From 61b9f2dfb3e20d2e2e7efda30cf459df5134d88f Mon Sep 17 00:00:00 2001 From: notaz Date: Sun, 22 Jun 2014 23:32:05 +0300 Subject: [PATCH] rice: ProcessVertex implemented in NEON --- source/gles2rice/projects/unix/Makefile | 5 + source/gles2rice/src/RenderBase.cpp | 180 ++++++++++++-- source/gles2rice/src/RenderBase.h | 1 + source/gles2rice/src/RenderBase_neon.S | 310 ++++++++++++++++++------ source/gles2rice/src/RenderBase_neon.h | 13 + source/gles2rice/src/arm_features.h | 52 ++++ 6 files changed, 467 insertions(+), 94 deletions(-) create mode 100644 source/gles2rice/src/RenderBase_neon.h create mode 100644 source/gles2rice/src/arm_features.h diff --git a/source/gles2rice/projects/unix/Makefile b/source/gles2rice/projects/unix/Makefile index 948238d..f2c1a61 100755 --- a/source/gles2rice/projects/unix/Makefile +++ b/source/gles2rice/projects/unix/Makefile @@ -332,6 +332,7 @@ SOURCE = \ $(SRCDIR)/OGLTexture.cpp \ $(SRCDIR)/Render.cpp \ $(SRCDIR)/RenderBase.cpp \ + $(SRCDIR)/RenderBase_neon.S \ $(SRCDIR)/RenderExt.cpp \ $(SRCDIR)/RenderTexture.cpp \ $(SRCDIR)/RSP_Parser.cpp \ @@ -362,6 +363,7 @@ endif # generate a list of object files build, make a temporary directory for them OBJECTS := $(patsubst $(SRCDIR)/%.c, $(OBJDIR)/%.o, $(filter %.c, $(SOURCE))) OBJECTS += $(patsubst $(SRCDIR)/%.cpp, $(OBJDIR)/%.o, $(filter %.cpp, $(SOURCE))) +OBJECTS += $(patsubst $(SRCDIR)/%.S, $(OBJDIR)/%.o, $(filter %.S, $(SOURCE))) OBJDIRS = $(dir $(OBJECTS)) $(shell $(MKDIR) $(OBJDIRS)) @@ -421,6 +423,9 @@ CXXFLAGS += $(CFLAGS) $(OBJDIR)/%.o: $(SRCDIR)/%.c $(COMPILE.c) -o $@ $< +$(OBJDIR)/%.o: $(SRCDIR)/%.S + $(COMPILE.c) -o $@ $< + $(OBJDIR)/%.o: $(SRCDIR)/%.cpp $(COMPILE.cc) -o $@ $< diff --git a/source/gles2rice/src/RenderBase.cpp b/source/gles2rice/src/RenderBase.cpp index ee5b76c..429f51d 100644 --- a/source/gles2rice/src/RenderBase.cpp +++ b/source/gles2rice/src/RenderBase.cpp @@ -1528,13 +1528,110 @@ void ProcessVertexDataNoSSE(uint32 dwAddr, uint32 dwV0, uint32 dwNum) DEBUGGER_PAUSE_AND_DUMP(NEXT_VERTEX_CMD,{TRACE0("Paused at Vertex Cmd");}); } +/* NEON code */ + +#include "RenderBase_neon.h" + extern "C" void pv_neon(XVECTOR4 *g_vtxTransformed, XVECTOR4 *g_vecProjected, uint32 *g_dwVtxDifColor, VECTOR2 *g_fVtxTxtCoords, float *g_fFogCoord, uint32 *g_clipFlag2, - uint32 dwNum, const FiddledVtx *vtx, + uint32 dwNum, int neon_state, + const FiddledVtx *vtx, const Light *gRSPlights, const float *fRSPAmbientLightRGBA, const XMATRIX *gRSPworldProject, const XMATRIX *gRSPmodelViewTop, - uint32 gRSPnumLights, float gRSPfFogMin); + uint32 gRSPnumLights, float gRSPfFogMin, + uint32 primitiveColor, uint32 primitiveColor_); + +// debug +//#define DO_CMP +#ifdef DO_CMP +// note: don't forget -fno-associative-math +static XVECTOR4 n_transformed[2], n_projected[2]; +static uint32 n_color[2]; +static VECTOR2 n_vtxcoords[2]; +static float n_fogcoord[2]; +static uint32 n_clipflag2[2]; + +static int do_cmp_f(void *a, void *b, int c) +{ + int *ia = (int *)a, *ib = (int *)b; + for (int i = 0; i < c; i++) { + int di = abs(ia[i] - ib[i]); + if (di > 7) { + printf("di: %d\n", di); + return 1; + } + } + return 0; +} + +static int do_cmp_c(uint32 a, uint32 b) +{ + if (abs(((a >> 0) & 0xff) - ((b >> 0) & 0xff)) > 1) + return 1; + if (abs(((a >> 8) & 0xff) - ((b >> 8) & 0xff)) > 1) + return 1; + if (abs(((a >> 16) & 0xff) - ((b >> 16) & 0xff)) > 1) + return 1; + if (abs(((a >> 24) & 0xff) - ((b >> 24) & 0xff)) > 1) + return 1; + + return 0; +} + +static void do_cmp(int i, int s, int neon_state) +{ + static int ccnt; + int bad = 0; + + // if (memcmp(&n_transformed, &g_vtxTransformed[i], sizeof(XVECTOR4))) + if (do_cmp_f(&n_transformed[s], &g_vtxTransformed[i], 4)) { + printf("transformed:\n%13.8e %13.8e %13.8e %13.8e\n" + "%13.8e %13.8e %13.8e %13.8e\n", + n_transformed[s].x, n_transformed[s].y, + n_transformed[s].z, n_transformed[s].w, + g_vtxTransformed[i].x, g_vtxTransformed[i].y, + g_vtxTransformed[i].z, g_vtxTransformed[i].w); + bad = 1; + } + if (do_cmp_f(&n_projected[s], &g_vecProjected[i], 4)) { + printf("projected:\n%13.8e %13.8e %13.8e %13.8e |%08x\n" + "%13.8e %13.8e %13.8e %13.8e |%08x\n", + n_projected[s].x, n_projected[s].y, + n_projected[s].z, n_projected[s].w, + *(uint32 *)&n_projected[s].w, + g_vecProjected[i].x, g_vecProjected[i].y, + g_vecProjected[i].z, g_vecProjected[i].w, + *(uint32 *)&g_vecProjected[i].w); + bad = 1; + } + if (n_vtxcoords[s].x != g_fVtxTxtCoords[i].x + || n_vtxcoords[s].y != g_fVtxTxtCoords[i].y) + { + printf("vtxcoords:\n%13.8e %13.8e\n%13.8e %13.8e\n", + n_vtxcoords[s].x, n_vtxcoords[s].y, + g_fVtxTxtCoords[i].x, g_fVtxTxtCoords[i].y); + bad = 1; + } + if (n_clipflag2[s] != g_clipFlag2[i]) { + printf("clipflag2: %08x %08x\n", n_clipflag2[s], g_clipFlag2[i]); + bad = 1; + } + if (do_cmp_c(n_color[s], g_dwVtxDifColor[i])) { + printf("n_color: %08x %08x\n", n_color[s], g_dwVtxDifColor[i]); + bad = 1; + } + if (!(neon_state & PV_NEON_ENABLE_SHADE)) + printf("!ENABLE_SHADE!\n"); + if (bad) { + printf("%d s=%d, state %02x\n", ccnt, s, neon_state); + printf(".w %08x %08x\n", + *(uint32 *)&n_projected[s].w, *(uint32 *)&g_vecProjected[i].w); + exit(ccnt); + } + ccnt++; +} +#endif void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) { @@ -1545,14 +1642,10 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) // assumtions: // - g_clipFlag is not used at all + // - g_fFogCoord is not used at all // - g_vtxNonTransformed is not used after ProcessVertexData*() returns // - g_normal - same -#define PV_NEON_ENABLE_LIGHT (1 << 0) -#define PV_NEON_ENABLE_SHADE (1 << 1) -#define PV_NEON_ENABLE_FOG (1 << 2) -#define PV_NEON_FOG_ALPHA (1 << 3) - int neon_state = 0; if ( gRSP.bLightingEnable ) neon_state |= PV_NEON_ENABLE_LIGHT; @@ -1564,6 +1657,9 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) neon_state |= PV_NEON_FOG_ALPHA; uint32 i; +#ifdef DO_CMP + uint32 s = 0; +#endif UpdateCombinedMatrix(); @@ -1578,22 +1674,55 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) // - g_vtxTransformed[i] // - g_dwVtxDifColor[i] -> vertex color // - g_fVtxTxtCoords[i] -> vertex texture cooridinates - // - g_fFogCoord[i] + // - g_fFogCoord[i] -> unused // - g_clipFlag2[i] const FiddledVtx * pVtxBase = (const FiddledVtx*)(g_pRDRAMu8 + dwAddr); g_pVtxBase = (FiddledVtx *)pVtxBase; + gRSPmodelViewTop._14 = gRSPmodelViewTop._24 = + gRSPmodelViewTop._34 = 0; + // SP_Timing(RSP_GBI0_Vtx); status.SPCycleCount += Timing_RSP_GBI0_Vtx * dwNum; - if (!(neon_state & (PV_NEON_ENABLE_LIGHT | PV_NEON_ENABLE_SHADE))) { - for (i = dwV0; i < dwV0 + dwNum; i++) - g_dwVtxDifColor[i] = gRDP.primitiveColor; // FLAT shade - } +//#define DO_CC +#ifdef DO_CC + asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(i)); + i |= 5; // master enable, ccnt reset + i &= ~8; // ccnt divider 0 + asm volatile("mcr p15, 0, %0, c9, c12, 0" :: "r"(i)); + // enable cycle counter + asm volatile("mcr p15, 0, %0, c9, c12, 1" :: "r"(1<<31)); + unsigned int cc_start; + asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(cc_start)); +#endif +#if 1 + i = dwV0; + pv_neon(&g_vtxTransformed[i], &g_vecProjected[i], + &g_dwVtxDifColor[i], &g_fVtxTxtCoords[i], + &g_fFogCoord[i], &g_clipFlag2[i], + dwNum, neon_state, &pVtxBase[i - dwV0], + gRSPlights, gRSP.fAmbientColors, + &gRSPworldProject, &gRSPmodelViewTop, + gRSPnumLights, gRSPfFogMin, + gRDP.primitiveColor, gRDP.primitiveColor); +#else for (i = dwV0; i < dwV0 + dwNum; i++) { +#ifdef DO_CMP + if (!(s & 1)) + pv_neon(n_transformed, n_projected, + n_color, n_vtxcoords, + n_fogcoord, n_clipflag2, + 1, neon_state, &pVtxBase[i - dwV0], + gRSPlights, gRSP.fAmbientColors, + &gRSPworldProject, &gRSPmodelViewTop, + gRSPnumLights, gRSPfFogMin, + gRDP.primitiveColor, gRDP.primitiveColor); +#endif + const FiddledVtx & vert = pVtxBase[i - dwV0]; XVECTOR3 vtx_raw; // was g_vtxNonTransformed @@ -1608,13 +1737,6 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) g_vecProjected[i].y = g_vtxTransformed[i].y * g_vecProjected[i].w; g_vecProjected[i].z = g_vtxTransformed[i].z * g_vecProjected[i].w; - if( neon_state & PV_NEON_ENABLE_FOG ) - { - g_fFogCoord[i] = g_vecProjected[i].z; - if( g_vecProjected[i].w < 0 || g_vecProjected[i].z < 0 || g_fFogCoord[i] < gRSPfFogMin ) - g_fFogCoord[i] = gRSPfFogMin; - } - // RSP_Vtx_Clipping(i); g_clipFlag2[i] = 0; if( g_vecProjected[i].w > 0 ) @@ -1664,6 +1786,8 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) color.r = vert.rgba.b; color.a = vert.rgba.a; } + else + g_dwVtxDifColor[i] = gRDP.primitiveColor; // FLAT shade // ReplaceAlphaWithFogFactor(i); if( neon_state & PV_NEON_FOG_ALPHA ) @@ -1671,7 +1795,8 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) // Use fog factor to replace vertex alpha if( g_vecProjected[i].z > 1 ) *(((uint8*)&(g_dwVtxDifColor[i]))+3) = 0xFF; - if( g_vecProjected[i].z < 0 ) + // missing 'else' in original code?? + else if( g_vecProjected[i].z < 0 ) *(((uint8*)&(g_dwVtxDifColor[i]))+3) = 0; else *(((uint8*)&(g_dwVtxDifColor[i]))+3) = (uint8)(g_vecProjected[i].z*255); @@ -1679,7 +1804,22 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum) g_fVtxTxtCoords[i].x = (float)vert.tu; g_fVtxTxtCoords[i].y = (float)vert.tv; +#ifdef DO_CMP + do_cmp(i, s++ & 1, neon_state); +#endif } +#endif +#ifdef DO_CC + static int total, total_c; + unsigned int cc; + asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(cc)); + total += cc - cc_start; + total_c += dwNum; + if (total_c > 20000) { + printf("%.u\n", total / total_c); + total = total_c = 0; + } +#endif } bool PrepareTriangle(uint32 dwV0, uint32 dwV1, uint32 dwV2) diff --git a/source/gles2rice/src/RenderBase.h b/source/gles2rice/src/RenderBase.h index ceeb385..b2e37bb 100644 --- a/source/gles2rice/src/RenderBase.h +++ b/source/gles2rice/src/RenderBase.h @@ -243,6 +243,7 @@ extern void (*ProcessVertexData)(uint32 dwAddr, uint32 dwV0, uint32 dwNum); void ProcessVertexDataSSE(uint32 dwAddr, uint32 dwV0, uint32 dwNum); #endif void ProcessVertexDataNoSSE(uint32 dwAddr, uint32 dwV0, uint32 dwNum); +void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum); void ProcessVertexDataExternal(uint32 dwAddr, uint32 dwV0, uint32 dwNum); void SetPrimitiveColor(uint32 dwCol, uint32 LODMin, uint32 LODFrac); void SetPrimitiveDepth(uint32 z, uint32 dwDZ); diff --git a/source/gles2rice/src/RenderBase_neon.S b/source/gles2rice/src/RenderBase_neon.S index 77003a5..08df333 100644 --- a/source/gles2rice/src/RenderBase_neon.S +++ b/source/gles2rice/src/RenderBase_neon.S @@ -6,46 +6,24 @@ */ #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: @@ -59,19 +37,21 @@ * 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] @@ -80,60 +60,242 @@ FUNCTION(pv_neon): @ 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 + /* write g_vtxTransformed */ vst1.32 {q0}, [r0, :128]! + vmovl.s16 q9, d18 + /* ... [1] */ vst1.32 {q3}, [r0, :128]! + 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 + /* wrt u,v to g_fVtxTxtCoords */ vst1.32 {d16}, [r3]! + /* ... [1] */ vst1.32 {d18}, [r3]! + 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 - vmul.f32 q7, q0, d2[1] - vshr.u64 d2, #32 - vsli.64 d15, d2, #32 - /* wrt g_vecProjected */ vst1.32 {q7}, [r1]! + 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,q1}, [r1]! @ wrt g_vecProjected + vcgt.f32 d8, d2, d20 + vcgt.f32 d9, d21, d2 + vld1.32 d0[0], [r4]! @ mem: [0] .azyx + vand q3, q11 + vand q4, q11 + 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 + +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 +@ vim:filetype=armasm:expandtab diff --git a/source/gles2rice/src/RenderBase_neon.h b/source/gles2rice/src/RenderBase_neon.h new file mode 100644 index 0000000..0c252d0 --- /dev/null +++ b/source/gles2rice/src/RenderBase_neon.h @@ -0,0 +1,13 @@ + +#define PV_NEON_ENABLE_LIGHT (1 << 0) +#define PV_NEON_ENABLE_SHADE (1 << 1) +#define PV_NEON_ENABLE_FOG (1 << 2) // unused +#define PV_NEON_FOG_ALPHA (1 << 3) + +#define X_CLIP_MAX 0x1 +#define X_CLIP_MIN 0x2 +#define Y_CLIP_MAX 0x4 +#define Y_CLIP_MIN 0x8 + +#define OFFSETOF_Light_fr 0x14 +#define SIZEOF_Light 0x44 diff --git a/source/gles2rice/src/arm_features.h b/source/gles2rice/src/arm_features.h new file mode 100644 index 0000000..fdec522 --- /dev/null +++ b/source/gles2rice/src/arm_features.h @@ -0,0 +1,52 @@ +#ifndef __ARM_FEATURES_H__ +#define __ARM_FEATURES_H__ + +#if defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) \ + || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) \ + || defined(__ARM_ARCH_7EM__) + +#define HAVE_ARMV7 +#define HAVE_ARMV6 +#define HAVE_ARMV5 + +#elif defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \ + || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) \ + || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) + +#define HAVE_ARMV6 +#define HAVE_ARMV5 + +#elif defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_5E__) \ + || defined(__ARM_ARCH_5T__) || defined(__ARM_ARCH_5TE__) || defined(__ARM_ARCH_5TEJ__) + +#define HAVE_ARMV5 + +#endif + +/* no need for HAVE_NEON - GCC defines __ARM_NEON__ consistently */ + +/* global function/external symbol */ +#ifndef __MACH__ +#define ESYM(name) name + +#define FUNCTION(name) \ + .globl name; \ + .type name, %function; \ + name + +#define EXTRA_UNSAVED_REGS + +#else +#define ESYM(name) _##name + +#define FUNCTION(name) \ + .globl ESYM(name); \ + name: \ + ESYM(name) + +// r7 is preserved, but add it for EABI alignment.. +#define EXTRA_UNSAVED_REGS r7, r9, + +#endif + +#endif /* __ARM_FEATURES_H__ */ -- 2.39.5