*/
#include "arm_features.h"
+#include "RenderBase_neon.h"
.syntax unified
.text
.align 3
-.macro do_mac_flags rr1 rr2 rr3
- cmp \rr1, #1
-.endm
-
-
/*
* ProcessVertexData register map:
*
* q | d | c code
- * 0 0
- * 1
- * 1 2
- * 3
- * 2 4
- * 5
- * 3 6
- * 7
- * 4 8
- * 9
- * 5 10
- * 11
- * 6 12
- * 13
- * 7 14 g_vecProjected
- * 15
- * 8 16
- * 17
* ...
- * 12 24 gRSPworldProject _11,_21,_31,_41
+ * 12 24 gRSPworldProject _11,_12,_13,_14
* 25
- * 13 26 gRSPworldProject _12,_22,_32,_42
+ * 13 26 gRSPworldProject _21,_22,_23,_24
* 27
- * 14 28 gRSPworldProject _13,_23,_33,_43
+ * 14 28 gRSPworldProject _31,_32,_33,_34
* 29
- * 15 30 gRSPworldProject _14,_24,_34,_44
+ * 15 30 gRSPworldProject _41,_42,_43,_44
* 31
*
* r4 vtx[], 16 bytes:
* r2 - uint32 *g_dwVtxDifColor
* r3 - VECTOR2 *g_fVtxTxtCoords
* sp+00 - float *g_fFogCoord
- * sp+04 - uint32 *g_clipFlag2
+ * r6 sp+04 - uint32 *g_clipFlag2
* inputs:
* r11 sp+08 - uint32 dwNum
- * sp+0c - int neon_flags
- * r4 sp+10 - FiddledVtx vtx[]
- * sp+14 - Light *gRSPlights
+ * r10 sp+0c - int neon_flags
+ * r4 sp+10 - FiddledVtx vtx[], (r4 [0], r5 [1])
+ * r7 sp+14 - Light *gRSPlights
* sp+18 - float *fRSPAmbientLightRGBA
- * sp+1c - XMATRIX *gRSPworldProjectTransported
+ * sp+1c - XMATRIX *gRSPworldProject
* sp+20 - XMATRIX *gRSPmodelViewTop
* sp+24 - uint32 gRSPnumLights
* sp+28 - float gRSPfFogMin
+ * sp+2c - uint32 primitiveColor
+ * sp+30 - uint32 primitiveColor
*/
-FUNCTION(pv_neon): @
+FUNCTION(pv_neon):
ldr r12, [sp, #0x10]
pld [r12]
mov r4, r12 @ vtx
ldr r12, [sp, #0x64+0x1c]
- ldr r11, [sp, #0x64+0x08]
vld1.32 {q12,q13}, [r12, :128]! @ load gRSPworldProject
vld1.32 {q14,q15}, [r12, :128]
+ ldr r6, [sp, #0x64+0x04] @ g_clipFlag2
+ add r5, r4, #16 @ vtx + 1
+ ldr r11, [sp, #0x64+0x08] @ dwNum
+ ldr r10, [sp, #0x64+0x0c] @ neon_flags
0:
- vld1.16 d4[1], [r4]! @ y
- vld1.16 d4[0], [r4]! @ x
- vld1.16 d4[3], [r4]! @ flag
- vld1.16 d4[2], [r4]! @ z
- vld1.16 d5[1], [r4]! @ v
- vld1.16 d5[0], [r4]! @ u
- vmovl.s16 q0, d4
- vmovl.s16 q1, d5
- mov r12, #0x3f800000 @ 1.0f
- vcvt.f32.s32 q2, q0 @ q2 = vtx_raw
- vcvt.f32.s32 q3, q1 @ d6 = float u, v
- vmov.32 d5[1], r12 @ q2 = { x, y, x, 1.0f }
- vmul.f32 q4, q2, q12
- vmul.f32 q5, q2, q13
- vmul.f32 q6, q2, q14
- vmul.f32 q7, q2, q15
- /* wrt u,v to g_fVtxTxtCoords */ vst1.32 {d6}, [r3]!
- vpadd.f32 d7, d14, d15
- vpadd.f32 d6, d12, d13
- vpadd.f32 d5, d10, d11
- vpadd.f32 d4, d8, d9
- vpadd.f32 d1, d6, d7 @ g_vtxTransformed .z .w
- vpadd.f32 d0, d4, d5 @ g_vtxTransformed .x .y
+ vld1.16 d12, [r4]! @ vtx[0] .z .flag .x .y (reg)
+ vmovl.s16 q6, d12
+ vld1.16 d14, [r5]! @ vtx[1] .z .flag .x .y
+ vmovl.s16 q7, d14
+ vcvt.f32.s32 q6, q6 @ q6 = vtx_raw0
+ vcvt.f32.s32 q7, q7 @ q7 = vtx_raw1
+ vdup.32 q0, d12[1] @ vtx_raw0.x (dup)
+ vdup.32 q1, d12[0] @ vtx_raw0.y (dup)
+ vdup.32 q2, d13[1] @ vtx_raw0.z (dup)
+ vdup.32 q3, d14[1] @ vtx_raw1.x (dup)
+ vdup.32 q4, d14[0] @ vtx_raw1.y (dup)
+ vdup.32 q5, d15[1] @ vtx_raw1.z (dup)
+ /* note: order of operations matters greatly,
+ * may cause like 20 fraction bits to differ! */
+ vmul.f32 q0, q0, q12
+ vmul.f32 q3, q3, q12
+ vmla.f32 q0, q1, q13
+ vmla.f32 q3, q4, q13
+ vmul.f32 q2, q2, q14 @ yes, mul+add is
+ vmul.f32 q5, q5, q14 @ faster than mla
+ vadd.f32 q0, q2
+ vadd.f32 q3, q5
+ vadd.f32 q0, q15 @ q0 = g_vtxTransformed[i]
+ vadd.f32 q3, q15 @ q3 = g_vtxTransformed[i + 1]
+ vld1.16 d16[1], [r4]! @ [0].v
+ vmov d2, d1
+ vld1.16 d16[0], [r4]! @ [0].u
+ vsri.64 d2, d7, #32
+ vld1.16 d18[1], [r5]! @ [0].v
#if 1
- vrecpe.f32 d2, d1 @ inv .z(unused) .w
- vrecps.f32 d3, d1, d2 @ step
- /* wrt g_vtxTransformed */ vst1.32 {q0}, [r0]!
- vmul.f32 d2, d3, d2 @ better inv
- vrecps.f32 d3, d1, d2 @ step
- vmul.f32 d2, d3, d2 @ better inv
+ vrecpe.f32 d4, d2 @ inv [0][1] .w
+ vld1.16 d18[0], [r5]! @ [0].u
+ vrecps.f32 d5, d2, d4 @ step
+ vmovl.s16 q8, d16
+ /* g_vtxTransformed[0] */ vst1.32 {q0}, [r0, :128]!
+ vmovl.s16 q9, d18
+ vcvt.f32.s32 d16, d16
+ vcvt.f32.s32 d18, d18
+ vmul.f32 d4, d5, d4 @ better inv
+ bic r9, r5, #63
+ pld [r9, #64]
+ vrecps.f32 d5, d2, d4 @ step
+ cmp r11, #1
+ /* u,v g_fVtxTxtCoords[0] */ vst1.32 {d16}, [r3]!
+ beq 99f
+ /* g_vtxTransformed[1] */ vst1.32 {q3}, [r0, :128]!
+ /* ... [1] */ vst1.32 {d18}, [r3]!
+ 99:
+ vmov.f32 d20, #1.0
+ vmov.f32 d21, #-1.0
+ vmul.f32 d4, d5, d4 @ better inv [0][1] .w
#if 0
- vrecps.f32 d3, d1, d2 @ step
- vmul.f32 d2, d3, d2 @ better inv
+ vrecps.f32 d5, d2, d4 @ step
+ vmul.f32 d4, d5, d4 @ better inv
#endif
#else
- vmov.f32 s5, r12
- /* wrt g_vtxTransformed */ vst1.32 {q0}, [r0]!
- vdiv.f32 s5, s5, s3
+ mov r12, #0x3f800000 @ 1.0f
+ vmov.f32 s6, r12
+ vdiv.f32 s8, s6, s4
+ vdiv.f32 s9, s6, s5
+ #error incomplete
+#endif
+
+ mov r8, #X_CLIP_MAX
+ mov r9, #Y_CLIP_MAX
+ vmov d22, r8, r9
+ vmul.f32 q0, q0, d4[1] @ .x .y .z .w *= [0] .w
+ vmul.f32 q1, q3, d4[0]
+ vshr.u64 d5, d4, #32 @ [0] .w
+ mov r8, #X_CLIP_MIN
+ mov r9, #Y_CLIP_MIN
+ vmov d23, r8, r9
+ vsli.64 d3, d4, #32 @ insert [1] .w
+ vsli.64 d1, d5, #32
+ vsli.u64 d5, d4, #32 @ [0] [1] .w
+ vcgt.f32 d6, d0, d20 @ .xy > 1.0?
+ vcgt.f32 d7, d21, d0
+ vcgt.f32 d4, d5, #0 @ .w > 0?
+ vst1.32 {q0}, [r1]! @ g_vecProjected[0]
+ vcgt.f32 d8, d2, d20
+ vcgt.f32 d9, d21, d2
+ vld1.32 d0[0], [r4]! @ mem: [0] .azyx
+ vand q3, q11
+ vand q4, q11
+ cmp r11, #1
+ beq 99f
+ vst1.32 {q1}, [r1]! @ g_vecProjected[1]
+99:
+ vorr d6, d6, d7
+ vorr d7, d8, d9
+ vld1.32 d0[1], [r5]! @ mem: [1] .azyx
+ vpadd.u32 d6, d7
+ vrev32.8 d0, d0 @ make 0xaazzyyxx [1][0]
+ vsli.u64 d1, d3, #32 @ d3 = [1] [0] .z
+ vmovl.s8 q4, d0
+ vand d6, d4
+ vmovl.s16 q1, d8
+ vmovl.s16 q2, d9
+ vst1.32 {d6}, [r6]! @ g_clipFlag2
+
+ tst r10, #PV_NEON_ENABLE_LIGHT
+ beq pv_neon_no_light
+@ pv_neon_light:
+ @ live NEON registers:
+ @ d1 = [1][0] .z (must preserve)
+ @ q1,q2 = azyx [1][0]
+ @ q12+ = gRSPworldProject
+ ldr r12, [sp, #0x64+0x20]
+ vcvt.f32.s32 q1, q1
+ vcvt.f32.s32 q2, q2
+ vld1.32 {q8,q9}, [r12, :128]! @ load gRSPmodelViewTop
+ vld1.32 {q10}, [r12, :128]
+
+ vdup.32 q5, d4[0] @ [1] .x (dup)
+ vdup.32 q6, d4[1] @ [1] .y (dup)
+ vdup.32 q7, d5[0] @ [1] .z (dup)
+ vdup.32 q2, d2[0] @ [0] .x (dup)
+ vdup.32 q3, d2[1] @ [0] .y (dup)
+ vdup.32 q4, d3[0] @ [0] .z (dup)
+ vmul.f32 q2, q2, q8
+ vmul.f32 q5, q5, q8
+ vmla.f32 q2, q3, q9
+ vmla.f32 q5, q6, q9
+ vmul.f32 q4, q4, q10
+ vmul.f32 q7, q7, q10
+ vadd.f32 q4, q2 @ q4 = temp[0] .xyz0
+ vadd.f32 q5, q7 @ q5 = temp[1] .xyz0
+ vmul.f32 q2, q4, q4 @ temp .xyz0 ^2
+ vmul.f32 q3, q5, q5
+ vpadd.f32 d2, d4, d5
+ vpadd.f32 d3, d6, d7
+ movw r8, #0x0000ffff
+ movt r8, #0x7f7f @ max normal float, ~3.4e+38
+ vdup.32 d4, r8
+ vpadd.f32 d2, d2, d3 @ d2 = [1][0] x^2 + y^2 + z^2
+ vcgt.f32 d5, d2, #0
+ vbif d2, d4, d5 @ if (d2 == 0) d2 = MAXFLOAT
+
+ vrsqrte.f32 d3, d2 @ ~ 1/sqrt(d2), d2 = [1][0] .sqrsum
+ vmul.f32 d4, d3, d2
+ ldr r9, [sp, #0x64+0x18] @ &fRSPAmbientLightRGBA
+ ldr r7, [sp, #0x64+0x14] @ gRSPlights
+ ldr r8, [sp, #0x64+0x24] @ gRSPnumLights
+ vrsqrts.f32 d4, d3, d4 @ step
+ vld1.32 {q6}, [r9] @ rgb
+ vld1.32 {q7}, [r9] @ rgb
+ vmul.f32 d3, d3, d4 @ 1/sqrt(d2)
+#if 0 /* not necessary? */
+ vmul.f32 d4, d3, d2
+ vrsqrts.f32 d4, d3, d4 @ step
+ vmul.f32 d3, d3, d4 @ 1/sqrt(d2)
#endif
+ vmul.f32 q2, q4, d3[0] @ q2 = normal[0] .xyz
+ vmul.f32 q3, q5, d3[1] @ q3 = normal[1] .xyz
- vmul.f32 q7, q0, d2[1]
- vshr.u64 d2, #32
- vsli.64 d15, d2, #32
- /* wrt g_vecProjected */ vst1.32 {q7}, [r1]!
+1:
+ vld1.32 {q8}, [r7]
+ vmul.f32 q4, q8, q2 @ gRSPlights[l] * normal
+ vmul.f32 q5, q8, q3
+ vpadd.f32 d8, d8, d9
+ vpadd.f32 d10, d10, d11
+ vpadd.f32 d8, d8, d10 @ d8 = [1][0] fCosT
+ vcgt.f32 d9, d8, #0 @ if (!(fCosT > 0))
+ vand d8, d9 @ fCosT = 0
+ add r9, r7, #OFFSETOF_Light_fr
+ vld1.32 {q8}, [r9] @ .fr .fg .fb
+ vdup.32 q5, d8[1] @ [1] fCosT (dup)
+ vdup.32 q4, d8[0] @
+ vmla.f32 q7, q8, q5 @ .rgb += frgb * fCosT
+ vmla.f32 q6, q8, q4
+ add r7, #SIZEOF_Light
+ subs r8, #1
+ bgt 1b
+
+ movt r8, #0x437f @ float 255
+ vdup.32 q8, r8
+ vcgt.f32 q4, q6, q8 @ if (.rgb > 255)
+ vcgt.f32 q5, q7, q8
+ vbit q6, q8, q4 @ .rgb = 255
+ vbit q7, q8, q5
+ vcvt.u32.f32 q6, q6
+ vcvt.u32.f32 q7, q7
+ ldrb r8, [r4, #-4] @ .a from vtx
+ ldrb r9, [r5, #-4]
+ vext.32 q4, q6, q6, #3 @ reg: .abgr -> .bgra
+ vext.32 q5, q7, q7, #3
+ vmov.32 d8[0], r8 @ use .a from input
+ vmov.32 d10[0], r9
+ vmovn.u32 d8, q4
+ vmovn.u32 d10, q5
+ vmovn.u16 d0, q4
+ vmovn.u16 d2, q5
+ vsli.u64 d0, d2, #32
+ vrev32.8 d0, d0 @ 0xbbggrraa -> 0xaarrggbb
+ b pv_neon_fog_alpha
+
+pv_neon_no_light:
+ tst r10, #PV_NEON_ENABLE_SHADE
+ vldr d0, [sp, #0x64+0x2c] @ primitiveColor [0] [1]
+ beq pv_neon_fog_alpha
+ @ easier to do with ARM
+ ldr r8, [r4, #-4]
+ ldr r9, [r5, #-4]
+ ror r8, #8 @ mem: .argb -> .rgba
+ ror r9, #8 @ reg: 0xbbggrraa -> ..
+ vmov d0, r8, r9
+
+pv_neon_fog_alpha:
+ tst r10, #PV_NEON_FOG_ALPHA
+ beq pv_neon_next
+ vmov.f32 d20, #1.0
+ vcgt.f32 d2, d1, d20 @ [0] [1] .z > 1.0?
+ vcgt.f32 d3, d1, #0 @ > 0?
+ movw r8, #0
+ movt r8, #0x4f7f @ r8 = (float)(255<<24)
+ vbit d1, d20, d2 @ make 1.0 if needed
+ vand d1, d3
+ vdup.32 d4, r8
+ vmul.f32 d1, d1, d4
+ vcvt.u32.f32 d1, d1
+ vmov.u32 d5, #0xff000000
+ vbit d0, d1, d5
+
+pv_neon_next:
+ subs r11, #2
+ vst1.32 {d0}, [r2]! @ g_dwVtxDifColor
+ add r4, #16
+ add r5, #16
+ bgt 0b
+ nop
vpop {q4-q7}
pop {r4-r11,pc}
- .size pv_neon, .-pv_neon
+ .size pv_neon, .-pv_neon
-@ vim:filetype=armasm
+@ (float *d, const float *m1, const float *m2, const float *s)
+FUNCTION(multiply_subtract2):
+ vld1.32 {d1}, [r1]
+ vld1.32 {d2}, [r2]
+ vmul.f32 d0, d1, d2
+ vld1.32 {d3}, [r3]
+ vsub.f32 d0, d3
+ vst1.32 {d0}, [r0]
+ bx lr
+ .size multiply_subtract2, .-multiply_subtract2
+
+
+@ (const XVECTOR4 *v0, const XVECTOR4 *v1, const XVECTOR4 *v2)
+FUNCTION(tv_direction):
+ vld1.32 {q0}, [r0]
+ vld1.32 {q2}, [r2]
+ vld1.32 {q1}, [r1]
+ vsub.f32 d6, d4, d0 @ d6 = V2,V1
+ vsub.f32 d7, d4, d2 @ d7 = W2,W1
+ vmul.f32 d1, d5 @ d1 = v0.w * v2.w
+ vrev64.32 d7, d7
+ vmul.f32 d6, d7 @ d6 = V2*W1,V1*W2
+ vmul.f32 d1, d3 @ d1 *= v1.w
+ vshr.u64 d7, d6, #32
+ vsub.f32 d6, d7 @ d6[0] = V1*W2 - V2*W1
+ vshr.u64 d1, d1, #32
+ vmul.f32 d0, d1, d6
+ vmov.32 r0, d0[0]
+ bx lr
+
+
+@ vim:filetype=armasm:expandtab