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 | #include <stdio.h> | |
16 | #include <stdlib.h> | |
17 | ||
18 | #include "SDL.h" | |
19 | #include "common.h" | |
20 | ||
21 | extern u32 span_pixels; | |
22 | extern u32 span_pixel_blocks; | |
75e28f62 E |
23 | extern u32 spans; |
24 | extern u32 triangles; | |
25 | extern u32 sprites; | |
26 | extern u32 sprites_4bpp; | |
27 | extern u32 sprites_8bpp; | |
28 | extern u32 sprites_16bpp; | |
29 | extern u32 sprites_untextured; | |
30 | extern u32 sprite_blocks; | |
31 | extern u32 lines; | |
32 | extern u32 texels_4bpp; | |
33 | extern u32 texels_8bpp; | |
34 | extern u32 texels_16bpp; | |
35 | extern u32 texel_blocks_4bpp; | |
36 | extern u32 texel_blocks_8bpp; | |
37 | extern u32 texel_blocks_16bpp; | |
38 | extern u32 texel_blocks_untextured; | |
39 | extern u32 blend_blocks; | |
75e28f62 E |
40 | extern u32 render_buffer_flushes; |
41 | extern u32 state_changes; | |
42 | extern u32 trivial_rejects; | |
43 | extern u32 left_split_triangles; | |
44 | extern u32 flat_triangles; | |
45 | extern u32 clipped_triangles; | |
46 | extern u32 zero_block_spans; | |
47 | extern u32 texture_cache_loads; | |
3867c6ef | 48 | extern u32 false_modulated_blocks; |
75e28f62 E |
49 | |
50 | static u32 mismatches; | |
51 | ||
52 | typedef struct | |
53 | { | |
54 | u16 vram[1024 * 512]; | |
55 | u32 gpu_register[15]; | |
56 | u32 status; | |
57 | } gpu_dump_struct; | |
58 | ||
59 | static gpu_dump_struct state; | |
60 | ||
61 | psx_gpu_struct __attribute__((aligned(256))) _psx_gpu; | |
3867c6ef | 62 | u16 __attribute__((aligned(256))) _vram[(1024 * 512) + 1024]; |
75e28f62 E |
63 | |
64 | #define percent_of(numerator, denominator) \ | |
65 | ((((double)(numerator)) / (denominator)) * 100.0) \ | |
66 | ||
67 | void clear_stats(void) | |
68 | { | |
69 | triangles = 0; | |
70 | sprites = 0; | |
71 | sprites_4bpp = 0; | |
72 | sprites_8bpp = 0; | |
73 | sprites_16bpp = 0; | |
74 | sprites_untextured = 0; | |
75 | sprite_blocks = 0; | |
76 | lines = 0; | |
77 | span_pixels = 0; | |
78 | span_pixel_blocks = 0; | |
75e28f62 E |
79 | spans = 0; |
80 | texels_4bpp = 0; | |
81 | texels_8bpp = 0; | |
82 | texels_16bpp = 0; | |
83 | texel_blocks_untextured = 0; | |
84 | texel_blocks_4bpp = 0; | |
85 | texel_blocks_8bpp = 0; | |
86 | texel_blocks_16bpp = 0; | |
87 | blend_blocks = 0; | |
75e28f62 E |
88 | render_buffer_flushes = 0; |
89 | state_changes = 0; | |
90 | trivial_rejects = 0; | |
91 | left_split_triangles = 0; | |
92 | flat_triangles = 0; | |
93 | clipped_triangles = 0; | |
94 | zero_block_spans = 0; | |
95 | texture_cache_loads = 0; | |
3867c6ef | 96 | false_modulated_blocks = 0; |
75e28f62 E |
97 | } |
98 | ||
99 | void update_screen(psx_gpu_struct *psx_gpu, SDL_Surface *screen) | |
100 | { | |
101 | u32 x, y; | |
102 | ||
103 | for(y = 0; y < 512; y++) | |
104 | { | |
105 | for(x = 0; x < 1024; x++) | |
106 | { | |
107 | u32 pixel = psx_gpu->vram_ptr[(y * 1024) + x]; | |
108 | ((u32 *)screen->pixels)[(y * 1024) + x] = | |
109 | ((pixel & 0x1F) << (16 + 3)) | | |
110 | (((pixel >> 5) & 0x1F) << (8 + 3)) | | |
111 | (((pixel >> 10) & 0x1F) << 3); | |
112 | } | |
113 | } | |
114 | ||
115 | SDL_Flip(screen); | |
116 | } | |
117 | ||
2bbbb7af | 118 | #ifdef NEON_BUILD |
75e28f62 E |
119 | |
120 | #include <fcntl.h> | |
121 | #include <linux/fb.h> | |
122 | #include <sys/mman.h> | |
123 | #include <sys/ioctl.h> | |
124 | ||
125 | #endif | |
126 | ||
127 | int main(int argc, char *argv[]) | |
128 | { | |
129 | psx_gpu_struct *psx_gpu = &_psx_gpu; | |
f82c6250 | 130 | SDL_Surface *screen = NULL; |
75e28f62 E |
131 | SDL_Event event; |
132 | ||
133 | u32 *list; | |
134 | int size; | |
135 | FILE *state_file; | |
136 | FILE *list_file; | |
137 | u32 no_display = 0; | |
138 | ||
139 | if((argc != 3) && (argc != 4)) | |
140 | { | |
141 | printf("usage:\n%s <state> <list>\n", argv[0]); | |
142 | return 1; | |
143 | } | |
144 | ||
145 | if((argc == 4) && !strcmp(argv[3], "-n")) | |
146 | no_display = 1; | |
147 | ||
148 | state_file = fopen(argv[1], "rb"); | |
149 | fread(&state, 1, sizeof(gpu_dump_struct), state_file); | |
150 | fclose(state_file); | |
151 | ||
152 | list_file = fopen(argv[2], "rb"); | |
153 | ||
154 | fseek(list_file, 0, SEEK_END); | |
155 | size = ftell(list_file); | |
156 | fseek(list_file, 0, SEEK_SET); | |
157 | //size = 0; | |
3867c6ef | 158 | |
75e28f62 E |
159 | list = malloc(size); |
160 | fread(list, 1, size, list_file); | |
161 | fclose(list_file); | |
162 | ||
163 | if(no_display == 0) | |
164 | { | |
165 | SDL_Init(SDL_INIT_EVERYTHING); | |
166 | screen = SDL_SetVideoMode(1024, 512, 32, 0); | |
f82c6250 | 167 | if (screen == 0) |
168 | { | |
169 | printf("can't set video mode: %s\n", SDL_GetError()); | |
170 | return 1; | |
171 | } | |
75e28f62 | 172 | } |
75e28f62 | 173 | |
2bbbb7af | 174 | #ifdef NEON_BUILD |
f82c6250 | 175 | u16 *vram_ptr; |
176 | #if 0 | |
75e28f62 E |
177 | system("ofbset -fb /dev/fb1 -mem 6291456 -en 0"); |
178 | u32 fbdev_handle = open("/dev/fb1", O_RDWR); | |
3867c6ef | 179 | vram_ptr = (mmap((void *)0x50000000, 1024 * 1024 * 2, PROT_READ | PROT_WRITE, |
75e28f62 | 180 | MAP_SHARED | 0xA0000000, fbdev_handle, 0)); |
f82c6250 | 181 | #elif 1 |
182 | #ifndef MAP_HUGETLB | |
183 | #define MAP_HUGETLB 0x40000 /* arch specific */ | |
184 | #endif | |
185 | vram_ptr = (mmap((void *)0x50000000, 1024 * 1024 * 2, PROT_READ | PROT_WRITE, | |
186 | MAP_ANONYMOUS | MAP_PRIVATE | MAP_HUGETLB, -1, 0)); | |
187 | #else | |
188 | vram_ptr = (mmap((void *)0x50000000, 1024 * 1024 * 2, PROT_READ | PROT_WRITE, | |
189 | MAP_ANONYMOUS | MAP_PRIVATE, -1, 0)); | |
190 | #endif | |
191 | if (vram_ptr == MAP_FAILED) | |
192 | { | |
193 | perror("mmap"); | |
194 | return 1; | |
195 | } | |
3867c6ef | 196 | vram_ptr += 64; |
75e28f62 | 197 | |
87c45ad1 | 198 | initialize_psx_gpu(psx_gpu, vram_ptr); |
3867c6ef E |
199 | #else |
200 | initialize_psx_gpu(psx_gpu, _vram + 64); | |
201 | #endif | |
75e28f62 | 202 | |
2bbbb7af | 203 | #ifdef NEON_BUILD |
75e28f62 E |
204 | //triangle_benchmark(psx_gpu); |
205 | //return 0; | |
206 | #endif | |
207 | ||
75e28f62 | 208 | memcpy(psx_gpu->vram_ptr, state.vram, 1024 * 512 * 2); |
75e28f62 E |
209 | |
210 | clear_stats(); | |
211 | ||
2bbbb7af | 212 | #ifdef NEON_BUILD |
75e28f62 E |
213 | init_counter(); |
214 | #endif | |
215 | ||
b243416b | 216 | gpu_parse(psx_gpu, list, size, NULL); |
75e28f62 E |
217 | flush_render_block_buffer(psx_gpu); |
218 | ||
219 | clear_stats(); | |
220 | ||
2bbbb7af | 221 | #ifdef NEON_BUILD |
75e28f62 E |
222 | u32 cycles = get_counter(); |
223 | #endif | |
224 | ||
b243416b | 225 | gpu_parse(psx_gpu, list, size, NULL); |
75e28f62 E |
226 | flush_render_block_buffer(psx_gpu); |
227 | ||
2bbbb7af | 228 | #ifdef NEON_BUILD |
75e28f62 E |
229 | u32 cycles_elapsed = get_counter() - cycles; |
230 | ||
87c45ad1 E |
231 | printf("%-64s: %d\n", argv[1], cycles_elapsed); |
232 | #else | |
233 | printf("%-64s: ", argv[1]); | |
75e28f62 E |
234 | #endif |
235 | ||
87c45ad1 | 236 | #if 1 |
75e28f62 E |
237 | u32 i; |
238 | ||
239 | for(i = 0; i < 1024 * 512; i++) | |
240 | { | |
241 | if((psx_gpu->vram_ptr[i] & 0x7FFF) != (state.vram[i] & 0x7FFF)) | |
242 | { | |
243 | printf("(%d %d %d) vs (%d %d %d) at (%d %d)\n", | |
244 | psx_gpu->vram_ptr[i] & 0x1F, | |
245 | (psx_gpu->vram_ptr[i] >> 5) & 0x1F, | |
246 | (psx_gpu->vram_ptr[i] >> 10) & 0x1F, | |
247 | state.vram[i] & 0x1F, | |
248 | (state.vram[i] >> 5) & 0x1F, | |
249 | (state.vram[i] >> 10) & 0x1F, i % 1024, i / 1024); | |
250 | ||
251 | mismatches++; | |
252 | } | |
253 | else | |
254 | { | |
255 | psx_gpu->vram_ptr[i] = | |
256 | ((psx_gpu->vram_ptr[i] & 0x1F) / 4) | | |
257 | ((((psx_gpu->vram_ptr[i] >> 5) & 0x1F) / 4) << 5) | | |
258 | ((((psx_gpu->vram_ptr[i] >> 10) & 0x1F) / 4) << 10); | |
259 | } | |
260 | } | |
261 | #endif | |
75e28f62 | 262 | |
87c45ad1 | 263 | #if 0 |
75e28f62 | 264 | printf("\n"); |
3867c6ef E |
265 | printf(" %d pixels, %d pixel blocks, %d spans\n" |
266 | " (%lf pixels per block, %lf pixels per span),\n" | |
75e28f62 | 267 | " %lf blocks per span (%lf per non-zero span), %lf overdraw)\n\n", |
3867c6ef | 268 | span_pixels, span_pixel_blocks, spans, |
75e28f62 | 269 | (double)span_pixels / span_pixel_blocks, |
75e28f62 E |
270 | (double)span_pixels / spans, |
271 | (double)span_pixel_blocks / spans, | |
272 | (double)span_pixel_blocks / (spans - zero_block_spans), | |
273 | (double)span_pixels / | |
274 | ((psx_gpu->viewport_end_x - psx_gpu->viewport_start_x) * | |
275 | (psx_gpu->viewport_end_y - psx_gpu->viewport_start_y))); | |
276 | ||
3867c6ef | 277 | printf(" %d triangles\n" |
75e28f62 E |
278 | " (%d trivial rejects, %lf%% flat, %lf%% left split, %lf%% clipped)\n" |
279 | " (%lf pixels per triangle, %lf rows per triangle)\n\n", | |
3867c6ef | 280 | triangles, trivial_rejects, |
75e28f62 E |
281 | percent_of(flat_triangles, triangles), |
282 | percent_of(left_split_triangles, triangles), | |
283 | percent_of(clipped_triangles, triangles), | |
284 | (double)span_pixels / triangles, | |
285 | (double)spans / triangles); | |
286 | ||
287 | printf(" Block data:\n"); | |
288 | printf(" %7d 4bpp texel blocks (%lf%%)\n", texel_blocks_4bpp, | |
289 | percent_of(texel_blocks_4bpp, span_pixel_blocks)); | |
290 | printf(" %7d 8bpp texel blocks (%lf%%)\n", texel_blocks_8bpp, | |
291 | percent_of(texel_blocks_8bpp, span_pixel_blocks)); | |
292 | printf(" %7d 16bpp texel blocks (%lf%%)\n", texel_blocks_16bpp, | |
293 | percent_of(texel_blocks_16bpp, span_pixel_blocks)); | |
294 | printf(" %7d untextured blocks (%lf%%)\n", texel_blocks_untextured, | |
295 | percent_of(texel_blocks_untextured, span_pixel_blocks)); | |
296 | printf(" %7d sprite blocks (%lf%%)\n", sprite_blocks, | |
297 | percent_of(sprite_blocks, span_pixel_blocks)); | |
298 | printf(" %7d blended blocks (%lf%%)\n", blend_blocks, | |
299 | percent_of(blend_blocks, span_pixel_blocks)); | |
3867c6ef E |
300 | printf(" %7d false-mod blocks (%lf%%)\n", false_modulated_blocks, |
301 | percent_of(false_modulated_blocks, span_pixel_blocks)); | |
75e28f62 E |
302 | printf("\n"); |
303 | printf(" %lf blocks per render buffer flush\n", (double)span_pixel_blocks / | |
304 | render_buffer_flushes); | |
305 | printf(" %d zero block spans\n", zero_block_spans); | |
306 | printf(" %d state changes, %d texture cache loads\n", state_changes, | |
307 | texture_cache_loads); | |
308 | if(sprites) | |
309 | { | |
310 | printf(" %d sprites\n" | |
311 | " 4bpp: %lf%%\n" | |
312 | " 8bpp: %lf%%\n" | |
313 | " 16bpp: %lf%%\n" | |
314 | " untextured: %lf%%\n", | |
315 | sprites, percent_of(sprites_4bpp, sprites), | |
316 | percent_of(sprites_8bpp, sprites), percent_of(sprites_16bpp, sprites), | |
317 | percent_of(sprites_untextured, sprites)); | |
318 | } | |
f9248bbf | 319 | printf(" %d lines\n", lines); |
75e28f62 E |
320 | printf("\n"); |
321 | printf(" %d mismatches\n\n\n", mismatches); | |
322 | #endif | |
323 | ||
324 | fflush(stdout); | |
325 | ||
326 | if(no_display == 0) | |
327 | { | |
328 | while(1) | |
329 | { | |
330 | update_screen(psx_gpu, screen); | |
331 | ||
332 | if(SDL_PollEvent(&event)) | |
333 | { | |
334 | if((event.type == SDL_QUIT) || | |
335 | ((event.type == SDL_KEYDOWN) && | |
336 | (event.key.keysym.sym == SDLK_ESCAPE))) | |
337 | { | |
338 | break; | |
339 | } | |
340 | } | |
341 | ||
342 | SDL_Delay(20); | |
343 | } | |
344 | } | |
345 | ||
346 | return (mismatches != 0); | |
347 | } |