
| 907 | ldrsh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ]; \ | |
| 908 | subs clip, temp, y_a; \ | |
| 909 | ble 0f; \ | |
| 910 | \ | |
| 911 | sub height, height, clip; \ | |
| 912 | add y_a, y_a, clip; \ | |
| 913 | setup_spans_clip(increment, alternate_active); \ | |
| 914 | \ | |
| 915 | 0: \ | |
| 916 | cmp height, #0; \ | |
| 917 | ble 1f; \ | |
| 918 | \ | |
| 919 | orr temp, y_a, y_a, lsl #16; \ | |
| 920 | add temp, temp, #(1 << 16); \ | |
| 921 | add y_a, temp, #2; \ | |
| 922 | add y_a, y_a, #(2 << 16); \ | |
| 923 | vmov.u32 y_x4, temp, y_a; \ | |
| 924 | \ | |
| 925 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
| 926 | right_index); \ | |
| 927 | setup_spans_prologue_b(); \ | |
| 928 | \ | |
| 929 | strh height, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
| 930 | \ | |
| 931 | 2: \ | |
| 932 | setup_spans_set_x4_alternate_##alternate_active(alternate, down); \ | |
| 933 | subs height, height, #4; \ | |
| 934 | bhi 2b; \ | |
| 935 | \ | |
| 936 | 1: \ | |
| 937 | ||
| 938 | ||
| 939 | #define setup_spans_alternate_pre_increment_yes() \ | |
| 940 | adds edge_alt_low, edge_alt_low, edge_dx_dy_alt; \ | |
| 941 | adc edge_alt_high, edge_alt_high, edge_dx_dy_alt, asr #31 \ | |
| 942 | ||
| 943 | #define setup_spans_alternate_pre_increment_no() \ | |
| 944 | ||
| 945 | ||
| 946 | #define setup_spans_up_decrement_yes() \ | |
| 947 | suble height, height, #1 \ | |
| 948 | ||
| 949 | #define setup_spans_up_decrement_no() \ | |
| 950 | ||
| 951 | ||
| 952 | #define setup_spans_up(left_index, right_index, alternate, alternate_active) \ | |
| 953 | setup_spans_alternate_adjust_##alternate_active(); \ | |
| 954 | setup_spans_load_b(); \ | |
| 955 | sub y_a, y_a, #1; \ | |
| 956 | \ | |
| 957 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ]; \ | |
| 958 | subs temp, temp, y_c; \ | |
| 959 | subgt height, height, temp; \ | |
| 960 | setup_spans_up_decrement_##alternate_active(); \ | |
| 961 | \ | |
| 962 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ]; \ | |
| 963 | subs clip, y_a, temp; \ | |
| 964 | ble 0f; \ | |
| 965 | \ | |
| 966 | sub height, height, clip; \ | |
| 967 | sub y_a, y_a, clip; \ | |
| 968 | setup_spans_clip(decrement, alternate_active); \ | |
| 969 | \ | |
| 970 | 0: \ | |
| 971 | cmp height, #0; \ | |
| 972 | ble 1f; \ | |
| 973 | \ | |
| 974 | orr temp, y_a, y_a, lsl #16; \ | |
| 975 | sub temp, temp, #(1 << 16); \ | |
| 976 | sub y_a, temp, #2; \ | |
| 977 | sub y_a, y_a, #(2 << 16); \ | |
| 978 | vmov.u32 y_x4, temp, y_a; \ | |
| 979 | \ | |
| 980 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy; \ | |
| 981 | \ | |
| 982 | setup_spans_alternate_pre_increment_##alternate_active(); \ | |
| 983 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
| 984 | right_index); \ | |
| 985 | setup_spans_adjust_interpolants_up(); \ | |
| 986 | setup_spans_prologue_b(); \ | |
| 987 | \ | |
| 988 | strh height, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
| 989 | \ | |
| 990 | 2: \ | |
| 991 | setup_spans_set_x4_alternate_##alternate_active(alternate, up); \ | |
| 992 | subs height, height, #4; \ | |
| 993 | bhi 2b; \ | |
| 994 | \ | |
| 995 | 1: \ | |
| 996 | ||
| 997 | ||
| 998 | #define setup_spans_epilogue() \ | |
| 999 | ldmia sp!, { r4 - r11, pc } \ | |
| 1000 | ||
| 1001 | ||
| 1002 | #define setup_spans_up_up(minor, major) \ | |
| 1003 | setup_spans_prologue(); \ | |
| 1004 | sub height_minor_a, y_a, y_b; \ | |
| 1005 | sub height_minor_b, y_b, y_c; \ | |
| 1006 | sub height, y_a, y_c; \ | |
| 1007 | \ | |
| 1008 | vdup.u32 x_starts, x_a; \ | |
| 1009 | vmov.u32 x_ends, x_c, x_b; \ | |
| 1010 | \ | |
| 1011 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
| 1012 | setup_spans_up(major, minor, minor, yes); \ | |
| 1013 | setup_spans_epilogue() \ | |
| 1014 | ||
| 1015 | function(setup_spans_up_left) | |
| 1016 | setup_spans_up_up(left, right) | |
| 1017 | ||
| 1018 | function(setup_spans_up_right) | |
| 1019 | setup_spans_up_up(right, left) | |
| 1020 | ||
| 1021 | ||
| 1022 | #define setup_spans_down_down(minor, major) \ | |
| 1023 | setup_spans_prologue(); \ | |
| 1024 | sub height_minor_a, y_b, y_a; \ | |
| 1025 | sub height_minor_b, y_c, y_b; \ | |
| 1026 | sub height, y_c, y_a; \ | |
| 1027 | \ | |
| 1028 | vdup.u32 x_starts, x_a; \ | |
| 1029 | vmov.u32 x_ends, x_c, x_b; \ | |
| 1030 | \ | |
| 1031 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
| 1032 | setup_spans_down(major, minor, minor, yes); \ | |
| 1033 | setup_spans_epilogue() \ | |
| 1034 | ||
| 1035 | function(setup_spans_down_left) | |
| 1036 | setup_spans_down_down(left, right) | |
| 1037 | ||
| 1038 | function(setup_spans_down_right) | |
| 1039 | setup_spans_down_down(right, left) | |
| 1040 | ||
| 1041 | ||
| 1042 | #define setup_spans_up_flat() \ | |
| 1043 | sub height, y_a, y_c; \ | |
| 1044 | \ | |
| 1045 | compute_edge_delta_x2(); \ | |
| 1046 | setup_spans_up(left, right, none, no); \ | |
| 1047 | setup_spans_epilogue() \ | |
| 1048 | ||
| 1049 | function(setup_spans_up_a) | |
| 1050 | setup_spans_prologue() | |
| 1051 | ||
| 1052 | vmov.u32 x_starts, x_a, x_b | |
| 1053 | vdup.u32 x_ends, x_c | |
| 1054 | ||
| 1055 | setup_spans_up_flat() | |
| 1056 | ||
| 1057 | function(setup_spans_up_b) | |
| 1058 | setup_spans_prologue() | |
| 1059 | ||
| 1060 | vdup.u32 x_starts, x_a | |
| 1061 | vmov.u32 x_ends, x_b, x_c | |
| 1062 | ||
| 1063 | setup_spans_up_flat() | |
| 1064 | ||
| 1065 | #define setup_spans_down_flat() \ | |
| 1066 | sub height, y_c, y_a; \ | |
| 1067 | \ | |
| 1068 | compute_edge_delta_x2(); \ | |
| 1069 | setup_spans_down(left, right, none, no); \ | |
| 1070 | setup_spans_epilogue() \ | |
| 1071 | ||
| 1072 | function(setup_spans_down_a) | |
| 1073 | setup_spans_prologue() | |
| 1074 | ||
| 1075 | vmov.u32 x_starts, x_a, x_b | |
| 1076 | vdup.u32 x_ends, x_c | |
| 1077 | ||
| 1078 | setup_spans_down_flat() | |
| 1079 | ||
| 1080 | function(setup_spans_down_b) | |
| 1081 | setup_spans_prologue() | |
| 1082 | ||
| 1083 | vdup.u32 x_starts, x_a | |
| 1084 | vmov.u32 x_ends, x_b, x_c | |
| 1085 | ||
| 1086 | setup_spans_down_flat() | |
| 1087 | ||
| 1088 | ||
| 1089 | #define middle_y r9 | |
| 1090 | ||
| 1091 | #define edges_xy_b q11 | |
| 1092 | #define edges_dx_dy_b d26 | |
| 1093 | #define edge_shifts_b d27 | |
| 1094 | #define edges_dx_dy_and_shifts_b q13 | |
| 1095 | #define height_increment d20 | |
| 1096 | ||
| 1097 | #define edges_dx_dy_and_shifts q1 | |
| 1098 | ||
| 1099 | #define edges_xy_b_left d22 | |
| 1100 | #define edges_xy_b_right d23 | |
| 1101 | ||
| 1102 | #define setup_spans_up_down_load_edge_set_b() \ | |
| 1103 | vmov edges_xy, edges_xy_b; \ | |
| 1104 | vmov edges_dx_dy_and_shifts, edges_dx_dy_and_shifts_b \ | |
| 1105 | ||
| 1106 | ||
| 1107 | function(setup_spans_up_down) | |
| 1108 | setup_spans_prologue() | |
| 1109 | ||
| 1110 | // s32 middle_y = y_a; | |
| 1111 | sub height_minor_a, y_a, y_b | |
| 1112 | sub height_minor_b, y_c, y_a | |
| 1113 | sub height_major, y_c, y_b | |
| 1114 | ||
| 1115 | vmov.u32 x_starts, x_a, x_c | |
| 1116 | vdup.u32 x_ends, x_b | |
| 1117 | ||
| 1118 | compute_edge_delta_x3(x_a, height_minor_a, height_major) | |
| 1119 | ||
| 1120 | mov temp, #0 | |
| 1121 | vmov.u32 height_increment, temp, height_minor_b | |
| 1122 | vmlal.s32 edges_xy, edges_dx_dy, height_increment | |
| 1123 | ||
| 1124 | vmov edges_xy_b_left, edge_alt_low, edge_alt_high | |
| 1125 | vmov edges_xy_b_right, edges_xy_right | |
| 1126 | ||
| 1127 | vmov edge_shifts_b, edge_shifts | |
| 1128 | vmov.u32 edge_shifts_b[0], edge_shift_alt | |
| 1129 | ||
| 1130 | vneg.s32 edges_dx_dy_b, edges_dx_dy | |
| 1131 | vmov.u32 edges_dx_dy_b[0], edge_dx_dy_alt | |
| 1132 | ||
| 1133 | mov middle_y, y_a | |
| 1134 | ||
| 1135 | setup_spans_load_b() | |
| 1136 | sub y_a, y_a, #1 | |
| 1137 | ||
| 1138 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ] | |
| 1139 | subs temp, temp, y_b | |
| 1140 | subgt height_minor_a, height_minor_a, temp | |
| 1141 | ||
| 1142 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ] | |
| 1143 | subs clip, y_a, temp | |
| 1144 | ble 0f | |
| 1145 | ||
| 1146 | sub height_minor_a, height_minor_a, clip | |
| 1147 | sub y_a, y_a, clip | |
| 1148 | setup_spans_clip(decrement, no) | |
| 1149 | ||
| 1150 | 0: | |
| 1151 | cmp height_minor_a, #0 | |
| 1152 | ble 3f | |
| 1153 | ||
| 1154 | orr temp, y_a, y_a, lsl #16 | |
| 1155 | sub temp, temp, #(1 << 16) | |
| 1156 | sub y_a, temp, #2 | |
| 1157 | sub y_a, y_a, #(2 << 16) | |
| 1158 | vmov.u32 y_x4, temp, y_a | |
| 1159 | ||
| 1160 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy | |
| 1161 | ||
| 1162 | strh height_minor_a, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
| 1163 | ||
| 1164 | setup_spans_adjust_edges_alternate_no(left, right); | |
| 1165 | setup_spans_adjust_interpolants_up() | |
| 1166 | setup_spans_up_down_load_edge_set_b() | |
| 1167 | ||
| 1168 | setup_spans_prologue_b() | |
| 1169 | ||
| 1170 | ||
| 1171 | 2: | |
| 1172 | setup_spans_set_x4_alternate_no(none, up) | |
| 1173 | subs height_minor_a, height_minor_a, #4 | |
| 1174 | bhi 2b | |
| 1175 | ||
| 1176 | add span_edge_data, span_edge_data, height_minor_a, lsl #3 | |
| 1177 | add span_uvrg_offset, span_uvrg_offset, height_minor_a, lsl #4 | |
| 1178 | add span_b_offset, span_b_offset, height_minor_a, lsl #2 | |
| 1179 | ||
| 1180 | 4: | |
| 1181 | add temp, psx_gpu, #psx_gpu_uvrg_offset | |
| 1182 | vld1.32 { uvrg }, [ temp ] | |
| 1183 | mov y_a, middle_y | |
| 1184 | ||
| 1185 | setup_spans_load_b() | |
| 1186 | ||
| 1187 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ] | |
| 1188 | subs y_c, y_c, temp | |
| 1189 | subgt height_minor_b, height_minor_b, y_c | |
| 1190 | addgt height_minor_b, height_minor_b, #1 | |
| 1191 | ||
| 1192 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ] | |
| 1193 | subs clip, temp, y_a | |
| 1194 | ble 0f | |
| 1195 | ||
| 1196 | sub height_minor_b, height_minor_b, clip | |
| 1197 | add y_a, y_a, clip | |
| 1198 | setup_spans_clip(increment, no) | |
| 1199 | ||
| 1200 | 0: | |
| 1201 | cmp height_minor_b, #0 | |
| 1202 | ble 1f | |
| 1203 | ||
| 1204 | orr temp, y_a, y_a, lsl #16 | |
| 1205 | add temp, temp, #(1 << 16) | |
| 1206 | add y_a, temp, #2 | |
| 1207 | add y_a, y_a, #(2 << 16) | |
| 1208 | vmov.u32 y_x4, temp, y_a | |
| 1209 | ||
| 1210 | setup_spans_adjust_edges_alternate_no(left, right) | |
| 1211 | ||
| 1212 | ldrh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
| 1213 | add temp, temp, height_minor_b | |
| 1214 | strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
| 1215 | ||
| 1216 | 2: | |
| 1217 | setup_spans_set_x4_alternate_no(none, down) | |
| 1218 | subs height_minor_b, height_minor_b, #4 | |
| 1219 | bhi 2b | |
| 1220 | ||
| 1221 | 1: | |
| 1222 | setup_spans_epilogue() | |
| 1223 | ||
| 1224 | 3: | |
| 1225 | setup_spans_up_down_load_edge_set_b() | |
| 1226 | setup_spans_prologue_b() | |
| 1227 | bal 4b | |
| 1228 | ||
| 1229 | ||
| 1230 | #undef span_uvrg_offset | |
| 1231 | #undef span_edge_data | |
| 1232 | #undef span_b_offset | |
| 1233 | #undef left_x | |
| 1234 | #undef b | |
| 1235 | ||
| 1236 | #define psx_gpu r0 | |
| 1237 | #define num_spans r1 | |
| 1238 | #define span_uvrg_offset r2 | |
| 1239 | #define span_edge_data r3 | |
| 1240 | #define span_b_offset r4 | |
| 1241 | #define b_dx r5 | |
| 1242 | #define span_num_blocks r6 | |
| 1243 | #define y r7 | |
| 1244 | #define left_x r8 | |
| 1245 | #define b r9 | |
| 1246 | #define dither_offset_ptr r10 | |
| 1247 | #define block_ptr_a r11 | |
| 1248 | #define fb_ptr r12 | |
| 1249 | #define num_blocks r14 | |
| 1250 | ||
| 1251 | #define uvrg_dx_ptr r2 | |
| 1252 | #define texture_mask_ptr r3 | |
| 1253 | #define dither_shift r8 | |
| 1254 | #define dither_row r10 | |
| 1255 | ||
| 1256 | #define c_32 r7 | |
| 1257 | #define b_dx4 r8 | |
| 1258 | #define b_dx8 r9 | |
| 1259 | #define block_ptr_b r10 | |
| 1260 | ||
| 1261 | #define block_span_ptr r10 | |
| 1262 | #define right_mask r8 | |
| 1263 | ||
| 1264 | #define color r2 | |
| 1265 | #define color_r r3 | |
| 1266 | #define color_g r4 | |
| 1267 | #define color_b r5 | |
| 1268 | ||
| 1269 | #undef uvrg | |
| 1270 | ||
| 1271 | #define u_block q0 | |
| 1272 | #define v_block q1 | |
| 1273 | #define r_block q2 | |
| 1274 | #define g_block q3 | |
| 1275 | #define b_block q4 | |
| 1276 | ||
| 1277 | #define uv_dx4 d10 | |
| 1278 | #define rg_dx4 d11 | |
| 1279 | #define uv_dx8 d12 | |
| 1280 | #define rg_dx8 d13 | |
| 1281 | #define b_whole_8 d14 | |
| 1282 | #define fb_mask_ptrs d15 | |
| 1283 | ||
| 1284 | #define uvrg_dx4 q5 | |
| 1285 | #define uvrg_dx8 q6 | |
| 1286 | #define uv_dx8 d12 | |
| 1287 | #define rg_dx8 d13 | |
| 1288 | ||
| 1289 | #define u_whole q8 | |
| 1290 | #define v_whole q9 | |
| 1291 | #define r_whole q10 | |
| 1292 | #define g_whole q11 | |
| 1293 | #define b_whole q12 | |
| 1294 | ||
| 1295 | #define u_whole_low d16 | |
| 1296 | #define u_whole_high d17 | |
| 1297 | #define v_whole_low d18 | |
| 1298 | #define v_whole_high d19 | |
| 1299 | #define r_whole_low d20 | |
| 1300 | #define r_whole_high d21 | |
| 1301 | #define g_whole_low d22 | |
| 1302 | #define g_whole_high d23 | |
| 1303 | #define b_whole_low d24 | |
| 1304 | #define b_whole_high d25 | |
| 1305 | ||
| 1306 | #define dx4 q13 | |
| 1307 | #define dx8 q13 | |
| 1308 | ||
| 1309 | #define u_whole_8 d26 | |
| 1310 | #define v_whole_8 d27 | |
| 1311 | #define u_whole_8b d24 | |
| 1312 | #define r_whole_8 d24 | |
| 1313 | #define g_whole_8 d25 | |
| 1314 | ||
| 1315 | #define uv_whole_8 q13 | |
| 1316 | #define uv_whole_8b q14 | |
| 1317 | ||
| 1318 | #define dither_offsets q14 | |
| 1319 | #define texture_mask q15 | |
| 1320 | #define texture_mask_u d30 | |
| 1321 | #define texture_mask_v d31 | |
| 1322 | ||
| 1323 | #define dither_offsets_short d28 | |
| 1324 | ||
| 1325 | #define v_left_x q8 | |
| 1326 | #define uvrg q9 | |
| 1327 | #define block_span q10 | |
| 1328 | ||
| 1329 | #define uv d18 | |
| 1330 | #define rg d19 | |
| 1331 | ||
| 1332 | #define draw_mask q1 | |
| 1333 | #define draw_mask_edge q13 | |
| 1334 | #define test_mask q0 | |
| 1335 | ||
| 1336 | #define uvrg_dx q3 | |
| 1337 | ||
| 1338 | #define colors q2 | |
| 1339 | ||
| 1340 | #define setup_blocks_texture_swizzled() \ | |
| 1341 | vand.u8 u_whole_8b, u_whole_8, texture_mask_u; \ | |
| 1342 | vsli.u8 u_whole_8, v_whole_8, #4; \ | |
| 1343 | vsri.u8 v_whole_8, u_whole_8b, #4 \ | |
| 1344 | ||
| 1345 | #define setup_blocks_texture_unswizzled() \ | |
| 1346 | ||
| 1347 | ||
| 1348 | #define setup_blocks_shaded_textured_builder(swizzling) \ | |
| 1349 | .align 3; \ | |
| 1350 | \ | |
| 1351 | function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ | |
| 1352 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
| 1353 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ | |
| 1354 | \ | |
| 1355 | vld1.u32 { uvrg_dx }, [ uvrg_dx_ptr, :128 ]; \ | |
| 1356 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ | |
| 1357 | \ | |
| 1358 | cmp num_spans, #0; \ | |
| 1359 | bxeq lr; \ | |
| 1360 | \ | |
| 1361 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 1362 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ | |
| 1363 | \ | |
| 1364 | ldr b_dx, [ psx_gpu, #psx_gpu_b_dx_offset ]; \ | |
| 1365 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
| 1366 | \ | |
| 1367 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [ texture_mask_ptr, :16 ]; \ | |
| 1368 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 1369 | \ | |
| 1370 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 1371 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 1372 | \ | |
| 1373 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 1374 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1375 | \ | |
| 1376 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 1377 | \ | |
| 1378 | 0: \ | |
| 1379 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1380 | \ | |
| 1381 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
| 1382 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
| 1383 | \ | |
| 1384 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
| 1385 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ | |
| 1386 | \ | |
| 1387 | cmp span_num_blocks, #0; \ | |
| 1388 | beq 1f; \ | |
| 1389 | \ | |
| 1390 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
| 1391 | add num_blocks, span_num_blocks, num_blocks; \ | |
| 1392 | \ | |
| 1393 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 1394 | bgt 2f; \ | |
| 1395 | \ | |
| 1396 | 3: \ | |
| 1397 | ldr b, [ span_b_offset ]; \ | |
| 1398 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 1399 | \ | |
| 1400 | vdup.u32 v_left_x, left_x; \ | |
| 1401 | and y, y, #0x3; \ | |
| 1402 | \ | |
| 1403 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
| 1404 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
| 1405 | \ | |
| 1406 | mla b, b_dx, left_x, b; \ | |
| 1407 | and dither_shift, left_x, #0x03; \ | |
| 1408 | \ | |
| 1409 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
| 1410 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ | |
| 1411 | \ | |
| 1412 | mov dither_shift, dither_shift, lsl #3; \ | |
| 1413 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
| 1414 | \ | |
| 1415 | mov c_32, #32; \ | |
| 1416 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1417 | \ | |
| 1418 | mov dither_row, dither_row, ror dither_shift; \ | |
| 1419 | mov b_dx4, b_dx, lsl #2; \ | |
| 1420 | \ | |
| 1421 | vdup.u32 dither_offsets_short, dither_row; \ | |
| 1422 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
| 1423 | \ | |
| 1424 | vdup.u32 b_block, b; \ | |
| 1425 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
| 1426 | \ | |
| 1427 | vdup.u32 u_block, uv[0]; \ | |
| 1428 | mov b_dx8, b_dx, lsl #3; \ | |
| 1429 | \ | |
| 1430 | vdup.u32 v_block, uv[1]; \ | |
| 1431 | vdup.u32 r_block, rg[0]; \ | |
| 1432 | vdup.u32 g_block, rg[1]; \ | |
| 1433 | \ | |
| 1434 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 1435 | \ | |
| 1436 | vadd.u32 u_block, u_block, block_span; \ | |
| 1437 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 1438 | \ | |
| 1439 | vadd.u32 v_block, v_block, block_span; \ | |
| 1440 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 1441 | \ | |
| 1442 | vadd.u32 r_block, r_block, block_span; \ | |
| 1443 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 1444 | \ | |
| 1445 | vadd.u32 g_block, g_block, block_span; \ | |
| 1446 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]; \ | |
| 1447 | \ | |
| 1448 | vadd.u32 b_block, b_block, block_span; \ | |
| 1449 | add block_ptr_b, block_ptr_a, #16; \ | |
| 1450 | \ | |
| 1451 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1452 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1453 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 1454 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 1455 | \ | |
| 1456 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1457 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 1458 | \ | |
| 1459 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1460 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1461 | \ | |
| 1462 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1463 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 1464 | \ | |
| 1465 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 1466 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 1467 | \ | |
| 1468 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 1469 | vdup.u32 dx4, b_dx4; \ | |
| 1470 | \ | |
| 1471 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 1472 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1473 | \ | |
| 1474 | vadd.u32 u_block, u_block, dx8; \ | |
| 1475 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1476 | \ | |
| 1477 | vadd.u32 v_block, v_block, dx8; \ | |
| 1478 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 1479 | \ | |
| 1480 | vadd.u32 r_block, r_block, dx8; \ | |
| 1481 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 1482 | \ | |
| 1483 | vadd.u32 g_block, g_block, dx8; \ | |
| 1484 | vdup.u32 dx8, b_dx8; \ | |
| 1485 | \ | |
| 1486 | vadd.u32 b_block, b_block, dx8; \ | |
| 1487 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1488 | \ | |
| 1489 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1490 | \ | |
| 1491 | vmovn.u16 b_whole_8, b_whole; \ | |
| 1492 | pld [ fb_ptr ]; \ | |
| 1493 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1494 | \ | |
| 1495 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1496 | setup_blocks_texture_##swizzling(); \ | |
| 1497 | \ | |
| 1498 | vmovn.u16 r_whole_8, r_whole; \ | |
| 1499 | beq 5f; \ | |
| 1500 | \ | |
| 1501 | 4: \ | |
| 1502 | vmovn.u16 g_whole_8, g_whole; \ | |
| 1503 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1504 | \ | |
| 1505 | vst2.u8 { u_whole_8, v_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
| 1506 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1507 | \ | |
| 1508 | vst1.u32 { r_whole_8, g_whole_8 }, [ block_ptr_b, :128 ], c_32; \ | |
| 1509 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 1510 | \ | |
| 1511 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
| 1512 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 1513 | \ | |
| 1514 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1515 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 1516 | \ | |
| 1517 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1518 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1519 | \ | |
| 1520 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1521 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 1522 | \ | |
| 1523 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 1524 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 1525 | \ | |
| 1526 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 1527 | vdup.u32 dx4, b_dx4; \ | |
| 1528 | \ | |
| 1529 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 1530 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1531 | \ | |
| 1532 | vadd.u32 u_block, u_block, dx8; \ | |
| 1533 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1534 | \ | |
| 1535 | vadd.u32 v_block, v_block, dx8; \ | |
| 1536 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 1537 | \ | |
| 1538 | vadd.u32 r_block, r_block, dx8; \ | |
| 1539 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 1540 | \ | |
| 1541 | vadd.u32 g_block, g_block, dx8; \ | |
| 1542 | vdup.u32 dx8, b_dx8; \ | |
| 1543 | \ | |
| 1544 | vadd.u32 b_block, b_block, dx8; \ | |
| 1545 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1546 | \ | |
| 1547 | add fb_ptr, fb_ptr, #16; \ | |
| 1548 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1549 | \ | |
| 1550 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
| 1551 | vmovn.u16 b_whole_8, b_whole; \ | |
| 1552 | \ | |
| 1553 | pld [ fb_ptr ]; \ | |
| 1554 | \ | |
| 1555 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1556 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1557 | \ | |
| 1558 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1559 | setup_blocks_texture_##swizzling(); \ | |
| 1560 | \ | |
| 1561 | vmovn.u16 r_whole_8, r_whole; \ | |
| 1562 | bne 4b; \ | |
| 1563 | \ | |
| 1564 | 5: \ | |
| 1565 | vmovn.u16 g_whole_8, g_whole; \ | |
| 1566 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
| 1567 | \ | |
| 1568 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
| 1569 | vdup.u8 draw_mask, right_mask; \ | |
| 1570 | \ | |
| 1571 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
| 1572 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 1573 | vzip.u8 u_whole_8, v_whole_8; \ | |
| 1574 | \ | |
| 1575 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
| 1576 | vst1.u32 { r_whole_8, g_whole_8 }, [ block_ptr_b, :128 ], c_32; \ | |
| 1577 | vst1.u32 { uv_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
| 1578 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
| 1579 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
| 1580 | \ | |
| 1581 | 1: \ | |
| 1582 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 1583 | add span_b_offset, span_b_offset, #4; \ | |
| 1584 | \ | |
| 1585 | add span_edge_data, span_edge_data, #8; \ | |
| 1586 | subs num_spans, num_spans, #1; \ | |
| 1587 | \ | |
| 1588 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 1589 | bne 0b; \ | |
| 1590 | \ | |
| 1591 | ldmia sp!, { r4 - r11, pc }; \ | |
| 1592 | \ | |
| 1593 | 2: \ | |
| 1594 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
| 1595 | vpush { texture_mask }; \ | |
| 1596 | vpush { uvrg_dx4 }; \ | |
| 1597 | \ | |
| 1598 | stmdb sp!, { r0 - r3, r12, r14 }; \ | |
| 1599 | bl flush_render_block_buffer; \ | |
| 1600 | ldmia sp!, { r0 - r3, r12, r14 }; \ | |
| 1601 | \ | |
| 1602 | vpop { uvrg_dx4 }; \ | |
| 1603 | vpop { texture_mask }; \ | |
| 1604 | \ | |
| 1605 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
| 1606 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1607 | \ | |
| 1608 | mov num_blocks, span_num_blocks; \ | |
| 1609 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1610 | bal 3b \ | |
| 1611 | ||
| 1612 | ||
| 1613 | setup_blocks_shaded_textured_builder(swizzled) | |
| 1614 | setup_blocks_shaded_textured_builder(unswizzled) | |
| 1615 | ||
| 1616 | ||
| 1617 | #define setup_blocks_unshaded_textured_builder(swizzling) \ | |
| 1618 | .align 3; \ | |
| 1619 | \ | |
| 1620 | function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ | |
| 1621 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
| 1622 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ | |
| 1623 | \ | |
| 1624 | vld1.u32 { uvrg_dx }, [ uvrg_dx_ptr, :128 ]; \ | |
| 1625 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ | |
| 1626 | \ | |
| 1627 | cmp num_spans, #0; \ | |
| 1628 | bxeq lr; \ | |
| 1629 | \ | |
| 1630 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 1631 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ | |
| 1632 | \ | |
| 1633 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
| 1634 | \ | |
| 1635 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [ texture_mask_ptr, :16 ]; \ | |
| 1636 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 1637 | \ | |
| 1638 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 1639 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 1640 | \ | |
| 1641 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1642 | \ | |
| 1643 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 1644 | \ | |
| 1645 | 0: \ | |
| 1646 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1647 | \ | |
| 1648 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
| 1649 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
| 1650 | \ | |
| 1651 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
| 1652 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ | |
| 1653 | \ | |
| 1654 | cmp span_num_blocks, #0; \ | |
| 1655 | beq 1f; \ | |
| 1656 | \ | |
| 1657 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
| 1658 | add num_blocks, span_num_blocks, num_blocks; \ | |
| 1659 | \ | |
| 1660 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 1661 | bgt 2f; \ | |
| 1662 | \ | |
| 1663 | 3: \ | |
| 1664 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 1665 | \ | |
| 1666 | vdup.u32 v_left_x, left_x; \ | |
| 1667 | and y, y, #0x3; \ | |
| 1668 | \ | |
| 1669 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
| 1670 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
| 1671 | \ | |
| 1672 | and dither_shift, left_x, #0x03; \ | |
| 1673 | \ | |
| 1674 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
| 1675 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ | |
| 1676 | \ | |
| 1677 | mov dither_shift, dither_shift, lsl #3; \ | |
| 1678 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
| 1679 | \ | |
| 1680 | mov c_32, #32; \ | |
| 1681 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1682 | \ | |
| 1683 | mov dither_row, dither_row, ror dither_shift; \ | |
| 1684 | \ | |
| 1685 | vdup.u32 dither_offsets_short, dither_row; \ | |
| 1686 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
| 1687 | \ | |
| 1688 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
| 1689 | \ | |
| 1690 | vdup.u32 u_block, uv[0]; \ | |
| 1691 | \ | |
| 1692 | vdup.u32 v_block, uv[1]; \ | |
| 1693 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 1694 | \ | |
| 1695 | vadd.u32 u_block, u_block, block_span; \ | |
| 1696 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 1697 | \ | |
| 1698 | vadd.u32 v_block, v_block, block_span; \ | |
| 1699 | add block_ptr_b, block_ptr_a, #16; \ | |
| 1700 | \ | |
| 1701 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1702 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1703 | \ | |
| 1704 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1705 | \ | |
| 1706 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1707 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1708 | \ | |
| 1709 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1710 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1711 | \ | |
| 1712 | vadd.u32 u_block, u_block, dx8; \ | |
| 1713 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1714 | \ | |
| 1715 | vadd.u32 v_block, v_block, dx8; \ | |
| 1716 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1717 | \ | |
| 1718 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1719 | \ | |
| 1720 | pld [ fb_ptr ]; \ | |
| 1721 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1722 | \ | |
| 1723 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1724 | setup_blocks_texture_##swizzling(); \ | |
| 1725 | \ | |
| 1726 | beq 5f; \ | |
| 1727 | \ | |
| 1728 | 4: \ | |
| 1729 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1730 | \ | |
| 1731 | vst2.u8 { u_whole_8, v_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
| 1732 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1733 | \ | |
| 1734 | add block_ptr_b, block_ptr_b, #32; \ | |
| 1735 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
| 1736 | \ | |
| 1737 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1738 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1739 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1740 | \ | |
| 1741 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1742 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1743 | \ | |
| 1744 | vadd.u32 u_block, u_block, dx8; \ | |
| 1745 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1746 | \ | |
| 1747 | vadd.u32 v_block, v_block, dx8; \ | |
| 1748 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1749 | \ | |
| 1750 | add fb_ptr, fb_ptr, #16; \ | |
| 1751 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1752 | \ | |
| 1753 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
| 1754 | pld [ fb_ptr ]; \ | |
| 1755 | \ | |
| 1756 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1757 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1758 | \ | |
| 1759 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1760 | setup_blocks_texture_##swizzling(); \ | |
| 1761 | \ | |
| 1762 | bne 4b; \ | |
| 1763 | \ | |
| 1764 | 5: \ | |
| 1765 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
| 1766 | \ | |
| 1767 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
| 1768 | vdup.u8 draw_mask, right_mask; \ | |
| 1769 | \ | |
| 1770 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
| 1771 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 1772 | vzip.u8 u_whole_8, v_whole_8; \ | |
| 1773 | \ | |
| 1774 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
| 1775 | add block_ptr_b, block_ptr_b, #32; \ | |
| 1776 | vst1.u32 { uv_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
| 1777 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
| 1778 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
| 1779 | \ | |
| 1780 | 1: \ | |
| 1781 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 1782 | add span_edge_data, span_edge_data, #8; \ | |
| 1783 | subs num_spans, num_spans, #1; \ | |
| 1784 | \ | |
| 1785 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 1786 | bne 0b; \ | |
| 1787 | \ | |
| 1788 | ldmia sp!, { r4 - r11, pc }; \ | |
| 1789 | \ | |
| 1790 | 2: \ | |
| 1791 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
| 1792 | vpush { texture_mask }; \ | |
| 1793 | vpush { uvrg_dx4 }; \ | |
| 1794 | \ | |
| 1795 | stmdb sp!, { r0 - r3, r12, r14 }; \ | |
| 1796 | bl flush_render_block_buffer; \ | |
| 1797 | ldmia sp!, { r0 - r3, r12, r14 }; \ | |
| 1798 | \ | |
| 1799 | vpop { uvrg_dx4 }; \ | |
| 1800 | vpop { texture_mask }; \ | |
| 1801 | \ | |
| 1802 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
| 1803 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1804 | \ | |
| 1805 | mov num_blocks, span_num_blocks; \ | |
| 1806 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1807 | bal 3b \ | |
| 1808 | ||
| 1809 | ||
| 1810 | setup_blocks_unshaded_textured_builder(swizzled) | |
| 1811 | setup_blocks_unshaded_textured_builder(unswizzled) | |
| 1812 | ||
| 1813 | ||
| 1814 | .align 3 | |
| 1815 | ||
| 1816 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) | |
| 1817 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
| 1818 | veor.u32 draw_mask, draw_mask, draw_mask | |
| 1819 | ||
| 1820 | cmp num_spans, #0 | |
| 1821 | bxeq lr | |
| 1822 | ||
| 1823 | stmdb sp!, { r4 - r11, r14 } | |
| 1824 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
| 1825 | ||
| 1826 | ldr color, [ psx_gpu, #psx_gpu_triangle_color_offset ] | |
| 1827 | ||
| 1828 | ubfx color_r, color, #3, #5 | |
| 1829 | ubfx color_g, color, #11, #5 | |
| 1830 | ubfx color_b, color, #19, #5 | |
| 1831 | ||
| 1832 | orr color, color_r, color_b, lsl #10 | |
| 1833 | orr color, color, color_g, lsl #5 | |
| 1834 | ||
| 1835 | vdup.u16 colors, color | |
| 1836 | ||
| 1837 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 1838 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
| 1839 | ||
| 1840 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
| 1841 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6 | |
| 1842 | ||
| 1843 | 0: | |
| 1844 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ] | |
| 1845 | ldrh y, [ span_edge_data, #edge_data_y_offset ] | |
| 1846 | ||
| 1847 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
| 1848 | ||
| 1849 | cmp span_num_blocks, #0 | |
| 1850 | beq 1f | |
| 1851 | ||
| 1852 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ] | |
| 1853 | add num_blocks, span_num_blocks, num_blocks | |
| 1854 | ||
| 1855 | cmp num_blocks, #MAX_BLOCKS | |
| 1856 | bgt 2f | |
| 1857 | ||
| 1858 | 3: | |
| 1859 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 1860 | and y, y, #0x3 | |
| 1861 | ||
| 1862 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
| 1863 | mov c_32, #32 | |
| 1864 | ||
| 1865 | subs span_num_blocks, span_num_blocks, #1 | |
| 1866 | ||
| 1867 | add block_ptr_b, block_ptr_a, #16 | |
| 1868 | pld [ fb_ptr ] | |
| 1869 | ||
| 1870 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
| 1871 | beq 5f | |
| 1872 | ||
| 1873 | 4: | |
| 1874 | vst1.u32 { draw_mask }, [ block_ptr_a, :128 ], c_32 | |
| 1875 | vst1.u32 { colors }, [ block_ptr_b, :128 ], c_32 | |
| 1876 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32 | |
| 1877 | ||
| 1878 | add fb_ptr, fb_ptr, #16 | |
| 1879 | add block_ptr_b, block_ptr_b, #32 | |
| 1880 | ||
| 1881 | pld [ fb_ptr ] | |
| 1882 | ||
| 1883 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
| 1884 | subs span_num_blocks, span_num_blocks, #1 | |
| 1885 | ||
| 1886 | bne 4b | |
| 1887 | ||
| 1888 | 5: | |
| 1889 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ] | |
| 1890 | ||
| 1891 | vdup.u8 draw_mask_edge, right_mask | |
| 1892 | vtst.u16 draw_mask_edge, draw_mask_edge, test_mask | |
| 1893 | ||
| 1894 | vst1.u32 { colors }, [ block_ptr_b, :128 ], c_32 | |
| 1895 | vst1.u32 { draw_mask_edge }, [ block_ptr_a, :128 ], c_32 | |
| 1896 | add block_ptr_b, block_ptr_b, #32 | |
| 1897 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32 | |
| 1898 | ||
| 1899 | 1: | |
| 1900 | add span_edge_data, span_edge_data, #8 | |
| 1901 | subs num_spans, num_spans, #1 | |
| 1902 | ||
| 1903 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 1904 | bne 0b | |
| 1905 | ||
| 1906 | ldmia sp!, { r4 - r11, pc } | |
| 1907 | ||
| 1908 | 2: | |
| 1909 | vpush { colors } | |
| 1910 | ||
| 1911 | stmdb sp!, { r0 - r3, r12, r14 } | |
| 1912 | bl flush_render_block_buffer | |
| 1913 | ldmia sp!, { r0 - r3, r12, r14 } | |
| 1914 | ||
| 1915 | vpop { colors } | |
| 1916 | ||
| 1917 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
| 1918 | veor.u32 draw_mask, draw_mask, draw_mask | |
| 1919 | ||
| 1920 | mov num_blocks, span_num_blocks | |
| 1921 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
| 1922 | bal 3b | |
| 1923 | ||
| 1924 | ||
| 1925 | #define mask_msb_scalar r14 | |
| 1926 | ||
| 1927 | #define msb_mask q15 | |
| 1928 | ||
| 1929 | #define pixels_low d16 | |
| 1930 | ||
| 1931 | #define msb_mask_low d30 | |
| 1932 | #define msb_mask_high d31 | |
| 1933 | ||
| 1934 | ||
| 1935 | .align 3 | |
| 1936 | ||
| 1937 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) | |
| 1938 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
| 1939 | ||
| 1940 | cmp num_spans, #0 | |
| 1941 | bxeq lr | |
| 1942 | ||
| 1943 | stmdb sp!, { r4 - r11, r14 } | |
| 1944 | ||
| 1945 | ldr color, [ psx_gpu, #psx_gpu_triangle_color_offset ] | |
| 1946 | ||
| 1947 | ubfx color_r, color, #3, #5 | |
| 1948 | ubfx color_g, color, #11, #5 | |
| 1949 | ||
| 1950 | ldrh mask_msb_scalar, [ psx_gpu, #psx_gpu_mask_msb_offset ] | |
| 1951 | ubfx color_b, color, #19, #5 | |
| 1952 | ||
| 1953 | orr color, color_r, color_b, lsl #10 | |
| 1954 | orr color, color, color_g, lsl #5 | |
| 1955 | orr color, color, mask_msb_scalar | |
| 1956 | ||
| 1957 | vdup.u16 colors, color | |
| 1958 | ||
| 1959 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
| 1960 | ||
| 1961 | 0: | |
| 1962 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ] | |
| 1963 | ldrh y, [ span_edge_data, #edge_data_y_offset ] | |
| 1964 | ||
| 1965 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
| 1966 | ||
| 1967 | cmp span_num_blocks, #0 | |
| 1968 | beq 1f | |
| 1969 | ||
| 1970 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ] | |
| 1971 | ||
| 1972 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 1973 | subs span_num_blocks, span_num_blocks, #1 | |
| 1974 | ||
| 1975 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
| 1976 | beq 3f | |
| 1977 | ||
| 1978 | 2: | |
| 1979 | vst1.u32 { colors }, [ fb_ptr ]! | |
| 1980 | subs span_num_blocks, span_num_blocks, #1 | |
| 1981 | ||
| 1982 | bne 2b | |
| 1983 | ||
| 1984 | 3: | |
| 1985 | ldrb right_mask, [ span_edge_data, #edge_data_right_mask_offset ] | |
| 1986 | eor right_mask, right_mask, #0xFF | |
| 1987 | ||
| 1988 | 4: | |
| 1989 | strh color, [ fb_ptr ], #2 | |
| 1990 | movs right_mask, right_mask, lsr #1 | |
| 1991 | bne 4b | |
| 1992 | ||
| 1993 | 1: | |
| 1994 | add span_edge_data, span_edge_data, #8 | |
| 1995 | subs num_spans, num_spans, #1 | |
| 1996 | ||
| 1997 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 1998 | bne 0b | |
| 1999 | ||
| 2000 | ldmia sp!, { r4 - r11, pc } | |
| 2001 | ||
| 2002 | ||
| 2003 | ||
| 2004 | #undef c_64 | |
| 2005 | ||
| 2006 | #define c_64 r7 | |
| 2007 | #define rg_dx_ptr r2 | |
| 2008 | ||
| 2009 | ||
| 2010 | #undef r_block | |
| 2011 | #undef g_block | |
| 2012 | #undef b_block | |
| 2013 | #undef r_whole | |
| 2014 | #undef g_whole | |
| 2015 | #undef b_whole | |
| 2016 | #undef r_whole_low | |
| 2017 | #undef r_whole_high | |
| 2018 | #undef g_whole_low | |
| 2019 | #undef g_whole_high | |
| 2020 | #undef b_whole_low | |
| 2021 | #undef b_whole_high | |
| 2022 | #undef r_whole_8 | |
| 2023 | #undef g_whole_8 | |
| 2024 | #undef b_whole_8 | |
| 2025 | #undef dither_offsets | |
| 2026 | #undef rg_dx4 | |
| 2027 | #undef rg_dx8 | |
| 2028 | #undef dx4 | |
| 2029 | #undef dx8 | |
| 2030 | #undef v_left_x | |
| 2031 | #undef uvrg | |
| 2032 | #undef block_span | |
| 2033 | #undef rg | |
| 2034 | #undef draw_mask | |
| 2035 | #undef test_mask | |
| 2036 | ||
| 2037 | #define r_block q0 | |
| 2038 | #define g_block q1 | |
| 2039 | #define b_block q2 | |
| 2040 | ||
| 2041 | #define r_whole q3 | |
| 2042 | #define g_whole q4 | |
| 2043 | #define b_whole q5 | |
| 2044 | ||
| 2045 | #define r_whole_low d6 | |
| 2046 | #define r_whole_high d7 | |
| 2047 | #define g_whole_low d8 | |
| 2048 | #define g_whole_high d9 | |
| 2049 | #define b_whole_low d10 | |
| 2050 | #define b_whole_high d11 | |
| 2051 | ||
| 2052 | #define gb_whole_8 q6 | |
| 2053 | ||
| 2054 | #define g_whole_8 d12 | |
| 2055 | #define b_whole_8 d13 | |
| 2056 | ||
| 2057 | #define r_whole_8 d14 | |
| 2058 | ||
| 2059 | #define pixels q8 | |
| 2060 | ||
| 2061 | #define rg_dx4 d18 | |
| 2062 | #define rg_dx8 d19 | |
| 2063 | ||
| 2064 | #define dx4 q10 | |
| 2065 | #define dx8 q10 | |
| 2066 | ||
| 2067 | #define v_left_x d6 | |
| 2068 | #define uvrg q4 | |
| 2069 | #define block_span q5 | |
| 2070 | ||
| 2071 | #define rg d9 | |
| 2072 | ||
| 2073 | #define d64_1 d22 | |
| 2074 | #define d64_128 d23 | |
| 2075 | ||
| 2076 | #define d128_4 q12 | |
| 2077 | #define d128_0x7 q13 | |
| 2078 | ||
| 2079 | #define d64_4 d24 | |
| 2080 | ||
| 2081 | #define dither_offsets q14 | |
| 2082 | #define draw_mask q15 | |
| 2083 | ||
| 2084 | #define dither_offsets_low d28 | |
| 2085 | ||
| 2086 | #define rg_dx d0 | |
| 2087 | #define test_mask q10 | |
| 2088 | ||
| 2089 | ||
| 2090 | #define setup_blocks_shaded_untextured_dither_a_dithered() \ | |
| 2091 | vqadd.u8 r_whole_8, r_whole_8, dither_offsets_low; \ | |
| 2092 | vqadd.u8 gb_whole_8, gb_whole_8, dither_offsets; \ | |
| 2093 | ||
| 2094 | #define setup_blocks_shaded_untextured_dither_b_dithered() \ | |
| 2095 | vqsub.u8 r_whole_8, r_whole_8, d64_4; \ | |
| 2096 | vqsub.u8 gb_whole_8, gb_whole_8, d128_4 \ | |
| 2097 | ||
| 2098 | #define setup_blocks_shaded_untextured_dither_a_undithered() \ | |
| 2099 | ||
| 2100 | #define setup_blocks_shaded_untextured_dither_b_undithered() \ | |
| 2101 | ||
| 2102 | ||
| 2103 | #define setup_blocks_shaded_untextured_indirect_builder(dithering) \ | |
| 2104 | .align 3; \ | |
| 2105 | \ | |
| 2106 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ | |
| 2107 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
| 2108 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ | |
| 2109 | \ | |
| 2110 | vld1.u32 { rg_dx }, [ rg_dx_ptr, :64 ]; \ | |
| 2111 | \ | |
| 2112 | cmp num_spans, #0; \ | |
| 2113 | bxeq lr; \ | |
| 2114 | \ | |
| 2115 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 2116 | vshl.u32 rg_dx4, rg_dx, #2; \ | |
| 2117 | \ | |
| 2118 | ldr b_dx, [ psx_gpu, #psx_gpu_b_dx_offset ]; \ | |
| 2119 | vshl.u32 rg_dx8, rg_dx, #3; \ | |
| 2120 | \ | |
| 2121 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 2122 | \ | |
| 2123 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 2124 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 2125 | \ | |
| 2126 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 2127 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 2128 | \ | |
| 2129 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 2130 | vmov.u8 d64_1, #1; \ | |
| 2131 | \ | |
| 2132 | vmov.u8 d128_4, #4; \ | |
| 2133 | vmov.u8 d64_128, #128; \ | |
| 2134 | \ | |
| 2135 | vmov.u8 d128_0x7, #0x7; \ | |
| 2136 | \ | |
| 2137 | 0: \ | |
| 2138 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
| 2139 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
| 2140 | \ | |
| 2141 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
| 2142 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ | |
| 2143 | \ | |
| 2144 | cmp span_num_blocks, #0; \ | |
| 2145 | beq 1f; \ | |
| 2146 | \ | |
| 2147 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
| 2148 | add num_blocks, span_num_blocks, num_blocks; \ | |
| 2149 | \ | |
| 2150 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 2151 | bgt 2f; \ | |
| 2152 | \ | |
| 2153 | 3: \ | |
| 2154 | ldr b, [ span_b_offset ]; \ | |
| 2155 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 2156 | \ | |
| 2157 | vdup.u32 v_left_x, left_x; \ | |
| 2158 | and y, y, #0x3; \ | |
| 2159 | \ | |
| 2160 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
| 2161 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
| 2162 | \ | |
| 2163 | mla b, b_dx, left_x, b; \ | |
| 2164 | and dither_shift, left_x, #0x03; \ | |
| 2165 | \ | |
| 2166 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
| 2167 | vshr.u32 rg_dx, rg_dx4, #2; \ | |
| 2168 | \ | |
| 2169 | mov dither_shift, dither_shift, lsl #3; \ | |
| 2170 | vmla.u32 rg, rg_dx, v_left_x; \ | |
| 2171 | \ | |
| 2172 | mov c_64, #64; \ | |
| 2173 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2174 | \ | |
| 2175 | mov dither_row, dither_row, ror dither_shift; \ | |
| 2176 | mov b_dx4, b_dx, lsl #2; \ | |
| 2177 | \ | |
| 2178 | vdup.u32 dither_offsets, dither_row; \ | |
| 2179 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
| 2180 | \ | |
| 2181 | vdup.u32 b_block, b; \ | |
| 2182 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
| 2183 | \ | |
| 2184 | mov b_dx8, b_dx, lsl #3; \ | |
| 2185 | vdup.u32 r_block, rg[0]; \ | |
| 2186 | vdup.u32 g_block, rg[1]; \ | |
| 2187 | \ | |
| 2188 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 2189 | \ | |
| 2190 | vadd.u32 r_block, r_block, block_span; \ | |
| 2191 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 2192 | \ | |
| 2193 | vadd.u32 g_block, g_block, block_span; \ | |
| 2194 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]; \ | |
| 2195 | \ | |
| 2196 | vadd.u32 b_block, b_block, block_span; \ | |
| 2197 | add block_ptr_b, block_ptr_a, #16; \ | |
| 2198 | \ | |
| 2199 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2200 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2201 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2202 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2203 | \ | |
| 2204 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2205 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2206 | \ | |
| 2207 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2208 | vdup.u32 dx4, b_dx4; \ | |
| 2209 | \ | |
| 2210 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2211 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2212 | \ | |
| 2213 | vadd.u32 r_block, r_block, dx8; \ | |
| 2214 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2215 | \ | |
| 2216 | vadd.u32 g_block, g_block, dx8; \ | |
| 2217 | vdup.u32 dx8, b_dx8; \ | |
| 2218 | \ | |
| 2219 | vadd.u32 b_block, b_block, dx8; \ | |
| 2220 | \ | |
| 2221 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2222 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2223 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2224 | \ | |
| 2225 | beq 5f; \ | |
| 2226 | veor.u32 draw_mask, draw_mask, draw_mask; \ | |
| 2227 | \ | |
| 2228 | 4: \ | |
| 2229 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2230 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2231 | \ | |
| 2232 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2233 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2234 | \ | |
| 2235 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2236 | str fb_ptr, [ block_ptr_a, #44 ]; \ | |
| 2237 | \ | |
| 2238 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2239 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2240 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2241 | \ | |
| 2242 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2243 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2244 | \ | |
| 2245 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2246 | vdup.u32 dx4, b_dx4; \ | |
| 2247 | \ | |
| 2248 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2249 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2250 | \ | |
| 2251 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
| 2252 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2253 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2254 | \ | |
| 2255 | vadd.u32 r_block, r_block, dx8; \ | |
| 2256 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2257 | \ | |
| 2258 | vadd.u32 g_block, g_block, dx8; \ | |
| 2259 | vdup.u32 dx8, b_dx8; \ | |
| 2260 | \ | |
| 2261 | vadd.u32 b_block, b_block, dx8; \ | |
| 2262 | add fb_ptr, fb_ptr, #16; \ | |
| 2263 | \ | |
| 2264 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2265 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2266 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2267 | \ | |
| 2268 | vst1.u32 { draw_mask }, [ block_ptr_a, :128 ], c_64; \ | |
| 2269 | vst1.u32 { pixels }, [ block_ptr_b, :128 ], c_64; \ | |
| 2270 | \ | |
| 2271 | pld [ fb_ptr ]; \ | |
| 2272 | \ | |
| 2273 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2274 | bne 4b; \ | |
| 2275 | \ | |
| 2276 | 5: \ | |
| 2277 | str fb_ptr, [ block_ptr_a, #44 ]; \ | |
| 2278 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2279 | \ | |
| 2280 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
| 2281 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2282 | \ | |
| 2283 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2284 | vdup.u8 draw_mask, right_mask; \ | |
| 2285 | \ | |
| 2286 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2287 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
| 2288 | \ | |
| 2289 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 2290 | \ | |
| 2291 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
| 2292 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2293 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2294 | \ | |
| 2295 | vst1.u32 { draw_mask }, [ block_ptr_a, :128 ], c_64; \ | |
| 2296 | vst1.u32 { pixels }, [ block_ptr_b, :128 ], c_64; \ | |
| 2297 | \ | |
| 2298 | 1: \ | |
| 2299 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 2300 | add span_b_offset, span_b_offset, #4; \ | |
| 2301 | \ | |
| 2302 | add span_edge_data, span_edge_data, #8; \ | |
| 2303 | subs num_spans, num_spans, #1; \ | |
| 2304 | \ | |
| 2305 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 2306 | bne 0b; \ | |
| 2307 | \ | |
| 2308 | ldmia sp!, { r4 - r11, pc }; \ | |
| 2309 | \ | |
| 2310 | 2: \ | |
| 2311 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
| 2312 | vpush { rg_dx4 }; \ | |
| 2313 | \ | |
| 2314 | stmdb sp!, { r0 - r3, r12, r14 }; \ | |
| 2315 | bl flush_render_block_buffer; \ | |
| 2316 | ldmia sp!, { r0 - r3, r12, r14 }; \ | |
| 2317 | \ | |
| 2318 | vpop { rg_dx4 }; \ | |
| 2319 | \ | |
| 2320 | vmov.u8 d64_1, #1; \ | |
| 2321 | vmov.u8 d128_4, #4; \ | |
| 2322 | vmov.u8 d64_128, #128; \ | |
| 2323 | vmov.u8 d128_0x7, #0x7; \ | |
| 2324 | \ | |
| 2325 | vadd.u32 rg_dx8, rg_dx4, rg_dx4; \ | |
| 2326 | \ | |
| 2327 | mov num_blocks, span_num_blocks; \ | |
| 2328 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 2329 | bal 3b \ | |
| 2330 | ||
| 2331 | ||
| 2332 | setup_blocks_shaded_untextured_indirect_builder(undithered) | |
| 2333 | setup_blocks_shaded_untextured_indirect_builder(dithered) | |
| 2334 | ||
| 2335 | ||
| 2336 | #undef draw_mask | |
| 2337 | ||
| 2338 | #define mask_msb_ptr r14 | |
| 2339 | ||
| 2340 | #define draw_mask q0 | |
| 2341 | #define pixels_low d16 | |
| 2342 | ||
| 2343 | ||
| 2344 | ||
| 2345 | #define setup_blocks_shaded_untextured_direct_builder(dithering) \ | |
| 2346 | .align 3; \ | |
| 2347 | \ | |
| 2348 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ | |
| 2349 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
| 2350 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ | |
| 2351 | \ | |
| 2352 | vld1.u32 { rg_dx }, [ rg_dx_ptr, :64 ]; \ | |
| 2353 | \ | |
| 2354 | cmp num_spans, #0; \ | |
| 2355 | bxeq lr; \ | |
| 2356 | \ | |
| 2357 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 2358 | vshl.u32 rg_dx4, rg_dx, #2; \ | |
| 2359 | \ | |
| 2360 | ldr b_dx, [ psx_gpu, #psx_gpu_b_dx_offset ]; \ | |
| 2361 | vshl.u32 rg_dx8, rg_dx, #3; \ | |
| 2362 | \ | |
| 2363 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 2364 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 2365 | \ | |
| 2366 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 2367 | vmov.u8 d64_1, #1; \ | |
| 2368 | \ | |
| 2369 | vmov.u8 d128_4, #4; \ | |
| 2370 | vmov.u8 d64_128, #128; \ | |
| 2371 | \ | |
| 2372 | vmov.u8 d128_0x7, #0x7; \ | |
| 2373 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 2374 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 2375 | \ | |
| 2376 | 0: \ | |
| 2377 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
| 2378 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
| 2379 | \ | |
| 2380 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
| 2381 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ | |
| 2382 | \ | |
| 2383 | cmp span_num_blocks, #0; \ | |
| 2384 | beq 1f; \ | |
| 2385 | \ | |
| 2386 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
| 2387 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 2388 | \ | |
| 2389 | ldr b, [ span_b_offset ]; \ | |
| 2390 | vdup.u32 v_left_x, left_x; \ | |
| 2391 | and y, y, #0x3; \ | |
| 2392 | \ | |
| 2393 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
| 2394 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
| 2395 | \ | |
| 2396 | mla b, b_dx, left_x, b; \ | |
| 2397 | and dither_shift, left_x, #0x03; \ | |
| 2398 | \ | |
| 2399 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
| 2400 | vshr.u32 rg_dx, rg_dx4, #2; \ | |
| 2401 | \ | |
| 2402 | mov dither_shift, dither_shift, lsl #3; \ | |
| 2403 | vmla.u32 rg, rg_dx, v_left_x; \ | |
| 2404 | \ | |
| 2405 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2406 | \ | |
| 2407 | mov dither_row, dither_row, ror dither_shift; \ | |
| 2408 | mov b_dx4, b_dx, lsl #2; \ | |
| 2409 | \ | |
| 2410 | vdup.u32 dither_offsets, dither_row; \ | |
| 2411 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
| 2412 | \ | |
| 2413 | vdup.u32 b_block, b; \ | |
| 2414 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
| 2415 | \ | |
| 2416 | mov b_dx8, b_dx, lsl #3; \ | |
| 2417 | vdup.u32 r_block, rg[0]; \ | |
| 2418 | vdup.u32 g_block, rg[1]; \ | |
| 2419 | \ | |
| 2420 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 2421 | \ | |
| 2422 | vadd.u32 r_block, r_block, block_span; \ | |
| 2423 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
| 2424 | \ | |
| 2425 | vadd.u32 g_block, g_block, block_span; \ | |
| 2426 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]; \ | |
| 2427 | \ | |
| 2428 | vadd.u32 b_block, b_block, block_span; \ | |
| 2429 | add block_ptr_b, block_ptr_a, #16; \ | |
| 2430 | \ | |
| 2431 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2432 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2433 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2434 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2435 | \ | |
| 2436 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2437 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2438 | \ | |
| 2439 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2440 | vdup.u32 dx4, b_dx4; \ | |
| 2441 | \ | |
| 2442 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2443 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2444 | \ | |
| 2445 | vadd.u32 r_block, r_block, dx8; \ | |
| 2446 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2447 | \ | |
| 2448 | vadd.u32 g_block, g_block, dx8; \ | |
| 2449 | vdup.u32 dx8, b_dx8; \ | |
| 2450 | \ | |
| 2451 | vadd.u32 b_block, b_block, dx8; \ | |
| 2452 | \ | |
| 2453 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2454 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2455 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2456 | \ | |
| 2457 | beq 3f; \ | |
| 2458 | \ | |
| 2459 | 2: \ | |
| 2460 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2461 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2462 | \ | |
| 2463 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2464 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2465 | \ | |
| 2466 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2467 | \ | |
| 2468 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2469 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2470 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2471 | \ | |
| 2472 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2473 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2474 | \ | |
| 2475 | vmov pixels, msb_mask; \ | |
| 2476 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2477 | vdup.u32 dx4, b_dx4; \ | |
| 2478 | \ | |
| 2479 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2480 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2481 | \ | |
| 2482 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
| 2483 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2484 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2485 | \ | |
| 2486 | vadd.u32 r_block, r_block, dx8; \ | |
| 2487 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2488 | \ | |
| 2489 | vadd.u32 g_block, g_block, dx8; \ | |
| 2490 | vdup.u32 dx8, b_dx8; \ | |
| 2491 | \ | |
| 2492 | vadd.u32 b_block, b_block, dx8; \ | |
| 2493 | \ | |
| 2494 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2495 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2496 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2497 | \ | |
| 2498 | vst1.u32 { pixels }, [ fb_ptr ]!; \ | |
| 2499 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2500 | bne 2b; \ | |
| 2501 | \ | |
| 2502 | 3: \ | |
| 2503 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2504 | \ | |
| 2505 | ldrb right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
| 2506 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2507 | \ | |
| 2508 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2509 | vmov pixels, msb_mask; \ | |
| 2510 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2511 | eor right_mask, right_mask, #0xFF; \ | |
| 2512 | \ | |
| 2513 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
| 2514 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2515 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2516 | \ | |
| 2517 | 4: \ | |
| 2518 | vst1.u16 { pixels_low[0] }, [ fb_ptr ]!; \ | |
| 2519 | vext.16 pixels, pixels, #1; \ | |
| 2520 | movs right_mask, right_mask, lsr #1; \ | |
| 2521 | bne 4b; \ | |
| 2522 | \ | |
| 2523 | 1: \ | |
| 2524 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 2525 | add span_b_offset, span_b_offset, #4; \ | |
| 2526 | \ | |
| 2527 | add span_edge_data, span_edge_data, #8; \ | |
| 2528 | subs num_spans, num_spans, #1; \ | |
| 2529 | \ | |
| 2530 | bne 0b; \ | |
| 2531 | \ | |
| 2532 | ldmia sp!, { r4 - r11, pc } \ | |
| 2533 | ||
| 2534 | setup_blocks_shaded_untextured_direct_builder(undithered) | |
| 2535 | setup_blocks_shaded_untextured_direct_builder(dithered) | |
| 2536 | ||
| 2537 | ||
| 2538 | #undef psx_gpu | |
| 2539 | #undef num_blocks | |
| 2540 | #undef triangle | |
| 2541 | #undef c_64 | |
| 2542 | ||
| 2543 | #define psx_gpu r0 | |
| 2544 | #define block_ptr r1 | |
| 2545 | #define num_blocks r2 | |
| 2546 | #define uv_01 r3 | |
| 2547 | #define uv_23 r4 | |
| 2548 | #define uv_45 r5 | |
| 2549 | #define uv_67 r6 | |
| 2550 | #define uv_0 r7 | |
| 2551 | #define uv_1 r3 | |
| 2552 | #define uv_2 r8 | |
| 2553 | #define uv_3 r4 | |
| 2554 | #define uv_4 r9 | |
| 2555 | #define uv_5 r5 | |
| 2556 | #define uv_6 r10 | |
| 2557 | #define uv_7 r6 | |
| 2558 | #define texture_ptr r11 | |
| 2559 | ||
| 2560 | #define pixel_0 r7 | |
| 2561 | #define pixel_1 r3 | |
| 2562 | #define pixel_2 r8 | |
| 2563 | #define pixel_3 r4 | |
| 2564 | #define pixel_4 r9 | |
| 2565 | #define pixel_5 r5 | |
| 2566 | #define pixel_6 r10 | |
| 2567 | #define pixel_7 r6 | |
| 2568 | ||
| 2569 | #define pixels_a r7 | |
| 2570 | #define pixels_b r9 | |
| 2571 | #define pixels_c r8 | |
| 2572 | #define pixels_d r10 | |
| 2573 | ||
| 2574 | #define c_64 r0 | |
| 2575 | ||
| 2576 | #define clut_ptr r12 | |
| 2577 | #define current_texture_mask r5 | |
| 2578 | #define dirty_textures_mask r6 | |
| 2579 | ||
| 2580 | #define texels d0 | |
| 2581 | ||
| 2582 | #define clut_low_a d2 | |
| 2583 | #define clut_low_b d3 | |
| 2584 | #define clut_high_a d4 | |
| 2585 | #define clut_high_b d5 | |
| 2586 | ||
| 2587 | #define clut_a q1 | |
| 2588 | #define clut_b q2 | |
| 2589 | ||
| 2590 | #define texels_low d6 | |
| 2591 | #define texels_high d7 | |
| 2592 | ||
| 2593 | .align 3 | |
| 2594 | ||
| 2595 | function(texture_blocks_untextured) | |
| 2596 | bx lr | |
| 2597 | ||
| 2598 | ||
| 2599 | .align 3 | |
| 2600 | ||
| 2601 | function(texture_blocks_4bpp) | |
| 2602 | stmdb sp!, { r3 - r11, r14 } | |
| 2603 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2604 | ||
| 2605 | ldr texture_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
| 2606 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 2607 | ||
| 2608 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ] | |
| 2609 | vld1.u32 { clut_a, clut_b }, [ clut_ptr, :128 ] | |
| 2610 | ||
| 2611 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ] | |
| 2612 | vuzp.u8 clut_a, clut_b | |
| 2613 | ||
| 2614 | ldr dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ] | |
| 2615 | tst dirty_textures_mask, current_texture_mask | |
| 2616 | ||
| 2617 | bne 1f | |
| 2618 | mov c_64, #64 | |
| 2619 | ||
| 2620 | 0: | |
| 2621 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
| 2622 | ||
| 2623 | uxtah uv_0, texture_ptr, uv_01 | |
| 2624 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
| 2625 | ||
| 2626 | uxtah uv_2, texture_ptr, uv_23 | |
| 2627 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
| 2628 | ||
| 2629 | uxtah uv_4, texture_ptr, uv_45 | |
| 2630 | ldrb pixel_0, [ uv_0 ] | |
| 2631 | ||
| 2632 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
| 2633 | ldrb pixel_1, [ uv_1 ] | |
| 2634 | ||
| 2635 | uxtah uv_6, texture_ptr, uv_67 | |
| 2636 | ldrb pixel_2, [ uv_2 ] | |
| 2637 | ||
| 2638 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
| 2639 | ldrb pixel_3, [ uv_3 ] | |
| 2640 | ||
| 2641 | ldrb pixel_4, [ uv_4 ] | |
| 2642 | subs num_blocks, num_blocks, #1 | |
| 2643 | ||
| 2644 | ldrb pixel_5, [ uv_5 ] | |
| 2645 | orr pixels_a, pixel_0, pixel_1, lsl #8 | |
| 2646 | ||
| 2647 | ldrb pixel_6, [ uv_6 ] | |
| 2648 | orr pixels_b, pixel_4, pixel_5, lsl #8 | |
| 2649 | ||
| 2650 | ldrb pixel_7, [ uv_7 ] | |
| 2651 | orr pixels_a, pixels_a, pixel_2, lsl #16 | |
| 2652 | ||
| 2653 | orr pixels_b, pixels_b, pixel_6, lsl #16 | |
| 2654 | orr pixels_a, pixels_a, pixel_3, lsl #24 | |
| 2655 | ||
| 2656 | orr pixels_b, pixels_b, pixel_7, lsl #24 | |
| 2657 | vmov.u32 texels, pixels_a, pixels_b | |
| 2658 | ||
| 2659 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels | |
| 2660 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels | |
| 2661 | ||
| 2662 | vst2.u8 { texels_low, texels_high }, [ block_ptr, :128 ], c_64 | |
| 2663 | bne 0b | |
| 2664 | ||
| 2665 | ldmia sp!, { r3 - r11, pc } | |
| 2666 | ||
| 2667 | 1: | |
| 2668 | stmdb sp!, { r1 - r2 } | |
| 2669 | bl update_texture_4bpp_cache | |
| 2670 | ||
| 2671 | mov c_64, #64 | |
| 2672 | ldmia sp!, { r1 - r2 } | |
| 2673 | bal 0b | |
| 2674 | ||
| 2675 | ||
| 2676 | .align 3 | |
| 2677 | ||
| 2678 | function(texture_blocks_8bpp) | |
| 2679 | stmdb sp!, { r3 - r11, r14 } | |
| 2680 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2681 | ||
| 2682 | ldr texture_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
| 2683 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 2684 | ||
| 2685 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ] | |
| 2686 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ] | |
| 2687 | ||
| 2688 | ldr dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset ] | |
| 2689 | tst dirty_textures_mask, current_texture_mask | |
| 2690 | ||
| 2691 | bne 1f | |
| 2692 | nop | |
| 2693 | ||
| 2694 | 0: | |
| 2695 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
| 2696 | ||
| 2697 | uxtah uv_0, texture_ptr, uv_01 | |
| 2698 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
| 2699 | ||
| 2700 | uxtah uv_2, texture_ptr, uv_23 | |
| 2701 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
| 2702 | ||
| 2703 | uxtah uv_4, texture_ptr, uv_45 | |
| 2704 | ldrb pixel_0, [ uv_0 ] | |
| 2705 | ||
| 2706 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
| 2707 | ldrb pixel_1, [ uv_1 ] | |
| 2708 | ||
| 2709 | uxtah uv_6, texture_ptr, uv_67 | |
| 2710 | ldrb pixel_2, [ uv_2 ] | |
| 2711 | ||
| 2712 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
| 2713 | ldrb pixel_3, [ uv_3 ] | |
| 2714 | ||
| 2715 | ldrb pixel_4, [ uv_4 ] | |
| 2716 | add pixel_0, pixel_0, pixel_0 | |
| 2717 | ||
| 2718 | ldrb pixel_5, [ uv_5 ] | |
| 2719 | add pixel_1, pixel_1, pixel_1 | |
| 2720 | ||
| 2721 | ldrb pixel_6, [ uv_6 ] | |
| 2722 | add pixel_2, pixel_2, pixel_2 | |
| 2723 | ||
| 2724 | ldrb pixel_7, [ uv_7 ] | |
| 2725 | add pixel_3, pixel_3, pixel_3 | |
| 2726 | ||
| 2727 | ldrh pixel_0, [ clut_ptr, pixel_0 ] | |
| 2728 | add pixel_4, pixel_4, pixel_4 | |
| 2729 | ||
| 2730 | ldrh pixel_1, [ clut_ptr, pixel_1 ] | |
| 2731 | add pixel_5, pixel_5, pixel_5 | |
| 2732 | ||
| 2733 | ldrh pixel_2, [ clut_ptr, pixel_2 ] | |
| 2734 | add pixel_6, pixel_6, pixel_6 | |
| 2735 | ||
| 2736 | ldrh pixel_3, [ clut_ptr, pixel_3 ] | |
| 2737 | add pixel_7, pixel_7, pixel_7 | |
| 2738 | ||
| 2739 | ldrh pixel_4, [ clut_ptr, pixel_4 ] | |
| 2740 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
| 2741 | ||
| 2742 | ldrh pixel_5, [ clut_ptr, pixel_5 ] | |
| 2743 | orr pixels_c, pixel_2, pixel_3, lsl #16 | |
| 2744 | ||
| 2745 | ldrh pixel_6, [ clut_ptr, pixel_6 ] | |
| 2746 | subs num_blocks, num_blocks, #1 | |
| 2747 | ||
| 2748 | ldrh pixel_7, [ clut_ptr, pixel_7 ] | |
| 2749 | orr pixels_b, pixel_4, pixel_5, lsl #16 | |
| 2750 | ||
| 2751 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
| 2752 | stm block_ptr, { pixels_a, pixels_c, pixels_b, pixels_d } | |
| 2753 | ||
| 2754 | add block_ptr, block_ptr, #64 | |
| 2755 | bne 0b | |
| 2756 | ||
| 2757 | ldmia sp!, { r3 - r11, pc } | |
| 2758 | ||
| 2759 | 1: | |
| 2760 | stmdb sp!, { r1 - r2, r12 } | |
| 2761 | ||
| 2762 | bl update_texture_8bpp_cache | |
| 2763 | ||
| 2764 | ldmia sp!, { r1 - r2, r12 } | |
| 2765 | bal 0b | |
| 2766 | ||
| 2767 | ||
| 2768 | #undef uv_0 | |
| 2769 | #undef uv_1 | |
| 2770 | #undef uv_2 | |
| 2771 | #undef uv_3 | |
| 2772 | #undef uv_4 | |
| 2773 | #undef uv_5 | |
| 2774 | #undef uv_6 | |
| 2775 | #undef uv_7 | |
| 2776 | ||
| 2777 | #undef pixel_0 | |
| 2778 | #undef pixel_1 | |
| 2779 | #undef pixel_2 | |
| 2780 | #undef pixel_3 | |
| 2781 | #undef pixel_4 | |
| 2782 | #undef pixel_5 | |
| 2783 | #undef pixel_6 | |
| 2784 | #undef pixel_7 | |
| 2785 | ||
| 2786 | #undef texture_ptr | |
| 2787 | ||
| 2788 | #undef pixels_a | |
| 2789 | #undef pixels_b | |
| 2790 | #undef pixels_c | |
| 2791 | #undef pixels_d | |
| 2792 | ||
| 2793 | #define psx_gpu r0 | |
| 2794 | #define block_ptr r1 | |
| 2795 | #define num_blocks r2 | |
| 2796 | ||
| 2797 | #define uv_0 r3 | |
| 2798 | #define uv_1 r4 | |
| 2799 | #define u_0 r3 | |
| 2800 | #define u_1 r4 | |
| 2801 | #define v_0 r5 | |
| 2802 | #define v_1 r6 | |
| 2803 | ||
| 2804 | #define uv_2 r5 | |
| 2805 | #define uv_3 r6 | |
| 2806 | #define u_2 r5 | |
| 2807 | #define u_3 r6 | |
| 2808 | #define v_2 r7 | |
| 2809 | #define v_3 r8 | |
| 2810 | ||
| 2811 | #define uv_4 r7 | |
| 2812 | #define uv_5 r8 | |
| 2813 | #define u_4 r7 | |
| 2814 | #define u_5 r8 | |
| 2815 | #define v_4 r9 | |
| 2816 | #define v_5 r10 | |
| 2817 | ||
| 2818 | #define uv_6 r9 | |
| 2819 | #define uv_7 r10 | |
| 2820 | #define u_6 r9 | |
| 2821 | #define u_7 r10 | |
| 2822 | #define v_6 r11 | |
| 2823 | #define v_7 r0 | |
| 2824 | ||
| 2825 | #define pixel_0 r3 | |
| 2826 | #define pixel_1 r4 | |
| 2827 | #define pixel_2 r5 | |
| 2828 | #define pixel_3 r6 | |
| 2829 | #define pixel_4 r7 | |
| 2830 | #define pixel_5 r8 | |
| 2831 | #define pixel_6 r9 | |
| 2832 | #define pixel_7 r10 | |
| 2833 | ||
| 2834 | #define pixels_a r3 | |
| 2835 | #define pixels_b r5 | |
| 2836 | #define pixels_c r7 | |
| 2837 | #define pixels_d r9 | |
| 2838 | ||
| 2839 | #define texture_ptr r12 | |
| 2840 | ||
| 2841 | ||
| 2842 | .align 3 | |
| 2843 | ||
| 2844 | function(texture_blocks_16bpp) | |
| 2845 | stmdb sp!, { r3 - r11, r14 } | |
| 2846 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2847 | ||
| 2848 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 2849 | ldr texture_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
| 2850 | ||
| 2851 | 0: | |
| 2852 | ldrh uv_0, [ block_ptr ] | |
| 2853 | subs num_blocks, num_blocks, #1 | |
| 2854 | ||
| 2855 | ldrh uv_1, [ block_ptr, #2 ] | |
| 2856 | ||
| 2857 | and v_0, uv_0, #0xFF00 | |
| 2858 | and v_1, uv_1, #0xFF00 | |
| 2859 | ||
| 2860 | and u_0, uv_0, #0xFF | |
| 2861 | and u_1, uv_1, #0xFF | |
| 2862 | ||
| 2863 | add uv_0, u_0, v_0, lsl #2 | |
| 2864 | ldrh uv_2, [ block_ptr, #4 ] | |
| 2865 | ||
| 2866 | add uv_1, u_1, v_1, lsl #2 | |
| 2867 | ldrh uv_3, [ block_ptr, #6 ] | |
| 2868 | ||
| 2869 | add uv_0, uv_0, uv_0 | |
| 2870 | add uv_1, uv_1, uv_1 | |
| 2871 | ||
| 2872 | and v_2, uv_2, #0xFF00 | |
| 2873 | and v_3, uv_3, #0xFF00 | |
| 2874 | ||
| 2875 | and u_2, uv_2, #0xFF | |
| 2876 | and u_3, uv_3, #0xFF | |
| 2877 | ||
| 2878 | add uv_2, u_2, v_2, lsl #2 | |
| 2879 | ldrh uv_4, [ block_ptr, #8 ] | |
| 2880 | ||
| 2881 | add uv_3, u_3, v_3, lsl #2 | |
| 2882 | ldrh uv_5, [ block_ptr, #10 ] | |
| 2883 | ||
| 2884 | add uv_2, uv_2, uv_2 | |
| 2885 | add uv_3, uv_3, uv_3 | |
| 2886 | ||
| 2887 | and v_4, uv_4, #0xFF00 | |
| 2888 | and v_5, uv_5, #0xFF00 | |
| 2889 | ||
| 2890 | and u_4, uv_4, #0xFF | |
| 2891 | and u_5, uv_5, #0xFF | |
| 2892 | ||
| 2893 | add uv_4, u_4, v_4, lsl #2 | |
| 2894 | ldrh uv_6, [ block_ptr, #12 ] | |
| 2895 | ||
| 2896 | add uv_5, u_5, v_5, lsl #2 | |
| 2897 | ldrh uv_7, [ block_ptr, #14 ] | |
| 2898 | ||
| 2899 | add uv_4, uv_4, uv_4 | |
| 2900 | ldrh pixel_0, [ texture_ptr, uv_0 ] | |
| 2901 | ||
| 2902 | add uv_5, uv_5, uv_5 | |
| 2903 | ldrh pixel_1, [ texture_ptr, uv_1 ] | |
| 2904 | ||
| 2905 | and v_6, uv_6, #0xFF00 | |
| 2906 | ldrh pixel_2, [ texture_ptr, uv_2 ] | |
| 2907 | ||
| 2908 | and v_7, uv_7, #0xFF00 | |
| 2909 | ldrh pixel_3, [ texture_ptr, uv_3 ] | |
| 2910 | ||
| 2911 | and u_6, uv_6, #0xFF | |
| 2912 | ldrh pixel_4, [ texture_ptr, uv_4 ] | |
| 2913 | ||
| 2914 | and u_7, uv_7, #0xFF | |
| 2915 | ldrh pixel_5, [ texture_ptr, uv_5 ] | |
| 2916 | ||
| 2917 | add uv_6, u_6, v_6, lsl #2 | |
| 2918 | add uv_7, u_7, v_7, lsl #2 | |
| 2919 | ||
| 2920 | add uv_6, uv_6, uv_6 | |
| 2921 | add uv_7, uv_7, uv_7 | |
| 2922 | ||
| 2923 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
| 2924 | orr pixels_b, pixel_2, pixel_3, lsl #16 | |
| 2925 | ||
| 2926 | ldrh pixel_6, [ texture_ptr, uv_6 ] | |
| 2927 | orr pixels_c, pixel_4, pixel_5, lsl #16 | |
| 2928 | ||
| 2929 | ldrh pixel_7, [ texture_ptr, uv_7 ] | |
| 2930 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
| 2931 | ||
| 2932 | stm block_ptr, { pixels_a, pixels_b, pixels_c, pixels_d } | |
| 2933 | add block_ptr, block_ptr, #64 | |
| 2934 | ||
| 2935 | bne 0b | |
| 2936 | ||
| 2937 | ldmia sp!, { r3 - r11, pc } | |
| 2938 | ||
| 2939 | ||
| 2940 | #undef num_blocks | |
| 2941 | ||
| 2942 | #undef test_mask | |
| 2943 | #undef texels | |
| 2944 | #undef pixels_b | |
| 2945 | #undef pixels | |
| 2946 | #undef d64_1 | |
| 2947 | #undef d64_4 | |
| 2948 | #undef d64_128 | |
| 2949 | #undef draw_mask | |
| 2950 | #undef msb_mask | |
| 2951 | #undef msb_mask_low | |
| 2952 | #undef msb_mask_high | |
| 2953 | #undef fb_pixels | |
| 2954 | ||
| 2955 | #undef c_32 | |
| 2956 | #undef fb_ptr | |
| 2957 | #undef mask_msb_ptr | |
| 2958 | ||
| 2959 | #define psx_gpu r0 | |
| 2960 | #define num_blocks r1 | |
| 2961 | #define color_ptr r2 | |
| 2962 | #define mask_msb_ptr r2 | |
| 2963 | ||
| 2964 | #define block_ptr_load_a r0 | |
| 2965 | #define block_ptr_store r3 | |
| 2966 | #define block_ptr_load_b r12 | |
| 2967 | #define c_32 r2 | |
| 2968 | ||
| 2969 | #define c_48 r4 | |
| 2970 | #define fb_ptr r14 | |
| 2971 | #define draw_mask_bits_scalar r5 | |
| 2972 | ||
| 2973 | #define d128_0x07 q0 | |
| 2974 | #define d128_0x1F q1 | |
| 2975 | #define d128_0x8000 q2 | |
| 2976 | #define test_mask q3 | |
| 2977 | #define texels q4 | |
| 2978 | #define colors_rg q5 | |
| 2979 | #define colors_b_dm_bits q6 | |
| 2980 | #define texels_rg q7 | |
| 2981 | #define pixels_r q8 | |
| 2982 | #define pixels_g q9 | |
| 2983 | #define pixels_b q10 | |
| 2984 | #define pixels q11 | |
| 2985 | #define zero_mask q4 | |
| 2986 | #define draw_mask q12 | |
| 2987 | #define msb_mask q13 | |
| 2988 | ||
| 2989 | #define fb_pixels q8 | |
| 2990 | ||
| 2991 | #define pixels_gb_low q9 | |
| 2992 | ||
| 2993 | #define colors_r d10 | |
| 2994 | #define colors_g d11 | |
| 2995 | #define colors_b d12 | |
| 2996 | #define draw_mask_bits d13 | |
| 2997 | #define texels_r d14 | |
| 2998 | #define texels_g d15 | |
| 2999 | #define pixels_r_low d16 | |
| 3000 | #define pixels_g_low d18 | |
| 3001 | #define pixels_b_low d19 | |
| 3002 | #define msb_mask_low d26 | |
| 3003 | #define msb_mask_high d27 | |
| 3004 | ||
| 3005 | #define d64_1 d28 | |
| 3006 | #define d64_4 d29 | |
| 3007 | #define d64_128 d30 | |
| 3008 | #define texels_b d31 | |
| 3009 | ||
| 3010 | #define shade_blocks_textured_modulated_prologue_indirect() \ | |
| 3011 | mov c_48, #48; \ | |
| 3012 | add block_ptr_store, psx_gpu, #psx_gpu_blocks_offset \ | |
| 3013 | ||
| 3014 | #define shade_blocks_textured_modulated_prologue_direct() \ | |
| 3015 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 3016 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] \ | |
| 3017 | ||
| 3018 | #define shade_blocks_textured_modulated_prologue_shaded() \ | |
| 3019 | ||
| 3020 | #define shade_blocks_textured_modulated_prologue_unshaded() \ | |
| 3021 | add color_ptr, psx_gpu, #psx_gpu_triangle_color_offset; \ | |
| 3022 | vld1.u32 { colors_r[] }, [ color_ptr, :32 ]; \ | |
| 3023 | vdup.u8 colors_g, colors_r[1]; \ | |
| 3024 | vdup.u8 colors_b, colors_r[2]; \ | |
| 3025 | vdup.u8 colors_r, colors_r[0] \ | |
| 3026 | ||
| 3027 | ||
| 3028 | #define shade_blocks_textured_modulated_load_dithered(target) \ | |
| 3029 | vld1.u32 { target }, [ block_ptr_load_b, :128 ] \ | |
| 3030 | ||
| 3031 | #define shade_blocks_textured_modulated_load_last_dithered(target) \ | |
| 3032 | vld1.u32 { target }, [ block_ptr_load_b, :128 ], c_32 \ | |
| 3033 | ||
| 3034 | #define shade_blocks_textured_modulated_load_undithered(target) \ | |
| 3035 | ||
| 3036 | #define shade_blocks_textured_modulated_load_last_undithered(target) \ | |
| 3037 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
| 3038 | ||
| 3039 | #define shade_blocks_textured_modulate_dithered(channel) \ | |
| 3040 | vmlal.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
| 3041 | ||
| 3042 | #define shade_blocks_textured_modulate_undithered(channel) \ | |
| 3043 | vmull.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
| 3044 | ||
| 3045 | ||
| 3046 | #define shade_blocks_textured_modulated_store_draw_mask_indirect(offset) \ | |
| 3047 | vst1.u32 { draw_mask }, [ block_ptr_store, :128 ]! \ | |
| 3048 | ||
| 3049 | #define shade_blocks_textured_modulated_store_draw_mask_direct(offset) \ | |
| 3050 | ldr fb_ptr, [ block_ptr_load_b, #(offset - 64) ]; \ | |
| 3051 | vld1.u32 { fb_pixels }, [ fb_ptr ]; \ | |
| 3052 | vbit.u16 pixels, fb_pixels, draw_mask \ | |
| 3053 | ||
| 3054 | #define shade_blocks_textured_modulated_store_pixels_indirect() \ | |
| 3055 | vst1.u32 { pixels }, [ block_ptr_store, :128 ], c_48 \ | |
| 3056 | ||
| 3057 | #define shade_blocks_textured_modulated_store_pixels_direct() \ | |
| 3058 | vst1.u32 { pixels }, [ fb_ptr ] \ | |
| 3059 | ||
| 3060 | ||
| 3061 | #define shade_blocks_textured_modulated_load_rg_shaded() \ | |
| 3062 | vld1.u32 { colors_r, colors_g }, [ block_ptr_load_b, :128 ], c_32 \ | |
| 3063 | ||
| 3064 | #define shade_blocks_textured_modulated_load_rg_unshaded() \ | |
| 3065 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
| 3066 | ||
| 3067 | #define shade_blocks_textured_modulated_load_bdm_shaded() \ | |
| 3068 | vld1.u32 { colors_b, draw_mask_bits }, [ block_ptr_load_a, :128 ], c_32 \ | |
| 3069 | ||
| 3070 | #define shade_blocks_textured_modulated_load_bdm_unshaded() \ | |
| 3071 | ldr draw_mask_bits_scalar, [ block_ptr_load_a, #8 ]; \ | |
| 3072 | add block_ptr_load_a, block_ptr_load_a, #32 \ | |
| 3073 | ||
| 3074 | #define shade_blocks_textured_modulated_expand_draw_mask_shaded() \ | |
| 3075 | vdup.u16 draw_mask, draw_mask_bits[0] \ | |
| 3076 | ||
| 3077 | #define shade_blocks_textured_modulated_expand_draw_mask_unshaded() \ | |
| 3078 | vdup.u16 draw_mask, draw_mask_bits_scalar \ | |
| 3079 | ||
| 3080 | ||
| 3081 | #define shade_blocks_textured_modulated_apply_msb_mask_indirect() \ | |
| 3082 | ||
| 3083 | #define shade_blocks_textured_modulated_apply_msb_mask_direct() \ | |
| 3084 | vorr.u16 pixels, pixels, msb_mask \ | |
| 3085 | ||
| 3086 | ||
| 3087 | #define shade_blocks_textured_modulated_builder(shading, dithering, target) \ | |
| 3088 | .align 3; \ | |
| 3089 | \ | |
| 3090 | function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ | |
| 3091 | stmdb sp!, { r4 - r5, lr }; \ | |
| 3092 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 3093 | \ | |
| 3094 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
| 3095 | \ | |
| 3096 | shade_blocks_textured_modulated_prologue_##target(); \ | |
| 3097 | shade_blocks_textured_modulated_prologue_##shading(); \ | |
| 3098 | \ | |
| 3099 | add block_ptr_load_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3100 | mov c_32, #32; \ | |
| 3101 | \ | |
| 3102 | add block_ptr_load_b, block_ptr_load_a, #16; \ | |
| 3103 | vmov.u8 d64_1, #1; \ | |
| 3104 | vmov.u8 d64_4, #4; \ | |
| 3105 | vmov.u8 d64_128, #128; \ | |
| 3106 | \ | |
| 3107 | vld1.u32 { texels }, [ block_ptr_load_a, :128 ], c_32; \ | |
| 3108 | vmov.u8 d128_0x07, #0x07; \ | |
| 3109 | \ | |
| 3110 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
| 3111 | vmov.u8 d128_0x1F, #0x1F; \ | |
| 3112 | \ | |
| 3113 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
| 3114 | vmov.u16 d128_0x8000, #0x8000; \ | |
| 3115 | \ | |
| 3116 | vmovn.u16 texels_r, texels; \ | |
| 3117 | vshrn.u16 texels_g, texels, #5; \ | |
| 3118 | \ | |
| 3119 | vshrn.u16 texels_b, texels, #7; \ | |
| 3120 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
| 3121 | \ | |
| 3122 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
| 3123 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 3124 | \ | |
| 3125 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
| 3126 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
| 3127 | \ | |
| 3128 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
| 3129 | vshr.u8 texels_b, texels_b, #3; \ | |
| 3130 | \ | |
| 3131 | shade_blocks_textured_modulate_##dithering(r); \ | |
| 3132 | shade_blocks_textured_modulate_##dithering(g); \ | |
| 3133 | shade_blocks_textured_modulate_##dithering(b); \ | |
| 3134 | \ | |
| 3135 | vand.u16 pixels, texels, d128_0x8000; \ | |
| 3136 | vceq.u16 zero_mask, texels, #0; \ | |
| 3137 | \ | |
| 3138 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
| 3139 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
| 3140 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
| 3141 | \ | |
| 3142 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
| 3143 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
| 3144 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
| 3145 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
| 3146 | \ | |
| 3147 | subs num_blocks, num_blocks, #1; \ | |
| 3148 | beq 1f; \ | |
| 3149 | \ | |
| 3150 | .align 3; \ | |
| 3151 | \ | |
| 3152 | 0: \ | |
| 3153 | vld1.u32 { texels }, [ block_ptr_load_a, :128 ], c_32; \ | |
| 3154 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
| 3155 | vshrn.u16 texels_g, texels, #5; \ | |
| 3156 | \ | |
| 3157 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
| 3158 | vshrn.u16 texels_b, texels, #7; \ | |
| 3159 | \ | |
| 3160 | vmovn.u16 texels_r, texels; \ | |
| 3161 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
| 3162 | \ | |
| 3163 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
| 3164 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
| 3165 | shade_blocks_textured_modulated_store_draw_mask_##target(-4); \ | |
| 3166 | \ | |
| 3167 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
| 3168 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
| 3169 | \ | |
| 3170 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
| 3171 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
| 3172 | \ | |
| 3173 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
| 3174 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 3175 | \ | |
| 3176 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
| 3177 | vshr.u8 texels_b, texels_b, #3; \ | |
| 3178 | \ | |
| 3179 | shade_blocks_textured_modulate_##dithering(r); \ | |
| 3180 | shade_blocks_textured_modulate_##dithering(g); \ | |
| 3181 | shade_blocks_textured_modulate_##dithering(b); \ | |
| 3182 | \ | |
| 3183 | vand.u16 pixels, texels, d128_0x8000; \ | |
| 3184 | vceq.u16 zero_mask, texels, #0; \ | |
| 3185 | \ | |
| 3186 | subs num_blocks, num_blocks, #1; \ | |
| 3187 | \ | |
| 3188 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
| 3189 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
| 3190 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
| 3191 | \ | |
| 3192 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
| 3193 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
| 3194 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
| 3195 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
| 3196 | \ | |
| 3197 | bne 0b; \ | |
| 3198 | \ | |
| 3199 | 1: \ | |
| 3200 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
| 3201 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
| 3202 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
| 3203 | \ | |
| 3204 | shade_blocks_textured_modulated_store_draw_mask_##target(28); \ | |
| 3205 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
| 3206 | \ | |
| 3207 | ldmia sp!, { r4 - r5, pc } \ | |
| 3208 | ||
| 3209 | ||
| 3210 | shade_blocks_textured_modulated_builder(shaded, dithered, direct); | |
| 3211 | shade_blocks_textured_modulated_builder(shaded, undithered, direct); | |
| 3212 | shade_blocks_textured_modulated_builder(unshaded, dithered, direct); | |
| 3213 | shade_blocks_textured_modulated_builder(unshaded, undithered, direct); | |
| 3214 | ||
| 3215 | shade_blocks_textured_modulated_builder(shaded, dithered, indirect); | |
| 3216 | shade_blocks_textured_modulated_builder(shaded, undithered, indirect); | |
| 3217 | shade_blocks_textured_modulated_builder(unshaded, dithered, indirect); | |
| 3218 | shade_blocks_textured_modulated_builder(unshaded, undithered, indirect); | |
| 3219 | ||
| 3220 | ||
| 3221 | #undef c_64 | |
| 3222 | #undef fb_ptr | |
| 3223 | #undef color_ptr | |
| 3224 | ||
| 3225 | #undef color_r | |
| 3226 | #undef color_g | |
| 3227 | #undef color_b | |
| 3228 | ||
| 3229 | #undef test_mask | |
| 3230 | #undef pixels | |
| 3231 | #undef draw_mask | |
| 3232 | #undef zero_mask | |
| 3233 | #undef fb_pixels | |
| 3234 | #undef msb_mask | |
| 3235 | #undef msb_mask_low | |
| 3236 | #undef msb_mask_high | |
| 3237 | ||
| 3238 | #define psx_gpu r0 | |
| 3239 | #define num_blocks r1 | |
| 3240 | #define mask_msb_ptr r2 | |
| 3241 | #define color_ptr r3 | |
| 3242 | ||
| 3243 | #define block_ptr_load r0 | |
| 3244 | #define draw_mask_store_ptr r3 | |
| 3245 | #define draw_mask_bits_ptr r12 | |
| 3246 | #define draw_mask_ptr r12 | |
| 3247 | #define pixel_store_ptr r14 | |
| 3248 | ||
| 3249 | #define fb_ptr_cmp r4 | |
| 3250 | ||
| 3251 | #define fb_ptr r3 | |
| 3252 | #define fb_ptr_next r14 | |
| 3253 | ||
| 3254 | #define c_64 r2 | |
| 3255 | ||
| 3256 | #define test_mask q0 | |
| 3257 | #define pixels q1 | |
| 3258 | #define draw_mask q2 | |
| 3259 | #define zero_mask q3 | |
| 3260 | #define draw_mask_combined q4 | |
| 3261 | #define fb_pixels q5 | |
| 3262 | #define fb_pixels_next q6 | |
| 3263 | #define msb_mask q7 | |
| 3264 | ||
| 3265 | #define draw_mask_low d4 | |
| 3266 | #define draw_mask_high d5 | |
| 3267 | #define msb_mask_low d14 | |
| 3268 | #define msb_mask_high d15 | |
| 3269 | ||
| 3270 | .align 3 | |
| 3271 | function(shade_blocks_textured_unmodulated_indirect) | |
| 3272 | str r14, [ sp, #-4 ] | |
| 3273 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) | |
| 3274 | ||
| 3275 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 3276 | add pixel_store_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
| 3277 | ||
| 3278 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
| 3279 | add draw_mask_store_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 3280 | ||
| 3281 | mov c_64, #64 | |
| 3282 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
| 3283 | ||
| 3284 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
| 3285 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| 3286 | [ draw_mask_bits_ptr, :16 ], c_64 | |
| 3287 | vceq.u16 zero_mask, pixels, #0 | |
| 3288 | ||
| 3289 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3290 | vst1.u32 { pixels }, [ pixel_store_ptr, :128 ], c_64 | |
| 3291 | ||
| 3292 | subs num_blocks, num_blocks, #1 | |
| 3293 | beq 1f | |
| 3294 | ||
| 3295 | 0: | |
| 3296 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
| 3297 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3298 | ||
| 3299 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| 3300 | [ draw_mask_bits_ptr, :16 ], c_64 | |
| 3301 | vceq.u16 zero_mask, pixels, #0 | |
| 3302 | ||
| 3303 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3304 | vst1.u32 { pixels }, [ pixel_store_ptr, :128 ], c_64 | |
| 3305 | ||
| 3306 | vst1.u32 { draw_mask_combined }, [ draw_mask_store_ptr, :128 ], c_64 | |
| 3307 | subs num_blocks, num_blocks, #1 | |
| 3308 | ||
| 3309 | bne 0b | |
| 3310 | ||
| 3311 | 1: | |
| 3312 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3313 | vst1.u32 { draw_mask_combined }, [ draw_mask_store_ptr, :128 ], c_64 | |
| 3314 | ||
| 3315 | ldr pc, [ sp, #-4 ] | |
| 3316 | ||
| 3317 | ||
| 3318 | .align 3 | |
| 3319 | ||
| 3320 | function(shade_blocks_textured_unmodulated_direct) | |
| 3321 | stmdb sp!, { r4, r14 } | |
| 3322 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) | |
| 3323 | ||
| 3324 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 3325 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset | |
| 3326 | ||
| 3327 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] | |
| 3328 | mov c_64, #64 | |
| 3329 | ||
| 3330 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
| 3331 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
| 3332 | ||
| 3333 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| 3334 | [ draw_mask_bits_ptr, :16 ], c_64 | |
| 3335 | ldr fb_ptr_next, [ block_ptr_load, #44 ] | |
| 3336 | ||
| 3337 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
| 3338 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3339 | vceq.u16 zero_mask, pixels, #0 | |
| 3340 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3341 | ||
| 3342 | subs num_blocks, num_blocks, #1 | |
| 3343 | beq 1f | |
| 3344 | ||
| 3345 | 0: | |
| 3346 | mov fb_ptr, fb_ptr_next | |
| 3347 | ldr fb_ptr_next, [ block_ptr_load, #44 ] | |
| 3348 | ||
| 3349 | vorr.u16 pixels, pixels, msb_mask | |
| 3350 | ||
| 3351 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3352 | vmov fb_pixels, fb_pixels_next | |
| 3353 | ||
| 3354 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| 3355 | [ draw_mask_bits_ptr, :16 ], c_64 | |
| 3356 | vbif.u16 fb_pixels, pixels, draw_mask_combined | |
| 3357 | ||
| 3358 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
| 3359 | ||
| 3360 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
| 3361 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
| 3362 | cmp fb_ptr_cmp, #28 | |
| 3363 | bls 4f | |
| 3364 | ||
| 3365 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3366 | vceq.u16 zero_mask, pixels, #0 | |
| 3367 | ||
| 3368 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
| 3369 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3370 | ||
| 3371 | 3: | |
| 3372 | subs num_blocks, num_blocks, #1 | |
| 3373 | bne 0b | |
| 3374 | ||
| 3375 | 1: | |
| 3376 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3377 | vbif.u16 fb_pixels_next, pixels, draw_mask_combined | |
| 3378 | ||
| 3379 | vst1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3380 | ||
| 3381 | ldmia sp!, { r4, pc } | |
| 3382 | ||
| 3383 | 4: | |
| 3384 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
| 3385 | vceq.u16 zero_mask, pixels, #0 | |
| 3386 | ||
| 3387 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3388 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3389 | ||
| 3390 | bal 3b | |
| 3391 | ||
| 3392 | ||
| 3393 | function(shade_blocks_unshaded_untextured_indirect) | |
| 3394 | bx lr | |
| 3395 | ||
| 3396 | .align 3 | |
| 3397 | ||
| 3398 | function(shade_blocks_unshaded_untextured_direct) | |
| 3399 | stmdb sp!, { r4, r14 } | |
| 3400 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 3401 | ||
| 3402 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 3403 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset | |
| 3404 | ||
| 3405 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] | |
| 3406 | add color_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
| 3407 | ||
| 3408 | add block_ptr_load, psx_gpu, #(psx_gpu_blocks_offset + 44) | |
| 3409 | vld1.u16 { pixels }, [ color_ptr, :128 ] | |
| 3410 | ||
| 3411 | mov c_64, #64 | |
| 3412 | vld1.u16 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
| 3413 | ||
| 3414 | vorr.u16 pixels, pixels, msb_mask | |
| 3415 | subs num_blocks, num_blocks, #1 | |
| 3416 | ||
| 3417 | ldr fb_ptr_next, [ block_ptr_load ], #64 | |
| 3418 | ||
| 3419 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3420 | beq 1f | |
| 3421 | ||
| 3422 | 0: | |
| 3423 | vmov fb_pixels, fb_pixels_next | |
| 3424 | mov fb_ptr, fb_ptr_next | |
| 3425 | ldr fb_ptr_next, [ block_ptr_load ], #64 | |
| 3426 | ||
| 3427 | vbif.u16 fb_pixels, pixels, draw_mask | |
| 3428 | vld1.u16 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
| 3429 | ||
| 3430 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
| 3431 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
| 3432 | cmp fb_ptr_cmp, #28 | |
| 3433 | bls 4f | |
| 3434 | ||
| 3435 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3436 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
| 3437 | ||
| 3438 | 3: | |
| 3439 | subs num_blocks, num_blocks, #1 | |
| 3440 | bne 0b | |
| 3441 | ||
| 3442 | 1: | |
| 3443 | vbif.u16 fb_pixels_next, pixels, draw_mask | |
| 3444 | vst1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3445 | ||
| 3446 | ldmia sp!, { r4, pc } | |
| 3447 | ||
| 3448 | 4: | |
| 3449 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
| 3450 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
| 3451 | bal 3b | |
| 3452 | ||
| 3453 | ||
| 3454 | #undef draw_mask_ptr | |
| 3455 | #undef c_64 | |
| 3456 | #undef fb_ptr | |
| 3457 | #undef fb_ptr_next | |
| 3458 | #undef fb_ptr_cmp | |
| 3459 | ||
| 3460 | #define psx_gpu r0 | |
| 3461 | #define num_blocks r1 | |
| 3462 | #define msb_mask_ptr r2 | |
| 3463 | #define pixel_ptr r3 | |
| 3464 | #define draw_mask_ptr r0 | |
| 3465 | #define c_64 r2 | |
| 3466 | #define fb_ptr r12 | |
| 3467 | #define fb_ptr_next r14 | |
| 3468 | #define fb_ptr_cmp r4 | |
| 3469 | ||
| 3470 | #undef msb_mask | |
| 3471 | #undef draw_mask | |
| 3472 | #undef pixels | |
| 3473 | #undef fb_pixels | |
| 3474 | #undef d128_0x8000 | |
| 3475 | #undef msb_mask_low | |
| 3476 | #undef msb_mask_high | |
| 3477 | #undef draw_mask_next | |
| 3478 | #undef pixels_g | |
| 3479 | #undef blend_pixels | |
| 3480 | #undef fb_pixels_next | |
| 3481 | ||
| 3482 | #define msb_mask q0 | |
| 3483 | #define draw_mask q1 | |
| 3484 | #define pixels q2 | |
| 3485 | #define fb_pixels q3 | |
| 3486 | #define blend_pixels q4 | |
| 3487 | #define pixels_no_msb q5 | |
| 3488 | #define blend_mask q6 | |
| 3489 | #define fb_pixels_no_msb q7 | |
| 3490 | #define d128_0x8000 q8 | |
| 3491 | #define d128_0x0421 q9 | |
| 3492 | #define fb_pixels_next q10 | |
| 3493 | #define blend_pixels_next q11 | |
| 3494 | #define pixels_next q12 | |
| 3495 | #define draw_mask_next q13 | |
| 3496 | #define write_mask q14 | |
| 3497 | ||
| 3498 | #define pixels_rb q5 | |
| 3499 | #define pixels_mg q7 | |
| 3500 | #define pixels_g q7 | |
| 3501 | #define d128_0x7C1F q8 | |
| 3502 | #define d128_0x03E0 q9 | |
| 3503 | #define fb_pixels_rb q10 | |
| 3504 | #define fb_pixels_g q11 | |
| 3505 | #define fb_pixels_masked q11 | |
| 3506 | #define d128_0x83E0 q15 | |
| 3507 | #define pixels_fourth q7 | |
| 3508 | #define d128_0x1C07 q12 | |
| 3509 | #define d128_0x00E0 q13 | |
| 3510 | #define d128_0x80E0 q13 | |
| 3511 | ||
| 3512 | #define msb_mask_low d0 | |
| 3513 | #define msb_mask_high d1 | |
| 3514 | ||
| 3515 | #define blend_blocks_average_set_blend_mask_textured(source) \ | |
| 3516 | vclt.s16 blend_mask, source, #0 \ | |
| 3517 | ||
| 3518 | #define blend_blocks_average_set_stp_bit_textured() \ | |
| 3519 | vorr.u16 blend_pixels, #0x8000 \ | |
| 3520 | ||
| 3521 | #define blend_blocks_average_combine_textured(source) \ | |
| 3522 | vbif.u16 blend_pixels, source, blend_mask \ | |
| 3523 | ||
| 3524 | #define blend_blocks_average_set_blend_mask_untextured(source) \ | |
| 3525 | ||
| 3526 | #define blend_blocks_average_set_stp_bit_untextured() \ | |
| 3527 | ||
| 3528 | #define blend_blocks_average_combine_untextured(source) \ | |
| 3529 | ||
| 3530 | #define blend_blocks_average_mask_set_on() \ | |
| 3531 | vclt.s16 write_mask, fb_pixels_next, #0 \ | |
| 3532 | ||
| 3533 | #define blend_blocks_average_mask_copy_on() \ | |
| 3534 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
| 3535 | ||
| 3536 | #define blend_blocks_average_mask_copy_b_on() \ | |
| 3537 | vorr.u16 draw_mask_next, draw_mask_next, write_mask \ | |
| 3538 | ||
| 3539 | #define blend_blocks_average_mask_set_off() \ | |
| 3540 | ||
| 3541 | #define blend_blocks_average_mask_copy_off() \ | |
| 3542 | vmov draw_mask, draw_mask_next \ | |
| 3543 | ||
| 3544 | #define blend_blocks_average_mask_copy_b_off() \ | |
| 3545 | ||
| 3546 | #define blend_blocks_average_builder(texturing, mask_evaluate) \ | |
| 3547 | .align 3; \ | |
| 3548 | \ | |
| 3549 | function(blend_blocks_##texturing##_average_##mask_evaluate) \ | |
| 3550 | stmdb sp!, { r4, r14 }; \ | |
| 3551 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 3552 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 3553 | \ | |
| 3554 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| 3555 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 3556 | \ | |
| 3557 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3558 | mov c_64, #64; \ | |
| 3559 | \ | |
| 3560 | vmov.u16 d128_0x8000, #0x8000; \ | |
| 3561 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3562 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3563 | \ | |
| 3564 | vmov.u16 d128_0x0421, #0x0400; \ | |
| 3565 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
| 3566 | \ | |
| 3567 | vorr.u16 d128_0x0421, #0x0021; \ | |
| 3568 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
| 3569 | \ | |
| 3570 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
| 3571 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3572 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3573 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3574 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
| 3575 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3576 | \ | |
| 3577 | subs num_blocks, num_blocks, #1; \ | |
| 3578 | beq 1f; \ | |
| 3579 | \ | |
| 3580 | 0: \ | |
| 3581 | mov fb_ptr, fb_ptr_next; \ | |
| 3582 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3583 | \ | |
| 3584 | vmov pixels, pixels_next; \ | |
| 3585 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
| 3586 | \ | |
| 3587 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
| 3588 | \ | |
| 3589 | blend_blocks_average_mask_copy_##mask_evaluate(); \ | |
| 3590 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3591 | \ | |
| 3592 | blend_blocks_average_set_blend_mask_##texturing(pixels); \ | |
| 3593 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
| 3594 | vmov fb_pixels, fb_pixels_next; \ | |
| 3595 | blend_blocks_average_combine_##texturing(pixels); \ | |
| 3596 | \ | |
| 3597 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3598 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3599 | cmp fb_ptr_cmp, #28; \ | |
| 3600 | bls 2f; \ | |
| 3601 | \ | |
| 3602 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
| 3603 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
| 3604 | \ | |
| 3605 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3606 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3607 | \ | |
| 3608 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3609 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
| 3610 | \ | |
| 3611 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3612 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3613 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
| 3614 | vst1.u16 { fb_pixels }, [ fb_ptr ]; \ | |
| 3615 | \ | |
| 3616 | 3: \ | |
| 3617 | subs num_blocks, num_blocks, #1; \ | |
| 3618 | bne 0b; \ | |
| 3619 | \ | |
| 3620 | 1: \ | |
| 3621 | blend_blocks_average_mask_copy_b_##mask_evaluate(); \ | |
| 3622 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
| 3623 | \ | |
| 3624 | blend_blocks_average_set_blend_mask_##texturing(pixels_next); \ | |
| 3625 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
| 3626 | blend_blocks_average_combine_##texturing(pixels_next); \ | |
| 3627 | \ | |
| 3628 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3629 | vbif.u16 fb_pixels_next, blend_pixels, draw_mask_next; \ | |
| 3630 | vst1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
| 3631 | \ | |
| 3632 | ldmia sp!, { r4, pc }; \ | |
| 3633 | \ | |
| 3634 | 2: \ | |
| 3635 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3636 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
| 3637 | vst1.u16 { fb_pixels }, [ fb_ptr ]; \ | |
| 3638 | \ | |
| 3639 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
| 3640 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
| 3641 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3642 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3643 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3644 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3645 | \ | |
| 3646 | bal 3b \ | |
| 3647 | ||
| 3648 | blend_blocks_average_builder(textured, off) | |
| 3649 | blend_blocks_average_builder(untextured, off) | |
| 3650 | blend_blocks_average_builder(textured, on) | |
| 3651 | blend_blocks_average_builder(untextured, on) | |
| 3652 | ||
| 3653 | ||
| 3654 | #define blend_blocks_add_mask_set_on() \ | |
| 3655 | vclt.s16 write_mask, fb_pixels, #0 \ | |
| 3656 | ||
| 3657 | #define blend_blocks_add_mask_copy_on() \ | |
| 3658 | vorr.u16 draw_mask, draw_mask, write_mask \ | |
| 3659 | ||
| 3660 | #define blend_blocks_add_mask_set_off() \ | |
| 3661 | ||
| 3662 | #define blend_blocks_add_mask_copy_off() \ | |
| 3663 | ||
| 3664 | ||
| 3665 | #define blend_blocks_add_textured_builder(mask_evaluate) \ | |
| 3666 | .align 3; \ | |
| 3667 | \ | |
| 3668 | function(blend_blocks_textured_add_##mask_evaluate) \ | |
| 3669 | stmdb sp!, { r4, r14 }; \ | |
| 3670 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 3671 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 3672 | \ | |
| 3673 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| 3674 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 3675 | \ | |
| 3676 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3677 | mov c_64, #64; \ | |
| 3678 | \ | |
| 3679 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3680 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3681 | vmov.u16 d128_0x83E0, #0x8000; \ | |
| 3682 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3683 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3684 | vorr.u16 d128_0x83E0, d128_0x83E0, d128_0x03E0; \ | |
| 3685 | \ | |
| 3686 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3687 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3688 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 3689 | vclt.s16 blend_mask, pixels, #0; \ | |
| 3690 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3691 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 3692 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3693 | \ | |
| 3694 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3695 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 3696 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3697 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
| 3698 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3699 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 3700 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3701 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 3702 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3703 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 3704 | \ | |
| 3705 | subs num_blocks, num_blocks, #1; \ | |
| 3706 | beq 1f; \ | |
| 3707 | \ | |
| 3708 | 0: \ | |
| 3709 | mov fb_ptr, fb_ptr_next; \ | |
| 3710 | \ | |
| 3711 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3712 | \ | |
| 3713 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 3714 | vclt.s16 blend_mask, pixels, #0; \ | |
| 3715 | \ | |
| 3716 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 3717 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3718 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
| 3719 | \ | |
| 3720 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 3721 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3722 | \ | |
| 3723 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3724 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3725 | cmp fb_ptr_cmp, #28; \ | |
| 3726 | bls 2f; \ | |
| 3727 | \ | |
| 3728 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3729 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 3730 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3731 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3732 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3733 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3734 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 3735 | \ | |
| 3736 | 3: \ | |
| 3737 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 3738 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3739 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 3740 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3741 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 3742 | \ | |
| 3743 | subs num_blocks, num_blocks, #1; \ | |
| 3744 | bne 0b; \ | |
| 3745 | \ | |
| 3746 | 1: \ | |
| 3747 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3748 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 3749 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
| 3750 | \ | |
| 3751 | ldmia sp!, { r4, pc }; \ | |
| 3752 | \ | |
| 3753 | 2: \ | |
| 3754 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 3755 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3756 | \ | |
| 3757 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3758 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 3759 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3760 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3761 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3762 | bal 3b \ | |
| 3763 | ||
| 3764 | ||
| 3765 | #define blend_blocks_add_untextured_builder(mask_evaluate) \ | |
| 3766 | .align 3; \ | |
| 3767 | \ | |
| 3768 | function(blend_blocks_untextured_add_##mask_evaluate) \ | |
| 3769 | stmdb sp!, { r4, r14 }; \ | |
| 3770 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 3771 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 3772 | \ | |
| 3773 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| 3774 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 3775 | \ | |
| 3776 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3777 | mov c_64, #64; \ | |
| 3778 | \ | |
| 3779 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3780 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3781 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3782 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3783 | \ | |
| 3784 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3785 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3786 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 3787 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3788 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 3789 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3790 | \ | |
| 3791 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3792 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
| 3793 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3794 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3795 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3796 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3797 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3798 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 3799 | \ | |
| 3800 | subs num_blocks, num_blocks, #1; \ | |
| 3801 | beq 1f; \ | |
| 3802 | \ | |
| 3803 | 0: \ | |
| 3804 | mov fb_ptr, fb_ptr_next; \ | |
| 3805 | \ | |
| 3806 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3807 | \ | |
| 3808 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 3809 | \ | |
| 3810 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3811 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3812 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
| 3813 | \ | |
| 3814 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 3815 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3816 | \ | |
| 3817 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3818 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3819 | cmp fb_ptr_cmp, #28; \ | |
| 3820 | bls 2f; \ | |
| 3821 | \ | |
| 3822 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3823 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 3824 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3825 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3826 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3827 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 3828 | \ | |
| 3829 | 3: \ | |
| 3830 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3831 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3832 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3833 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3834 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 3835 | \ | |
| 3836 | subs num_blocks, num_blocks, #1; \ | |
| 3837 | bne 0b; \ | |
| 3838 | \ | |
| 3839 | 1: \ | |
| 3840 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3841 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3842 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 3843 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
| 3844 | \ | |
| 3845 | ldmia sp!, { r4, pc }; \ | |
| 3846 | \ | |
| 3847 | 2: \ | |
| 3848 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 3849 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3850 | \ | |
| 3851 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3852 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 3853 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3854 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3855 | bal 3b \ | |
| 3856 | ||
| 3857 | ||
| 3858 | blend_blocks_add_textured_builder(off) | |
| 3859 | blend_blocks_add_textured_builder(on) | |
| 3860 | blend_blocks_add_untextured_builder(off) | |
| 3861 | blend_blocks_add_untextured_builder(on) | |
| 3862 | ||
| 3863 | #define blend_blocks_subtract_set_blend_mask_textured() \ | |
| 3864 | vclt.s16 blend_mask, pixels_next, #0 \ | |
| 3865 | ||
| 3866 | #define blend_blocks_subtract_combine_textured() \ | |
| 3867 | vbif.u16 blend_pixels, pixels, blend_mask \ | |
| 3868 | ||
| 3869 | #define blend_blocks_subtract_set_stb_textured() \ | |
| 3870 | vorr.u16 blend_pixels, #0x8000 \ | |
| 3871 | ||
| 3872 | #define blend_blocks_subtract_msb_mask_textured() \ | |
| 3873 | vorr.u16 pixels, pixels_next, msb_mask \ | |
| 3874 | ||
| 3875 | #define blend_blocks_subtract_set_blend_mask_untextured() \ | |
| 3876 | ||
| 3877 | #define blend_blocks_subtract_combine_untextured() \ | |
| 3878 | ||
| 3879 | #define blend_blocks_subtract_set_stb_untextured() \ | |
| 3880 | vorr.u16 blend_pixels, blend_pixels, msb_mask \ | |
| 3881 | ||
| 3882 | #define blend_blocks_subtract_msb_mask_untextured() \ | |
| 3883 | ||
| 3884 | ||
| 3885 | #define blend_blocks_subtract_mask_set_on() \ | |
| 3886 | vclt.s16 write_mask, fb_pixels, #0 \ | |
| 3887 | ||
| 3888 | #define blend_blocks_subtract_mask_copy_on() \ | |
| 3889 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
| 3890 | ||
| 3891 | #define blend_blocks_subtract_mask_set_off() \ | |
| 3892 | ||
| 3893 | #define blend_blocks_subtract_mask_copy_off() \ | |
| 3894 | vmov draw_mask, draw_mask_next \ | |
| 3895 | ||
| 3896 | ||
| 3897 | #define blend_blocks_subtract_builder(texturing, mask_evaluate) \ | |
| 3898 | .align 3; \ | |
| 3899 | \ | |
| 3900 | function(blend_blocks_##texturing##_subtract_##mask_evaluate) \ | |
| 3901 | stmdb sp!, { r4, r14 }; \ | |
| 3902 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 3903 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 3904 | \ | |
| 3905 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| 3906 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 3907 | \ | |
| 3908 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3909 | mov c_64, #64; \ | |
| 3910 | \ | |
| 3911 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3912 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3913 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3914 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3915 | \ | |
| 3916 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3917 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3918 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
| 3919 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
| 3920 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3921 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ | |
| 3922 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
| 3923 | \ | |
| 3924 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
| 3925 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3926 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3927 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3928 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3929 | \ | |
| 3930 | subs num_blocks, num_blocks, #1; \ | |
| 3931 | beq 1f; \ | |
| 3932 | \ | |
| 3933 | 0: \ | |
| 3934 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
| 3935 | mov fb_ptr, fb_ptr_next; \ | |
| 3936 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 3937 | \ | |
| 3938 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 3939 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
| 3940 | \ | |
| 3941 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
| 3942 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3943 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
| 3944 | blend_blocks_subtract_set_stb_##texturing(); \ | |
| 3945 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
| 3946 | blend_blocks_subtract_combine_##texturing(); \ | |
| 3947 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
| 3948 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 3949 | \ | |
| 3950 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3951 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3952 | cmp fb_ptr_cmp, #28; \ | |
| 3953 | bls 2f; \ | |
| 3954 | \ | |
| 3955 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3956 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ | |
| 3957 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3958 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3959 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3960 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 3961 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3962 | \ | |
| 3963 | 3: \ | |
| 3964 | subs num_blocks, num_blocks, #1; \ | |
| 3965 | bne 0b; \ | |
| 3966 | \ | |
| 3967 | 1: \ | |
| 3968 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
| 3969 | \ | |
| 3970 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
| 3971 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3972 | blend_blocks_subtract_set_stb_##texturing(); \ | |
| 3973 | blend_blocks_subtract_combine_##texturing(); \ | |
| 3974 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 3975 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
| 3976 | \ | |
| 3977 | ldmia sp!, { r4, pc }; \ | |
| 3978 | \ | |
| 3979 | 2: \ | |
| 3980 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 3981 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 3982 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ | |
| 3983 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3984 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3985 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3986 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3987 | bal 3b \ | |
| 3988 | ||
| 3989 | ||
| 3990 | blend_blocks_subtract_builder(textured, off) | |
| 3991 | blend_blocks_subtract_builder(textured, on) | |
| 3992 | blend_blocks_subtract_builder(untextured, off) | |
| 3993 | blend_blocks_subtract_builder(untextured, on) | |
| 3994 | ||
| 3995 | ||
| 3996 | #define blend_blocks_add_fourth_textured_builder(mask_evaluate) \ | |
| 3997 | .align 3; \ | |
| 3998 | \ | |
| 3999 | function(blend_blocks_textured_add_fourth_##mask_evaluate) \ | |
| 4000 | stmdb sp!, { r4, r14 }; \ | |
| 4001 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 4002 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 4003 | \ | |
| 4004 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| 4005 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 4006 | \ | |
| 4007 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4008 | mov c_64, #64; \ | |
| 4009 | \ | |
| 4010 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4011 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 4012 | vmov.u16 d128_0x83E0, #0x8300; \ | |
| 4013 | vmov.u16 d128_0x1C07, #0x1C00; \ | |
| 4014 | vmov.u16 d128_0x80E0, #0x8000; \ | |
| 4015 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 4016 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 4017 | vorr.u16 d128_0x83E0, #0x00E0; \ | |
| 4018 | vorr.u16 d128_0x1C07, #0x0007; \ | |
| 4019 | vorr.u16 d128_0x80E0, #0x00E0; \ | |
| 4020 | \ | |
| 4021 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 4022 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 4023 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 4024 | vclt.s16 blend_mask, pixels, #0; \ | |
| 4025 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 4026 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 4027 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4028 | \ | |
| 4029 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4030 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 4031 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4032 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 4033 | vand.u16 pixels_mg, pixels_fourth, d128_0x80E0; \ | |
| 4034 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 4035 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 4036 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4037 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 4038 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4039 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 4040 | \ | |
| 4041 | subs num_blocks, num_blocks, #1; \ | |
| 4042 | beq 1f; \ | |
| 4043 | \ | |
| 4044 | 0: \ | |
| 4045 | mov fb_ptr, fb_ptr_next; \ | |
| 4046 | \ | |
| 4047 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 4048 | \ | |
| 4049 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 4050 | vclt.s16 blend_mask, pixels, #0; \ | |
| 4051 | \ | |
| 4052 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4053 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 4054 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4055 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4056 | \ | |
| 4057 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 4058 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 4059 | \ | |
| 4060 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4061 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4062 | cmp fb_ptr_cmp, #28; \ | |
| 4063 | bls 2f; \ | |
| 4064 | \ | |
| 4065 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 4066 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 4067 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 4068 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4069 | vand.u16 pixels_mg, pixels_fourth, d128_0x80E0; \ | |
| 4070 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 4071 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 4072 | \ | |
| 4073 | 3: \ | |
| 4074 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 4075 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4076 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 4077 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4078 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 4079 | \ | |
| 4080 | subs num_blocks, num_blocks, #1; \ | |
| 4081 | bne 0b; \ | |
| 4082 | \ | |
| 4083 | 1: \ | |
| 4084 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4085 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 4086 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
| 4087 | \ | |
| 4088 | ldmia sp!, { r4, pc }; \ | |
| 4089 | \ | |
| 4090 | 2: \ | |
| 4091 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 4092 | vand.u16 pixels_mg, pixels_fourth, d128_0x80E0; \ | |
| 4093 | \ | |
| 4094 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 4095 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 4096 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 4097 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4098 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 4099 | bal 3b \ | |
| 4100 | ||
| 4101 | ||
| 4102 | #define blend_blocks_add_fourth_untextured_builder(mask_evaluate) \ | |
| 4103 | .align 3; \ | |
| 4104 | \ | |
| 4105 | function(blend_blocks_untextured_add_fourth_##mask_evaluate) \ | |
| 4106 | stmdb sp!, { r4, r14 }; \ | |
| 4107 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| 4108 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 4109 | \ | |
| 4110 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| 4111 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
| 4112 | \ | |
| 4113 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4114 | mov c_64, #64; \ | |
| 4115 | \ | |
| 4116 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4117 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 4118 | vmov.u16 d128_0x83E0, #0x8300; \ | |
| 4119 | vmov.u16 d128_0x1C07, #0x1C00; \ | |
| 4120 | vmov.u16 d128_0x00E0, #0x00E0; \ | |
| 4121 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 4122 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 4123 | vorr.u16 d128_0x83E0, #0x00E0; \ | |
| 4124 | vorr.u16 d128_0x1C07, #0x0007; \ | |
| 4125 | \ | |
| 4126 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 4127 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 4128 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 4129 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 4130 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 4131 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4132 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4133 | \ | |
| 4134 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4135 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4136 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4137 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4138 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4139 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4140 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4141 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 4142 | \ | |
| 4143 | subs num_blocks, num_blocks, #1; \ | |
| 4144 | beq 1f; \ | |
| 4145 | \ | |
| 4146 | 0: \ | |
| 4147 | mov fb_ptr, fb_ptr_next; \ | |
| 4148 | \ | |
| 4149 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
| 4150 | \ | |
| 4151 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
| 4152 | \ | |
| 4153 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4154 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4155 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 4156 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4157 | \ | |
| 4158 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 4159 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
| 4160 | \ | |
| 4161 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4162 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4163 | cmp fb_ptr_cmp, #28; \ | |
| 4164 | bls 2f; \ | |
| 4165 | \ | |
| 4166 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 4167 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 4168 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4169 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4170 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4171 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 4172 | \ | |
| 4173 | 3: \ | |
| 4174 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4175 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4176 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4177 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4178 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 4179 | \ | |
| 4180 | subs num_blocks, num_blocks, #1; \ | |
| 4181 | bne 0b; \ | |
| 4182 | \ | |
| 4183 | 1: \ | |
| 4184 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4185 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 4186 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 4187 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
| 4188 | \ | |
| 4189 | ldmia sp!, { r4, pc }; \ | |
| 4190 | \ | |
| 4191 | 2: \ | |
| 4192 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
| 4193 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4194 | \ | |
| 4195 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
| 4196 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
| 4197 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4198 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4199 | bal 3b \ | |
| 4200 | ||
| 4201 | ||
| 4202 | blend_blocks_add_fourth_textured_builder(off) | |
| 4203 | blend_blocks_add_fourth_textured_builder(on) | |
| 4204 | blend_blocks_add_fourth_untextured_builder(off) | |
| 4205 | blend_blocks_add_fourth_untextured_builder(on) | |
| 4206 | ||
| 4207 | // TODO: Optimize this more. Need a scene that actually uses it for | |
| 4208 | // confirmation.. | |
| 4209 | ||
| 4210 | .align 3 | |
| 4211 | ||
| 4212 | function(blend_blocks_textured_unblended_on) | |
| 4213 | stmdb sp!, { r4, r14 } | |
| 4214 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset | |
| 4215 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 4216 | ||
| 4217 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
| 4218 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] | |
| 4219 | ||
| 4220 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 4221 | mov c_64, #64 | |
| 4222 | ||
| 4223 | ldr fb_ptr, [ pixel_ptr, #28 ] | |
| 4224 | vld1.u16 { fb_pixels }, [ fb_ptr ] | |
| 4225 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
| 4226 | vclt.s16 write_mask, fb_pixels, #0 | |
| 4227 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64 | |
| 4228 | ||
| 4229 | subs num_blocks, num_blocks, #1 | |
| 4230 | beq 1f | |
| 4231 | ||
| 4232 | 0: | |
| 4233 | vorr.u16 draw_mask, draw_mask, write_mask | |
| 4234 | vbif.u16 fb_pixels, pixels, draw_mask | |
| 4235 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
| 4236 | ||
| 4237 | ldr fb_ptr, [ pixel_ptr, #28 ] | |
| 4238 | vld1.u16 { fb_pixels }, [ fb_ptr ] | |
| 4239 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
| 4240 | vclt.s16 write_mask, fb_pixels, #0 | |
| 4241 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64 | |
| 4242 | ||
| 4243 | subs num_blocks, num_blocks, #1 | |
| 4244 | bne 0b | |
| 4245 | ||
| 4246 | 1: | |
| 4247 | vorr.u16 draw_mask, draw_mask, write_mask | |
| 4248 | vbif.u16 fb_pixels, pixels, draw_mask | |
| 4249 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
| 4250 | ||
| 4251 | ldmia sp!, { r4, pc } | |
| 4252 | ||
| 4253 | ||
| 4254 | function(blend_blocks_textured_unblended_off) | |
| 4255 | bx lr | |
| 4256 | ||
| 4257 | ||
| 4258 | function(warmup) | |
| 4259 | mov r3, #64 | |
| 4260 | cmp r0, #0 | |
| 4261 | bxeq lr | |
| 4262 | ||
| 4263 | 0: | |
| 4264 | vld1.u32 { u_whole_8, v_whole_8 }, [ r1, :128 ], r3 | |
| 4265 | ||
| 4266 | subs r0, r0, #1 | |
| 4267 | bne 0b | |
| 4268 | ||
| 4269 | bx lr | |
| 4270 | ||
| 4271 | #undef color | |
| 4272 | #undef y | |
| 4273 | #undef height | |
| 4274 | ||
| 4275 | #define psx_gpu r0 | |
| 4276 | #define color r1 | |
| 4277 | #define x r2 | |
| 4278 | #define y r3 | |
| 4279 | ||
| 4280 | #define vram_ptr r0 | |
| 4281 | #define width r3 | |
| 4282 | #define height r12 | |
| 4283 | ||
| 4284 | #define parameter_width_offset 0 | |
| 4285 | #define parameter_height_offset 4 | |
| 4286 | ||
| 4287 | #define color_r r14 | |
| 4288 | #define color_g r4 | |
| 4289 | #define color_b r5 | |
| 4290 | ||
| 4291 | #define left_unaligned r14 | |
| 4292 | #define right_unaligned r4 | |
| 4293 | #define pitch r5 | |
| 4294 | #define num_unaligned r2 | |
| 4295 | #define num_width r6 | |
| 4296 | ||
| 4297 | #undef colors | |
| 4298 | ||
| 4299 | #define colors q0 | |
| 4300 | ||
| 4301 | .align 3 | |
| 4302 | ||
| 4303 | function(render_block_fill_body) | |
| 4304 | ldr vram_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
| 4305 | ldr height, [ sp, #parameter_height_offset ] | |
| 4306 | ||
| 4307 | add vram_ptr, vram_ptr, y, lsl #11 | |
| 4308 | ldr width, [ sp, #parameter_width_offset ] | |
| 4309 | ||
| 4310 | add vram_ptr, vram_ptr, x, lsl #1 | |
| 4311 | stmdb sp!, { r4 - r6, r14 } | |
| 4312 | ||
| 4313 | ubfx color_r, color, #3, #5 | |
| 4314 | ubfx color_g, color, #11, #5 | |
| 4315 | ||
| 4316 | ubfx color_b, color, #19, #5 | |
| 4317 | orr color, color_r, color_g, lsl #5 | |
| 4318 | ||
| 4319 | orr color, color, color_b, lsl #10 | |
| 4320 | add left_unaligned, x, #0x7 | |
| 4321 | ||
| 4322 | bic left_unaligned, left_unaligned, #0x7 | |
| 4323 | vdup.u16 colors, color | |
| 4324 | ||
| 4325 | sub left_unaligned, left_unaligned, x | |
| 4326 | mov pitch, #2048 | |
| 4327 | ||
| 4328 | sub pitch, pitch, width, lsl #1 | |
| 4329 | sub width, width, left_unaligned | |
| 4330 | ||
| 4331 | and right_unaligned, width, #0x7 | |
| 4332 | bic width, width, #0x7 | |
| 4333 | ||
| 4334 | 0: | |
| 4335 | mov num_width, width, lsr #3 | |
| 4336 | ||
| 4337 | movs num_unaligned, left_unaligned | |
| 4338 | beq 2f | |
| 4339 | ||
| 4340 | 1: | |
| 4341 | strh color, [ vram_ptr ], #2 | |
| 4342 | ||
| 4343 | subs num_unaligned, num_unaligned, #1 | |
| 4344 | bne 1b | |
| 4345 | ||
| 4346 | 2: | |
| 4347 | vst1.u32 { colors }, [ vram_ptr, :128 ]! | |
| 4348 | subs num_width, num_width, #1 | |
| 4349 | bne 2b | |
| 4350 | ||
| 4351 | movs num_unaligned, right_unaligned | |
| 4352 | beq 4f | |
| 4353 | ||
| 4354 | 3: | |
| 4355 | strh color, [ vram_ptr ], #2 | |
| 4356 | ||
| 4357 | subs num_unaligned, num_unaligned, #1 | |
| 4358 | bne 3b | |
| 4359 | ||
| 4360 | 4: | |
| 4361 | add vram_ptr, vram_ptr, pitch | |
| 4362 | subs height, height, #1 | |
| 4363 | bne 0b | |
| 4364 | ||
| 4365 | ldmia sp!, { r4 - r6, pc } | |
| 4366 | ||
| 4367 | ||
| 4368 | #undef x | |
| 4369 | #undef y | |
| 4370 | #undef width | |
| 4371 | #undef height | |
| 4372 | #undef fb_ptr | |
| 4373 | #undef texture_mask | |
| 4374 | #undef num_blocks | |
| 4375 | #undef temp | |
| 4376 | #undef dirty_textures_mask | |
| 4377 | #undef clut_ptr | |
| 4378 | #undef current_texture_mask | |
| 4379 | ||
| 4380 | #define psx_gpu r0 | |
| 4381 | #define x r1 | |
| 4382 | #define y r2 | |
| 4383 | #define u r3 | |
| 4384 | #define v r4 | |
| 4385 | #define width r5 | |
| 4386 | #define height r6 | |
| 4387 | #define offset_u r8 | |
| 4388 | #define offset_v r9 | |
| 4389 | #define offset_u_right r10 | |
| 4390 | #define width_rounded r11 | |
| 4391 | #define height_rounded r12 | |
| 4392 | ||
| 4393 | #define texture_offset_base r1 | |
| 4394 | #define tile_width r2 | |
| 4395 | #define tile_height r3 | |
| 4396 | #define num_blocks r4 | |
| 4397 | #define block r5 | |
| 4398 | #define sub_tile_height r6 | |
| 4399 | #define fb_ptr r7 | |
| 4400 | #define texture_mask r8 | |
| 4401 | #define column_data r9 | |
| 4402 | #define texture_offset r10 | |
| 4403 | #define tiles_remaining r11 | |
| 4404 | #define fb_ptr_advance_column r12 | |
| 4405 | #define texture_block_ptr r14 | |
| 4406 | ||
| 4407 | #define texture_page_ptr r3 | |
| 4408 | #define left_block_mask r4 | |
| 4409 | #define right_block_mask r5 | |
| 4410 | #define texture_mask_rev r10 | |
| 4411 | #define control_mask r11 | |
| 4412 | ||
| 4413 | #define dirty_textures_mask r4 | |
| 4414 | #define clut_ptr r5 | |
| 4415 | #define current_texture_mask r6 | |
| 4416 | ||
| 4417 | ||
| 4418 | #undef texels | |
| 4419 | #undef clut_low_a | |
| 4420 | #undef clut_low_b | |
| 4421 | #undef clut_high_a | |
| 4422 | #undef clut_high_b | |
| 4423 | #undef clut_a | |
| 4424 | #undef clut_b | |
| 4425 | #undef texels_low | |
| 4426 | #undef texels_high | |
| 4427 | ||
| 4428 | #define texels d0 | |
| 4429 | #define draw_masks_fb_ptrs q1 | |
| 4430 | ||
| 4431 | #define draw_mask_fb_ptr_left d2 | |
| 4432 | #define draw_mask_fb_ptr_right d3 | |
| 4433 | ||
| 4434 | #define clut_low_a d4 | |
| 4435 | #define clut_low_b d5 | |
| 4436 | #define clut_high_a d6 | |
| 4437 | #define clut_high_b d7 | |
| 4438 | ||
| 4439 | #define block_masks d8 | |
| 4440 | #define block_masks_shifted d9 | |
| 4441 | ||
| 4442 | #define clut_a q2 | |
| 4443 | #define clut_b q3 | |
| 4444 | ||
| 4445 | #define texels_low d10 | |
| 4446 | #define texels_high d11 | |
| 4447 | ||
| 4448 | ||
| 4449 | setup_sprite_flush_blocks_single: | |
| 4450 | vpush { q1 - q4 } | |
| 4451 | ||
| 4452 | stmdb sp!, { r0 - r3, r12, r14 } | |
| 4453 | bl flush_render_block_buffer | |
| 4454 | ldmia sp!, { r0 - r3, r12, r14 } | |
| 4455 | ||
| 4456 | vpop { q1 - q4 } | |
| 4457 | ||
| 4458 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 4459 | ||
| 4460 | mov num_blocks, sub_tile_height | |
| 4461 | bx lr | |
| 4462 | ||
| 4463 | ||
| 4464 | setup_sprite_flush_blocks_double: | |
| 4465 | vpush { q1 - q4 } | |
| 4466 | ||
| 4467 | stmdb sp!, { r0 - r3, r12, r14 } | |
| 4468 | bl flush_render_block_buffer | |
| 4469 | ldmia sp!, { r0 - r3, r12, r14 } | |
| 4470 | ||
| 4471 | vpop { q1 - q4 } | |
| 4472 | ||
| 4473 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 4474 | ||
| 4475 | mov num_blocks, sub_tile_height, lsl #1 | |
| 4476 | bx lr | |
| 4477 | ||
| 4478 | ||
| 4479 | setup_sprite_update_texture_4bpp_cache: | |
| 4480 | stmdb sp!, { r0 - r3, r14 } | |
| 4481 | bl update_texture_4bpp_cache | |
| 4482 | ldmia sp!, { r0 - r3, pc } | |
| 4483 | ||
| 4484 | ||
| 4485 | setup_sprite_update_texture_8bpp_cache: | |
| 4486 | stmdb sp!, { r0 - r3, r14 } | |
| 4487 | bl update_texture_8bpp_cache | |
| 4488 | ldmia sp!, { r0 - r3, pc } | |
| 4489 | ||
| 4490 | ||
| 4491 | #define setup_sprite_tiled_initialize_4bpp() \ | |
| 4492 | ldr dirty_textures_mask, \ | |
| 4493 | [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ]; \ | |
| 4494 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ]; \ | |
| 4495 | \ | |
| 4496 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ]; \ | |
| 4497 | vld1.u32 { clut_a, clut_b }, [ clut_ptr, :128 ]; \ | |
| 4498 | \ | |
| 4499 | tst current_texture_mask, dirty_textures_mask; \ | |
| 4500 | vuzp.u8 clut_a, clut_b; \ | |
| 4501 | \ | |
| 4502 | blne setup_sprite_update_texture_4bpp_cache \ | |
| 4503 | ||
| 4504 | #define setup_sprite_tiled_initialize_8bpp() \ | |
| 4505 | ldr dirty_textures_mask, \ | |
| 4506 | [ psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset ]; \ | |
| 4507 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ]; \ | |
| 4508 | \ | |
| 4509 | tst current_texture_mask, dirty_textures_mask; \ | |
| 4510 | blne setup_sprite_update_texture_8bpp_cache \ | |
| 4511 | ||
| 4512 | ||
| 4513 | #define setup_sprite_tile_setup_block_no(side, offset, texture_mode) \ | |
| 4514 | ||
| 4515 | #define setup_sprite_block_count_single() \ | |
| 4516 | sub_tile_height \ | |
| 4517 | ||
| 4518 | #define setup_sprite_block_count_double() \ | |
| 4519 | sub_tile_height, lsl #1 \ | |
| 4520 | ||
| 4521 | #define setup_sprite_tile_add_blocks(type) \ | |
| 4522 | add num_blocks, num_blocks, setup_sprite_block_count_##type(); \ | |
| 4523 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 4524 | \ | |
| 4525 | blgt setup_sprite_flush_blocks_##type \ | |
| 4526 | ||
| 4527 | ||
| 4528 | #define setup_sprite_tile_full_4bpp(edge) \ | |
| 4529 | setup_sprite_tile_add_blocks(double); \ | |
| 4530 | \ | |
| 4531 | 4: \ | |
| 4532 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4533 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
| 4534 | \ | |
| 4535 | pld [ fb_ptr ]; \ | |
| 4536 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4537 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
| 4538 | \ | |
| 4539 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 4540 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 4541 | \ | |
| 4542 | vst2.u8 { texels_low, texels_high }, [ block, :128 ]; \ | |
| 4543 | add texture_block_ptr, texture_offset, #8; \ | |
| 4544 | \ | |
| 4545 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
| 4546 | add block, block, #40; \ | |
| 4547 | \ | |
| 4548 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4549 | add fb_ptr, fb_ptr, #16; \ | |
| 4550 | \ | |
| 4551 | vst1.u32 { draw_mask_fb_ptr_left }, [ block, :64 ]; \ | |
| 4552 | add block, block, #24; \ | |
| 4553 | \ | |
| 4554 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
| 4555 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 4556 | \ | |
| 4557 | pld [ fb_ptr ]; \ | |
| 4558 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
| 4559 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 4560 | \ | |
| 4561 | vst2.u8 { texels_low, texels_high }, [ block, :128 ]; \ | |
| 4562 | add block, block, #40; \ | |
| 4563 | \ | |
| 4564 | add texture_offset, texture_offset, #0x10; \ | |
| 4565 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
| 4566 | \ | |
| 4567 | vst1.u32 { draw_mask_fb_ptr_right }, [ block, :64 ]; \ | |
| 4568 | add block, block, #24; \ | |
| 4569 | \ | |
| 4570 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4571 | bne 4b; \ | |
| 4572 | \ | |
| 4573 | add texture_offset, texture_offset, #0xF00; \ | |
| 4574 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
| 4575 | ||
| 4576 | ||
| 4577 | #define setup_sprite_tile_half_4bpp(edge) \ | |
| 4578 | setup_sprite_tile_add_blocks(single); \ | |
| 4579 | \ | |
| 4580 | 4: \ | |
| 4581 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4582 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
| 4583 | \ | |
| 4584 | pld [ fb_ptr ]; \ | |
| 4585 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4586 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
| 4587 | \ | |
| 4588 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 4589 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 4590 | \ | |
| 4591 | vst2.u8 { texels_low, texels_high }, [ block, :128 ]; \ | |
| 4592 | add block, block, #40; \ | |
| 4593 | \ | |
| 4594 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4595 | vst1.u32 { draw_mask_fb_ptr_##edge }, [ block, :64 ]; \ | |
| 4596 | \ | |
| 4597 | add block, block, #24; \ | |
| 4598 | add texture_offset, texture_offset, #0x10; \ | |
| 4599 | \ | |
| 4600 | add fb_ptr, fb_ptr, #2048; \ | |
| 4601 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4602 | \ | |
| 4603 | bne 4b; \ | |
| 4604 | \ | |
| 4605 | add texture_offset, texture_offset, #0xF00; \ | |
| 4606 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
| 4607 | ||
| 4608 | ||
| 4609 | #define setup_sprite_tile_full_8bpp(edge) \ | |
| 4610 | setup_sprite_tile_add_blocks(double); \ | |
| 4611 | add block, block, #16; \ | |
| 4612 | \ | |
| 4613 | 4: \ | |
| 4614 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4615 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
| 4616 | \ | |
| 4617 | pld [ fb_ptr ]; \ | |
| 4618 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4619 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
| 4620 | \ | |
| 4621 | add texture_block_ptr, texture_offset, #8; \ | |
| 4622 | vst1.u32 { texels }, [ block, :64 ]; \ | |
| 4623 | \ | |
| 4624 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
| 4625 | add block, block, #24; \ | |
| 4626 | \ | |
| 4627 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4628 | \ | |
| 4629 | add fb_ptr, fb_ptr, #16; \ | |
| 4630 | vst1.u32 { draw_mask_fb_ptr_left }, [ block, :64 ]; \ | |
| 4631 | \ | |
| 4632 | add block, block, #40; \ | |
| 4633 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
| 4634 | pld [ fb_ptr ]; \ | |
| 4635 | \ | |
| 4636 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
| 4637 | vst1.u32 { texels }, [ block, :64 ]; \ | |
| 4638 | add block, block, #24; \ | |
| 4639 | \ | |
| 4640 | add texture_offset, texture_offset, #0x10; \ | |
| 4641 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
| 4642 | \ | |
| 4643 | vst1.u32 { draw_mask_fb_ptr_right }, [ block, :64 ]; \ | |
| 4644 | add block, block, #40; \ | |
| 4645 | \ | |
| 4646 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4647 | bne 4b; \ | |
| 4648 | \ | |
| 4649 | sub block, block, #16; \ | |
| 4650 | add texture_offset, texture_offset, #0xF00; \ | |
| 4651 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
| 4652 | ||
| 4653 | ||
| 4654 | #define setup_sprite_tile_half_8bpp(edge) \ | |
| 4655 | setup_sprite_tile_add_blocks(single); \ | |
| 4656 | add block, block, #16; \ | |
| 4657 | \ | |
| 4658 | 4: \ | |
| 4659 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4660 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
| 4661 | pld [ fb_ptr ]; \ | |
| 4662 | \ | |
| 4663 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4664 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
| 4665 | \ | |
| 4666 | vst1.u32 { texels }, [ block, :64 ]; \ | |
| 4667 | add block, block, #24; \ | |
| 4668 | \ | |
| 4669 | vst1.u32 { draw_mask_fb_ptr_##edge }, [ block, :64 ]; \ | |
| 4670 | add block, block, #40; \ | |
| 4671 | \ | |
| 4672 | add texture_offset, texture_offset, #0x10; \ | |
| 4673 | add fb_ptr, fb_ptr, #2048; \ | |
| 4674 | \ | |
| 4675 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4676 | bne 4b; \ | |
| 4677 | \ | |
| 4678 | sub block, block, #16; \ | |
| 4679 | add texture_offset, texture_offset, #0xF00; \ | |
| 4680 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
| 4681 | ||
| 4682 | ||
| 4683 | #define setup_sprite_tile_column_edge_pre_adjust_half_right() \ | |
| 4684 | add texture_offset, texture_offset_base, #8; \ | |
| 4685 | add fb_ptr, fb_ptr, #16 \ | |
| 4686 | ||
| 4687 | #define setup_sprite_tile_column_edge_pre_adjust_half_left() \ | |
| 4688 | mov texture_offset, texture_offset_base \ | |
| 4689 | ||
| 4690 | #define setup_sprite_tile_column_edge_pre_adjust_half(edge) \ | |
| 4691 | setup_sprite_tile_column_edge_pre_adjust_half_##edge() \ | |
| 4692 | ||
| 4693 | #define setup_sprite_tile_column_edge_pre_adjust_full(edge) \ | |
| 4694 | mov texture_offset, texture_offset_base \ | |
| 4695 | ||
| 4696 | #define setup_sprite_tile_column_edge_post_adjust_half_right() \ | |
| 4697 | sub fb_ptr, fb_ptr, #16 \ | |
| 4698 | ||
| 4699 | #define setup_sprite_tile_column_edge_post_adjust_half_left() \ | |
| 4700 | ||
| 4701 | #define setup_sprite_tile_column_edge_post_adjust_half(edge) \ | |
| 4702 | setup_sprite_tile_column_edge_post_adjust_half_##edge() \ | |
| 4703 | ||
| 4704 | #define setup_sprite_tile_column_edge_post_adjust_full(edge) \ | |
| 4705 | ||
| 4706 | ||
| 4707 | #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode) \ | |
| 4708 | mov sub_tile_height, column_data; \ | |
| 4709 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode(edge); \ | |
| 4710 | setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ | |
| 4711 | setup_sprite_tile_column_edge_post_adjust_##edge_mode(edge) \ | |
| 4712 | ||
| 4713 | #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode) \ | |
| 4714 | and sub_tile_height, column_data, #0xFF; \ | |
| 4715 | mov tiles_remaining, column_data, lsr #16; \ | |
| 4716 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode(edge); \ | |
| 4717 | setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ | |
| 4718 | \ | |
| 4719 | subs tiles_remaining, tiles_remaining, #1; \ | |
| 4720 | beq 2f; \ | |
| 4721 | \ | |
| 4722 | 3: \ | |
| 4723 | mov sub_tile_height, #16; \ | |
| 4724 | setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ | |
| 4725 | subs tiles_remaining, tiles_remaining, #1; \ | |
| 4726 | bne 3b; \ | |
| 4727 | \ | |
| 4728 | 2: \ | |
| 4729 | uxtb sub_tile_height, column_data, ror #8; \ | |
| 4730 | setup_sprite_tile_##edge_mode##_##texture_mode(edge); \ | |
| 4731 | setup_sprite_tile_column_edge_post_adjust_##edge_mode(edge) \ | |
| 4732 | ||
| 4733 | ||
| 4734 | #define setup_sprite_column_data_single() \ | |
| 4735 | mov column_data, height; \ | |
| 4736 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] \ | |
| 4737 | ||
| 4738 | #define setup_sprite_column_data_multi() \ | |
| 4739 | and height_rounded, height_rounded, #0xF; \ | |
| 4740 | rsb column_data, offset_v, #16; \ | |
| 4741 | \ | |
| 4742 | add height_rounded, height_rounded, #1; \ | |
| 4743 | sub tile_height, tile_height, #1; \ | |
| 4744 | \ | |
| 4745 | orr column_data, column_data, tile_height, lsl #16; \ | |
| 4746 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ]; \ | |
| 4747 | \ | |
| 4748 | orr column_data, column_data, height_rounded, lsl #8 \ | |
| 4749 | ||
| 4750 | #define setup_sprite_tile_column_width_single(texture_mode, multi_height, \ | |
| 4751 | edge_mode, edge) \ | |
| 4752 | setup_sprite_##texture_mode##_single_##multi_height##_##edge_mode##_##edge: \ | |
| 4753 | setup_sprite_column_data_##multi_height(); \ | |
| 4754 | vext.32 block_masks_shifted, block_masks, block_masks, #1; \ | |
| 4755 | vorr.u32 block_masks, block_masks, block_masks_shifted; \ | |
| 4756 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
| 4757 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1]; \ | |
| 4758 | \ | |
| 4759 | setup_sprite_tile_column_height_##multi_height(edge_mode, edge, \ | |
| 4760 | texture_mode); \ | |
| 4761 | ldmia sp!, { r4 - r11, pc } \ | |
| 4762 | ||
| 4763 | #define setup_sprite_tiled_advance_column() \ | |
| 4764 | add texture_offset_base, texture_offset_base, #0x100; \ | |
| 4765 | tst texture_offset_base, #0xF00; \ | |
| 4766 | subeq texture_offset_base, texture_offset_base, #(0x100 + 0xF00) \ | |
| 4767 | ||
| 4768 | #define setup_sprite_tile_column_width_multi(tm, multi_height, left_mode, \ | |
| 4769 | right_mode) \ | |
| 4770 | setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode: \ | |
| 4771 | setup_sprite_column_data_##multi_height(); \ | |
| 4772 | mov fb_ptr_advance_column, #32; \ | |
| 4773 | \ | |
| 4774 | sub fb_ptr_advance_column, height, lsl #11; \ | |
| 4775 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
| 4776 | \ | |
| 4777 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1]; \ | |
| 4778 | setup_sprite_tile_column_height_##multi_height(left_mode, right, tm); \ | |
| 4779 | \ | |
| 4780 | subs tile_width, tile_width, #2; \ | |
| 4781 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ | |
| 4782 | \ | |
| 4783 | vmov.u8 draw_masks_fb_ptrs, #0; \ | |
| 4784 | beq 1f; \ | |
| 4785 | \ | |
| 4786 | 0: \ | |
| 4787 | setup_sprite_tiled_advance_column(); \ | |
| 4788 | setup_sprite_tile_column_height_##multi_height(full, none, tm); \ | |
| 4789 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ | |
| 4790 | subs tile_width, tile_width, #1; \ | |
| 4791 | bne 0b; \ | |
| 4792 | \ | |
| 4793 | 1: \ | |
| 4794 | vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ | |
| 4795 | vdup.u8 draw_mask_fb_ptr_right, block_masks[5]; \ | |
| 4796 | \ | |
| 4797 | setup_sprite_tiled_advance_column(); \ | |
| 4798 | setup_sprite_tile_column_height_##multi_height(right_mode, left, tm); \ | |
| 4799 | ldmia sp!, { r4 - r11, pc } \ | |
| 4800 | ||
| 4801 | ||
| 4802 | // r0: psx_gpu | |
| 4803 | // r1: x | |
| 4804 | // r2: y | |
| 4805 | // r3: u | |
| 4806 | // [ sp ]: v | |
| 4807 | // [ sp + 4 ]: width | |
| 4808 | // [ sp + 8 ]: height | |
| 4809 | // [ sp + 12 ]: color (unused) | |
| 4810 | ||
| 4811 | #define setup_sprite_tiled_builder(texture_mode) \ | |
| 4812 | \ | |
| 4813 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, full); \ | |
| 4814 | setup_sprite_tile_column_width_single(texture_mode, multi, full, none); \ | |
| 4815 | setup_sprite_tile_column_width_multi(texture_mode, single, full, full); \ | |
| 4816 | setup_sprite_tile_column_width_single(texture_mode, single, full, none); \ | |
| 4817 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, full); \ | |
| 4818 | setup_sprite_tile_column_width_single(texture_mode, multi, half, right); \ | |
| 4819 | setup_sprite_tile_column_width_multi(texture_mode, single, half, full); \ | |
| 4820 | setup_sprite_tile_column_width_single(texture_mode, single, half, right); \ | |
| 4821 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, half); \ | |
| 4822 | setup_sprite_tile_column_width_single(texture_mode, multi, half, left); \ | |
| 4823 | setup_sprite_tile_column_width_multi(texture_mode, single, full, half); \ | |
| 4824 | setup_sprite_tile_column_width_single(texture_mode, single, half, left); \ | |
| 4825 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, half); \ | |
| 4826 | setup_sprite_tile_column_width_multi(texture_mode, single, half, half); \ | |
| 4827 | \ | |
| 4828 | .align 4; \ | |
| 4829 | \ | |
| 4830 | function(setup_sprite_##texture_mode) \ | |
| 4831 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 4832 | setup_sprite_tiled_initialize_##texture_mode(); \ | |
| 4833 | \ | |
| 4834 | ldr v, [ sp, #36 ]; \ | |
| 4835 | and offset_u, u, #0xF; \ | |
| 4836 | \ | |
| 4837 | ldr width, [ sp, #40 ]; \ | |
| 4838 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ]; \ | |
| 4839 | \ | |
| 4840 | ldr height, [ sp, #44 ]; \ | |
| 4841 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 4842 | \ | |
| 4843 | add fb_ptr, fb_ptr, x, lsl #1; \ | |
| 4844 | and offset_v, v, #0xF; \ | |
| 4845 | \ | |
| 4846 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
| 4847 | add width_rounded, offset_u, width; \ | |
| 4848 | \ | |
| 4849 | add height_rounded, offset_v, height; \ | |
| 4850 | add width_rounded, width_rounded, #15; \ | |
| 4851 | \ | |
| 4852 | add height_rounded, height_rounded, #15; \ | |
| 4853 | mov tile_width, width_rounded, lsr #4; \ | |
| 4854 | \ | |
| 4855 | /* texture_offset_base = VH-VL-00-00 */\ | |
| 4856 | mov texture_offset_base, v, lsl #8; \ | |
| 4857 | and offset_u_right, width_rounded, #0xF; \ | |
| 4858 | \ | |
| 4859 | /* texture_offset_base = VH-UH-UL-00 */\ | |
| 4860 | bfi texture_offset_base, u, #4, #8; \ | |
| 4861 | movw right_block_mask, #0xFFFE; \ | |
| 4862 | \ | |
| 4863 | /* texture_offset_base = VH-UH-VL-00 */\ | |
| 4864 | bfi texture_offset_base, v, #4, #4; \ | |
| 4865 | movw left_block_mask, #0xFFFF; \ | |
| 4866 | \ | |
| 4867 | mov tile_height, height_rounded, lsr #4; \ | |
| 4868 | mvn left_block_mask, left_block_mask, lsl offset_u; \ | |
| 4869 | \ | |
| 4870 | /* texture_mask = HH-HL-WH-WL */\ | |
| 4871 | ldrh texture_mask, [ psx_gpu, #psx_gpu_texture_mask_width_offset ]; \ | |
| 4872 | mov right_block_mask, right_block_mask, lsl offset_u_right; \ | |
| 4873 | \ | |
| 4874 | /* texture_mask_rev = WH-WL-HH-HL */\ | |
| 4875 | rev16 texture_mask_rev, texture_mask; \ | |
| 4876 | vmov block_masks, left_block_mask, right_block_mask; \ | |
| 4877 | \ | |
| 4878 | /* texture_mask = HH-HL-HL-WL */\ | |
| 4879 | bfi texture_mask, texture_mask_rev, #4, #4; \ | |
| 4880 | /* texture_mask_rev = 00-00-00-WH */\ | |
| 4881 | mov texture_mask_rev, texture_mask_rev, lsr #12; \ | |
| 4882 | \ | |
| 4883 | /* texture_mask = HH-WH-HL-WL */\ | |
| 4884 | bfi texture_mask, texture_mask_rev, #8, #4; \ | |
| 4885 | and left_block_mask, left_block_mask, #0xFF; \ | |
| 4886 | \ | |
| 4887 | mov control_mask, #0; \ | |
| 4888 | cmp left_block_mask, #0xFF; \ | |
| 4889 | \ | |
| 4890 | uxtb right_block_mask, right_block_mask, ror #8; \ | |
| 4891 | orreq control_mask, control_mask, #0x4; \ | |
| 4892 | \ | |
| 4893 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
| 4894 | cmp right_block_mask, #0xFF; \ | |
| 4895 | \ | |
| 4896 | orreq control_mask, control_mask, #0x8; \ | |
| 4897 | cmp tile_width, #1; \ | |
| 4898 | \ | |
| 4899 | add block, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4900 | orreq control_mask, control_mask, #0x1; \ | |
| 4901 | \ | |
| 4902 | cmp tile_height, #1; \ | |
| 4903 | add block, block, num_blocks, lsl #6; \ | |
| 4904 | \ | |
| 4905 | orreq control_mask, control_mask, #0x2; \ | |
| 4906 | ldr pc, [ pc, control_mask, lsl #2 ]; \ | |
| 4907 | nop; \ | |
| 4908 | \ | |
| 4909 | .word setup_sprite_##texture_mode##_multi_multi_full_full; \ | |
| 4910 | .word setup_sprite_##texture_mode##_single_multi_full_none; \ | |
| 4911 | .word setup_sprite_##texture_mode##_multi_single_full_full; \ | |
| 4912 | .word setup_sprite_##texture_mode##_single_single_full_none; \ | |
| 4913 | .word setup_sprite_##texture_mode##_multi_multi_half_full; \ | |
| 4914 | .word setup_sprite_##texture_mode##_single_multi_half_right; \ | |
| 4915 | .word setup_sprite_##texture_mode##_multi_single_half_full; \ | |
| 4916 | .word setup_sprite_##texture_mode##_single_single_half_right; \ | |
| 4917 | .word setup_sprite_##texture_mode##_multi_multi_full_half; \ | |
| 4918 | .word setup_sprite_##texture_mode##_single_multi_half_left; \ | |
| 4919 | .word setup_sprite_##texture_mode##_multi_single_full_half; \ | |
| 4920 | .word setup_sprite_##texture_mode##_single_single_half_left; \ | |
| 4921 | .word setup_sprite_##texture_mode##_multi_multi_half_half; \ | |
| 4922 | .word 0x00000000; \ | |
| 4923 | .word setup_sprite_##texture_mode##_multi_single_half_half \ | |
| 4924 | ||
| 4925 | ||
| 4926 | setup_sprite_tiled_builder(4bpp); | |
| 4927 | setup_sprite_tiled_builder(8bpp); | |
| 4928 | ||
| 4929 | ||
| 4930 | #undef block_ptr | |
| 4931 | #undef num_blocks | |
| 4932 | #undef clut_ptr | |
| 4933 | ||
| 4934 | #define psx_gpu r0 | |
| 4935 | #define block_ptr r0 | |
| 4936 | #define num_blocks r1 | |
| 4937 | #define clut_ptr r2 | |
| 4938 | #define texel_shift_mask r3 | |
| 4939 | #define block_pixels_a r4 | |
| 4940 | #define block_pixels_b r5 | |
| 4941 | #define texel_0 r6 | |
| 4942 | #define texel_2 r7 | |
| 4943 | #define texel_4 r8 | |
| 4944 | #define texel_6 r9 | |
| 4945 | #define texel_1 r10 | |
| 4946 | #define texel_3 r11 | |
| 4947 | #define texel_5 r12 | |
| 4948 | #define texel_7 r14 | |
| 4949 | #define texels_01 r6 | |
| 4950 | #define texels_23 r7 | |
| 4951 | #define texels_45 r8 | |
| 4952 | #define texels_67 r9 | |
| 4953 | ||
| 4954 | function(texture_sprite_blocks_8bpp) | |
| 4955 | stmdb sp!, { r4 - r11, r14 } | |
| 4956 | movw texel_shift_mask, #(0xFF << 1) | |
| 4957 | ||
| 4958 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 4959 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ] | |
| 4960 | ||
| 4961 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 4962 | ldr block_pixels_a, [ block_ptr, #16 ] | |
| 4963 | ||
| 4964 | 0: | |
| 4965 | and texel_0, texel_shift_mask, block_pixels_a, lsl #1 | |
| 4966 | ldr block_pixels_b, [ block_ptr, #20 ] | |
| 4967 | ||
| 4968 | and texel_1, texel_shift_mask, block_pixels_a, lsr #7 | |
| 4969 | ldrh texel_0, [ clut_ptr, texel_0 ] | |
| 4970 | ||
| 4971 | and texel_2, texel_shift_mask, block_pixels_a, lsr #15 | |
| 4972 | ldrh texel_1, [ clut_ptr, texel_1 ] | |
| 4973 | ||
| 4974 | and texel_3, texel_shift_mask, block_pixels_a, lsr #23 | |
| 4975 | ldr block_pixels_a, [ block_ptr, #(64 + 16) ] | |
| 4976 | ||
| 4977 | ldrh texel_2, [ clut_ptr, texel_2 ] | |
| 4978 | and texel_4, texel_shift_mask, block_pixels_b, lsl #1 | |
| 4979 | ||
| 4980 | ldrh texel_3, [ clut_ptr, texel_3 ] | |
| 4981 | and texel_5, texel_shift_mask, block_pixels_b, lsr #7 | |
| 4982 | ||
| 4983 | ldrh texel_4, [ clut_ptr, texel_4 ] | |
| 4984 | and texel_6, texel_shift_mask, block_pixels_b, lsr #15 | |
| 4985 | ||
| 4986 | ldrh texel_5, [ clut_ptr, texel_5 ] | |
| 4987 | and texel_7, texel_shift_mask, block_pixels_b, lsr #23 | |
| 4988 | ||
| 4989 | ldrh texel_6, [ clut_ptr, texel_6 ] | |
| 4990 | orr texels_01, texel_0, texel_1, lsl #16 | |
| 4991 | ||
| 4992 | ldrh texel_7, [ clut_ptr, texel_7 ] | |
| 4993 | orr texels_23, texel_2, texel_3, lsl #16 | |
| 4994 | ||
| 4995 | orr texels_45, texel_4, texel_5, lsl #16 | |
| 4996 | str texels_01, [ block_ptr, #0 ] | |
| 4997 | ||
| 4998 | orr texels_67, texel_6, texel_7, lsl #16 | |
| 4999 | str texels_23, [ block_ptr, #4 ] | |
| 5000 | ||
| 5001 | subs num_blocks, num_blocks, #1 | |
| 5002 | str texels_45, [ block_ptr, #8 ] | |
| 5003 | ||
| 5004 | str texels_67, [ block_ptr, #12 ] | |
| 5005 | add block_ptr, block_ptr, #64 | |
| 5006 | ||
| 5007 | bne 0b | |
| 5008 | ||
| 5009 | ldmia sp!, { r4 - r11, pc } | |
| 5010 | ||
| 5011 | ||
| 5012 | #undef width_rounded | |
| 5013 | #undef texture_mask | |
| 5014 | #undef num_blocks | |
| 5015 | #undef texture_offset | |
| 5016 | ||
| 5017 | #define psx_gpu r0 | |
| 5018 | #define x r1 | |
| 5019 | #define y r2 | |
| 5020 | #define u r3 | |
| 5021 | #define v r4 | |
| 5022 | #define width r5 | |
| 5023 | #define height r6 | |
| 5024 | #define left_offset r8 | |
| 5025 | #define width_rounded r9 | |
| 5026 | #define right_width r10 | |
| 5027 | #define block_width r11 | |
| 5028 | ||
| 5029 | #define texture_offset_base r1 | |
| 5030 | #define texture_mask r2 | |
| 5031 | #define texture_page_ptr r3 | |
| 5032 | #define num_blocks r4 | |
| 5033 | #define block r5 | |
| 5034 | #define fb_ptr r7 | |
| 5035 | #define texture_offset r8 | |
| 5036 | #define blocks_remaining r9 | |
| 5037 | #define fb_ptr_pitch r12 | |
| 5038 | #define texture_block_ptr r14 | |
| 5039 | ||
| 5040 | #define texture_mask_width r2 | |
| 5041 | #define texture_mask_height r3 | |
| 5042 | #define left_mask_bits r4 | |
| 5043 | #define right_mask_bits r5 | |
| 5044 | ||
| 5045 | ||
| 5046 | #undef block_masks | |
| 5047 | #undef block_masks_shifted | |
| 5048 | #undef texels | |
| 5049 | ||
| 5050 | #define block_masks d0 | |
| 5051 | #define block_masks_shifted d1 | |
| 5052 | #define draw_mask_fb_ptr d2 | |
| 5053 | #define texels q2 | |
| 5054 | ||
| 5055 | ||
| 5056 | setup_sprites_16bpp_flush_single: | |
| 5057 | vpush { d0 - d2 } | |
| 5058 | ||
| 5059 | stmdb sp!, { r0 - r3, r12, r14 } | |
| 5060 | bl flush_render_block_buffer | |
| 5061 | ldmia sp!, { r0 - r3, r12, r14 } | |
| 5062 | ||
| 5063 | vpop { d0 - d2 } | |
| 5064 | ||
| 5065 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5066 | mov num_blocks, #1 | |
| 5067 | ||
| 5068 | bx lr | |
| 5069 | ||
| 5070 | setup_sprites_16bpp_flush_row: | |
| 5071 | vpush { d0 - d2 } | |
| 5072 | ||
| 5073 | stmdb sp!, { r0 - r3, r12, r14 } | |
| 5074 | bl flush_render_block_buffer | |
| 5075 | ldmia sp!, { r0 - r3, r12, r14 } | |
| 5076 | ||
| 5077 | vpop { d0 - d2 } | |
| 5078 | ||
| 5079 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5080 | mov num_blocks, block_width | |
| 5081 | ||
| 5082 | bx lr | |
| 5083 | ||
| 5084 | function(setup_sprite_16bpp) | |
| 5085 | stmdb sp!, { r4 - r11, r14 } | |
| 5086 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
| 5087 | ||
| 5088 | ldr v, [ sp, #36 ] | |
| 5089 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 5090 | ||
| 5091 | ldr width, [ sp, #40 ] | |
| 5092 | add fb_ptr, fb_ptr, x, lsl #1 | |
| 5093 | ||
| 5094 | ldr height, [ sp, #44 ] | |
| 5095 | and left_offset, u, #0x7 | |
| 5096 | ||
| 5097 | add texture_offset_base, u, u | |
| 5098 | add width_rounded, width, #7 | |
| 5099 | ||
| 5100 | add texture_offset_base, v, lsl #11 | |
| 5101 | mov left_mask_bits, #0xFF | |
| 5102 | ||
| 5103 | ldrb texture_mask_width, [ psx_gpu, #psx_gpu_texture_mask_width_offset ] | |
| 5104 | add width_rounded, width_rounded, left_offset | |
| 5105 | ||
| 5106 | ldrb texture_mask_height, [ psx_gpu, #psx_gpu_texture_mask_height_offset ] | |
| 5107 | sub fb_ptr, fb_ptr, left_offset, lsl #1 | |
| 5108 | ||
| 5109 | add texture_mask, texture_mask_width, texture_mask_width | |
| 5110 | mov right_mask_bits, #0xFE | |
| 5111 | ||
| 5112 | and right_width, width_rounded, #0x7 | |
| 5113 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
| 5114 | ||
| 5115 | add texture_mask, texture_mask_height, lsl #11 | |
| 5116 | mov block_width, width_rounded, lsr #3 | |
| 5117 | ||
| 5118 | mov right_mask_bits, right_mask_bits, lsl right_width | |
| 5119 | movw fb_ptr_pitch, #(2048 + 16) | |
| 5120 | ||
| 5121 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4 | |
| 5122 | vmov block_masks, left_mask_bits, right_mask_bits | |
| 5123 | ||
| 5124 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 5125 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5126 | ||
| 5127 | bic texture_offset_base, texture_offset_base, #0x7 | |
| 5128 | cmp block_width, #1 | |
| 5129 | ||
| 5130 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
| 5131 | add block, block, num_blocks, lsl #6 | |
| 5132 | ||
| 5133 | bne 0f | |
| 5134 | ||
| 5135 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
| 5136 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
| 5137 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
| 5138 | ||
| 5139 | 1: | |
| 5140 | add num_blocks, num_blocks, #1 | |
| 5141 | cmp num_blocks, #MAX_BLOCKS | |
| 5142 | blgt setup_sprites_16bpp_flush_single | |
| 5143 | ||
| 5144 | and texture_block_ptr, texture_offset_base, texture_mask | |
| 5145 | subs height, height, #1 | |
| 5146 | ||
| 5147 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5148 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
| 5149 | ||
| 5150 | vst1.u32 { texels }, [ block, :128 ] | |
| 5151 | add block, block, #40 | |
| 5152 | ||
| 5153 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| 5154 | pld [ fb_ptr ] | |
| 5155 | ||
| 5156 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
| 5157 | ||
| 5158 | add block, block, #24 | |
| 5159 | add texture_offset_base, texture_offset_base, #2048 | |
| 5160 | add fb_ptr, fb_ptr, #2048 | |
| 5161 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 5162 | bne 1b | |
| 5163 | ||
| 5164 | ldmia sp!, { r4 - r11, pc } | |
| 5165 | ||
| 5166 | 0: | |
| 5167 | add num_blocks, num_blocks, block_width | |
| 5168 | mov texture_offset, texture_offset_base | |
| 5169 | ||
| 5170 | cmp num_blocks, #MAX_BLOCKS | |
| 5171 | blgt setup_sprites_16bpp_flush_row | |
| 5172 | ||
| 5173 | add texture_offset_base, texture_offset_base, #2048 | |
| 5174 | and texture_block_ptr, texture_offset, texture_mask | |
| 5175 | ||
| 5176 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5177 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
| 5178 | ||
| 5179 | vst1.u32 { texels }, [ block, :128 ] | |
| 5180 | add block, block, #40 | |
| 5181 | ||
| 5182 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
| 5183 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| 5184 | pld [ fb_ptr ] | |
| 5185 | ||
| 5186 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
| 5187 | subs blocks_remaining, block_width, #2 | |
| 5188 | ||
| 5189 | add texture_offset, texture_offset, #16 | |
| 5190 | add fb_ptr, fb_ptr, #16 | |
| 5191 | ||
| 5192 | vmov.u8 draw_mask_fb_ptr, #0 | |
| 5193 | ||
| 5194 | add block, block, #24 | |
| 5195 | beq 2f | |
| 5196 | ||
| 5197 | 1: | |
| 5198 | and texture_block_ptr, texture_offset, texture_mask | |
| 5199 | subs blocks_remaining, blocks_remaining, #1 | |
| 5200 | ||
| 5201 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5202 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
| 5203 | ||
| 5204 | vst1.u32 { texels }, [ block, :128 ] | |
| 5205 | add block, block, #40 | |
| 5206 | ||
| 5207 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| 5208 | pld [ fb_ptr ] | |
| 5209 | ||
| 5210 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
| 5211 | ||
| 5212 | add texture_offset, texture_offset, #16 | |
| 5213 | add fb_ptr, fb_ptr, #16 | |
| 5214 | ||
| 5215 | add block, block, #24 | |
| 5216 | bne 1b | |
| 5217 | ||
| 5218 | 2: | |
| 5219 | and texture_block_ptr, texture_offset, texture_mask | |
| 5220 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5221 | ||
| 5222 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
| 5223 | vdup.u8 draw_mask_fb_ptr, block_masks[4] | |
| 5224 | ||
| 5225 | vst1.u32 { texels }, [ block, :128 ] | |
| 5226 | add block, block, #40 | |
| 5227 | ||
| 5228 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| 5229 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
| 5230 | ||
| 5231 | add block, block, #24 | |
| 5232 | subs height, height, #1 | |
| 5233 | ||
| 5234 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
| 5235 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
| 5236 | ||
| 5237 | bne 0b | |
| 5238 | ||
| 5239 | ldmia sp!, { r4 - r11, pc } | |
| 5240 | ||
| 5241 | ||
| 5242 | #undef texture_page_ptr | |
| 5243 | #undef vram_ptr | |
| 5244 | #undef dirty_textures_mask | |
| 5245 | #undef current_texture_mask | |
| 5246 | ||
| 5247 | #define psx_gpu r0 | |
| 5248 | #define current_texture_page r1 | |
| 5249 | #define texture_page_ptr r2 | |
| 5250 | #define vram_ptr_a r3 | |
| 5251 | #define current_texture_page_x r12 | |
| 5252 | #define current_texture_page_y r4 | |
| 5253 | #define dirty_textures_mask r5 | |
| 5254 | #define tile_y r6 | |
| 5255 | #define tile_x r7 | |
| 5256 | #define sub_y r8 | |
| 5257 | #define current_texture_mask r9 | |
| 5258 | #define c_4096 r10 | |
| 5259 | #define vram_ptr_b r11 | |
| 5260 | ||
| 5261 | #define texel_block_a d0 | |
| 5262 | #define texel_block_b d1 | |
| 5263 | #define texel_block_expanded_a q1 | |
| 5264 | #define texel_block_expanded_b q2 | |
| 5265 | #define texel_block_expanded_ab q2 | |
| 5266 | #define texel_block_expanded_c q3 | |
| 5267 | #define texel_block_expanded_d q4 | |
| 5268 | #define texel_block_expanded_cd q3 | |
| 5269 | ||
| 5270 | function(update_texture_4bpp_cache) | |
| 5271 | stmdb sp!, { r4 - r11, r14 } | |
| 5272 | vpush { q0 - q3 } | |
| 5273 | ||
| 5274 | ldrb current_texture_page, [ psx_gpu, #psx_gpu_current_texture_page_offset ] | |
| 5275 | ||
| 5276 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
| 5277 | ldr vram_ptr_a, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
| 5278 | ||
| 5279 | and current_texture_page_x, current_texture_page, #0xF | |
| 5280 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ] | |
| 5281 | ||
| 5282 | mov current_texture_page_y, current_texture_page, lsr #4 | |
| 5283 | ldr dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ] | |
| 5284 | ||
| 5285 | add vram_ptr_a, vram_ptr_a, current_texture_page_y, lsl #19 | |
| 5286 | mov tile_y, #16 | |
| 5287 | ||
| 5288 | add vram_ptr_a, vram_ptr_a, current_texture_page_x, lsl #7 | |
| 5289 | bic dirty_textures_mask, current_texture_mask | |
| 5290 | ||
| 5291 | mov tile_x, #16 | |
| 5292 | str dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ] | |
| 5293 | ||
| 5294 | mov sub_y, #8 | |
| 5295 | movw c_4096, #4096 | |
| 5296 | ||
| 5297 | add vram_ptr_b, vram_ptr_a, #2048 | |
| 5298 | ||
| 5299 | 0: | |
| 5300 | vld1.u32 { texel_block_a }, [ vram_ptr_a, :64 ], c_4096 | |
| 5301 | vld1.u32 { texel_block_b }, [ vram_ptr_b, :64 ], c_4096 | |
| 5302 | ||
| 5303 | vmovl.u8 texel_block_expanded_a, texel_block_a | |
| 5304 | vshll.u8 texel_block_expanded_b, texel_block_a, #4 | |
| 5305 | vmovl.u8 texel_block_expanded_c, texel_block_b | |
| 5306 | vshll.u8 texel_block_expanded_d, texel_block_b, #4 | |
| 5307 | ||
| 5308 | vbic.u16 texel_block_expanded_a, #0x00F0 | |
| 5309 | vbic.u16 texel_block_expanded_b, #0x00F0 | |
| 5310 | vbic.u16 texel_block_expanded_c, #0x00F0 | |
| 5311 | vbic.u16 texel_block_expanded_d, #0x00F0 | |
| 5312 | ||
| 5313 | vorr.u16 texel_block_expanded_ab, texel_block_expanded_a, \ | |
| 5314 | texel_block_expanded_b | |
| 5315 | vorr.u16 texel_block_expanded_cd, texel_block_expanded_c, \ | |
| 5316 | texel_block_expanded_d | |
| 5317 | ||
| 5318 | vst1.u32 { texel_block_expanded_ab, texel_block_expanded_cd }, \ | |
| 5319 | [ texture_page_ptr, :256 ]! | |
| 5320 | ||
| 5321 | subs sub_y, sub_y, #1 | |
| 5322 | bne 0b | |
| 5323 | ||
| 5324 | mov sub_y, #8 | |
| 5325 | add vram_ptr_a, vram_ptr_a, #8 | |
| 5326 | add vram_ptr_b, vram_ptr_b, #8 | |
| 5327 | ||
| 5328 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 5329 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 5330 | ||
| 5331 | subs tile_x, tile_x, #1 | |
| 5332 | bne 0b | |
| 5333 | ||
| 5334 | mov tile_x, #16 | |
| 5335 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 5336 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 5337 | ||
| 5338 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
| 5339 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
| 5340 | ||
| 5341 | subs tile_y, tile_y, #1 | |
| 5342 | bne 0b | |
| 5343 | ||
| 5344 | vpop { q0 - q3 } | |
| 5345 | ldmia sp!, { r4 - r11, pc } | |
| 5346 | ||
| 5347 | ||
| 5348 | #undef current_texture_page | |
| 5349 | ||
| 5350 | #define psx_gpu r0 | |
| 5351 | #define texture_page r1 | |
| 5352 | #define texture_page_ptr r2 | |
| 5353 | #define vram_ptr_a r3 | |
| 5354 | #define texture_page_x r12 | |
| 5355 | #define texture_page_y r4 | |
| 5356 | #define current_texture_page r5 | |
| 5357 | #define tile_y r6 | |
| 5358 | #define tile_x r7 | |
| 5359 | #define sub_y r8 | |
| 5360 | #define c_4096 r10 | |
| 5361 | #define vram_ptr_b r11 | |
| 5362 | ||
| 5363 | ||
| 5364 | #undef texels_a | |
| 5365 | #undef texels_b | |
| 5366 | ||
| 5367 | #define texels_a q0 | |
| 5368 | #define texels_b q1 | |
| 5369 | #define texels_c q2 | |
| 5370 | #define texels_d q3 | |
| 5371 | ||
| 5372 | ||
| 5373 | function(update_texture_8bpp_cache_slice) | |
| 5374 | stmdb sp!, { r4 - r11, r14 } | |
| 5375 | vpush { q0 - q3 } | |
| 5376 | ||
| 5377 | ldrb current_texture_page, [ psx_gpu, #psx_gpu_current_texture_page_offset ] | |
| 5378 | ldr vram_ptr_a, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
| 5379 | ||
| 5380 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
| 5381 | mov tile_y, #16 | |
| 5382 | ||
| 5383 | and texture_page_x, texture_page, #0xF | |
| 5384 | mov texture_page_y, texture_page, lsr #4 | |
| 5385 | ||
| 5386 | add vram_ptr_a, vram_ptr_a, texture_page_x, lsl #7 | |
| 5387 | mov tile_x, #8 | |
| 5388 | ||
| 5389 | add vram_ptr_a, vram_ptr_a, texture_page_y, lsl #19 | |
| 5390 | eor current_texture_page, current_texture_page, texture_page | |
| 5391 | ||
| 5392 | ands current_texture_page, current_texture_page, #0x1 | |
| 5393 | mov sub_y, #4 | |
| 5394 | ||
| 5395 | addne texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
| 5396 | movw c_4096, #4096 | |
| 5397 | ||
| 5398 | add vram_ptr_b, vram_ptr_a, #2048 | |
| 5399 | ||
| 5400 | 0: | |
| 5401 | vld1.u32 { texels_a }, [ vram_ptr_a, :128 ], c_4096 | |
| 5402 | vld1.u32 { texels_b }, [ vram_ptr_b, :128 ], c_4096 | |
| 5403 | vld1.u32 { texels_c }, [ vram_ptr_a, :128 ], c_4096 | |
| 5404 | vld1.u32 { texels_d }, [ vram_ptr_b, :128 ], c_4096 | |
| 5405 | ||
| 5406 | vst1.u32 { texels_a, texels_b }, [ texture_page_ptr, :256 ]! | |
| 5407 | vst1.u32 { texels_c, texels_d }, [ texture_page_ptr, :256 ]! | |
| 5408 | ||
| 5409 | subs sub_y, sub_y, #1 | |
| 5410 | bne 0b | |
| 5411 | ||
| 5412 | mov sub_y, #4 | |
| 5413 | ||
| 5414 | add vram_ptr_a, vram_ptr_a, #16 | |
| 5415 | add vram_ptr_b, vram_ptr_b, #16 | |
| 5416 | ||
| 5417 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 5418 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 5419 | ||
| 5420 | subs tile_x, tile_x, #1 | |
| 5421 | bne 0b | |
| 5422 | ||
| 5423 | mov tile_x, #8 | |
| 5424 | ||
| 5425 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 5426 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 5427 | ||
| 5428 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
| 5429 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
| 5430 | ||
| 5431 | subs tile_y, tile_y, #1 | |
| 5432 | add texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
| 5433 | ||
| 5434 | bne 0b | |
| 5435 | ||
| 5436 | vpop { q0 - q3 } | |
| 5437 | ldmia sp!, { r4 - r11, pc } | |
| 5438 |