From 92eab56ae97cb8cd2a198f960b61cc8f1086eaf8 Mon Sep 17 00:00:00 2001 From: Justin Weiss Date: Sun, 1 Mar 2020 15:40:25 -0800 Subject: [PATCH] Preserve uSrc MSB across lighting and blending This saves a few cycles because gcc stores / loads srcMSB when using ARM-optimized lighting. --- plugins/gpu_unai/gpu_inner.h | 62 ++++++++++++++------------ plugins/gpu_unai/gpu_inner_blend.h | 8 ++-- plugins/gpu_unai/gpu_inner_blend_arm.h | 10 ++++- plugins/gpu_unai/gpu_inner_light.h | 34 +++++++------- plugins/gpu_unai/gpu_inner_light_arm.h | 16 ++++--- 5 files changed, 72 insertions(+), 58 deletions(-) diff --git a/plugins/gpu_unai/gpu_inner.h b/plugins/gpu_unai/gpu_inner.h index 197d7ade..83147903 100644 --- a/plugins/gpu_unai/gpu_inner.h +++ b/plugins/gpu_unai/gpu_inner.h @@ -49,26 +49,25 @@ //#endif #include "gpu_inner_blend.h" - -#ifdef __arm__ -#include "gpu_inner_blend_arm.h" -#define gpuBlending gpuBlendingARM -#else -#define gpuBlending gpuBlendingGeneric -#endif - #include "gpu_inner_quantization.h" #include "gpu_inner_light.h" #ifdef __arm__ +#include "gpu_inner_blend_arm.h" #include "gpu_inner_light_arm.h" +#define gpuBlending gpuBlendingARM #define gpuLightingRGB gpuLightingRGBARM #define gpuLightingTXT gpuLightingTXTARM #define gpuLightingTXTGouraud gpuLightingTXTGouraudARM +// Non-dithering lighting and blending functions preserve uSrc +// MSB. This saves a few operations and useless load/stores. +#define MSB_PRESERVED (!CF_DITHER) #else +#define gpuBlending gpuBlendingGeneric #define gpuLightingRGB gpuLightingRGBGeneric #define gpuLightingTXT gpuLightingTXTGeneric #define gpuLightingTXTGouraud gpuLightingTXTGouraudGeneric +#define MSB_PRESERVED 0 #endif @@ -158,10 +157,10 @@ static u8* gpuPixelSpanFn(u8* pDst, uintptr_t data, ptrdiff_t incr, size_t len) else { *(u16*)pDst = col; } } } else { - u16 uDst = *(u16*)pDst; + uint_fast16_t uDst = *(u16*)pDst; if (CF_MASKCHECK) { if (uDst & 0x8000) goto endpixel; } - u16 uSrc = col; + uint_fast16_t uSrc = col; if (CF_BLEND) uSrc = gpuBlending(uSrc, uDst); @@ -184,11 +183,11 @@ static u8* gpuPixelSpanFn(u8* pDst, uintptr_t data, ptrdiff_t incr, size_t len) else { *(u16*)pDst = col; } } } else { - u16 uDst = *(u16*)pDst; + uint_fast16_t uDst = *(u16*)pDst; if (CF_MASKCHECK) { if (uDst & 0x8000) goto endpixel; } col = gpuGouraudColor15bpp(r, g, b); - u16 uSrc = col; + uint_fast16_t uSrc = col; // Blend func can save an operation if it knows uSrc MSB is // unset. For untextured prims, this is always true. @@ -294,7 +293,7 @@ static void gpuTileSpanFn(u16 *pDst, u32 count, u16 data) // unset. For untextured prims, this is always true. const bool skip_uSrc_mask = true; - u16 uSrc, uDst; + uint_fast16_t uSrc, uDst; do { if (CF_MASKCHECK || CF_BLEND) { uDst = *pDst; } @@ -357,10 +356,11 @@ static void gpuSpriteSpanFn(u16 *pDst, u32 count, u8* pTxt, u32 u0) { // Blend func can save an operation if it knows uSrc MSB is unset. // Untextured prims can always skip (source color always comes with MSB=0). - // For textured prims, lighting funcs always return it unset. (bonus!) - const bool skip_uSrc_mask = (!CF_TEXTMODE) || CF_LIGHT; + // For textured prims, the generic lighting funcs always return it unset. (bonus!) + const bool skip_uSrc_mask = MSB_PRESERVED ? (!CF_TEXTMODE) : (!CF_TEXTMODE) || CF_LIGHT; - u16 uSrc, uDst, srcMSB; + uint_fast16_t uSrc, uDst, srcMSB; + bool should_blend; u32 u0_mask = gpu_unai.TextureWindow[2]; u8 r5, g5, b5; @@ -402,12 +402,14 @@ static void gpuSpriteSpanFn(u16 *pDst, u32 count, u8* pTxt, u32 u0) if (CF_LIGHT) uSrc = gpuLightingTXT(uSrc, r5, g5, b5); - if (CF_BLEND && srcMSB) + should_blend = MSB_PRESERVED ? uSrc & 0x8000 : srcMSB; + + if (CF_BLEND && should_blend) uSrc = gpuBlending(uSrc, uDst); - if (CF_MASKSET) { *pDst = uSrc | 0x8000; } - else if (CF_BLEND || CF_LIGHT) { *pDst = uSrc | srcMSB; } - else { *pDst = uSrc; } + if (CF_MASKSET) { *pDst = uSrc | 0x8000; } + else if (!MSB_PRESERVED && (CF_BLEND || CF_LIGHT)) { *pDst = uSrc | srcMSB; } + else { *pDst = uSrc; } endsprite: u0 += (CF_TEXTMODE==3) ? 2 : 1; @@ -484,8 +486,9 @@ static void gpuPolySpanFn(const gpu_unai_t &gpu_unai, u16 *pDst, u32 count) { // Blend func can save an operation if it knows uSrc MSB is unset. // Untextured prims can always skip this (src color MSB is always 0). - // For textured prims, lighting funcs always return it unset. (bonus!) - const bool skip_uSrc_mask = (!CF_TEXTMODE) || CF_LIGHT; + // For textured prims, the generic lighting funcs always return it unset. (bonus!) + const bool skip_uSrc_mask = MSB_PRESERVED ? (!CF_TEXTMODE) : (!CF_TEXTMODE) || CF_LIGHT; + bool should_blend; u32 bMsk; if (CF_BLITMASK) bMsk = gpu_unai.blit_mask; @@ -496,7 +499,7 @@ static void gpuPolySpanFn(const gpu_unai_t &gpu_unai, u16 *pDst, u32 count) // UNTEXTURED, NO GOURAUD const u16 pix15 = gpu_unai.PixelData; do { - u16 uSrc, uDst; + uint_fast16_t uSrc, uDst; // NOTE: Don't enable CF_BLITMASK pixel skipping (speed hack) // on untextured polys. It seems to do more harm than good: see @@ -525,7 +528,7 @@ endpolynotextnogou: u32 l_gInc = gpu_unai.gInc; do { - u16 uDst, uSrc; + uint_fast16_t uDst, uSrc; // See note in above loop regarding CF_BLITMASK //if (CF_BLITMASK) { if ((bMsk>>((((uintptr_t)pDst)>>1)&7))&1) goto endpolynotextgou; } @@ -563,7 +566,7 @@ endpolynotextgou: { // TEXTURED - u16 uDst, uSrc, srcMSB; + uint_fast16_t uDst, uSrc, srcMSB; //senquack - note: original UNAI code had gpu_unai.{u4/v4} packed into // one 32-bit unsigned int, but this proved to lose too much accuracy @@ -650,13 +653,14 @@ endpolynotextgou: uSrc = gpuLightingTXT(uSrc, r5, g5, b5); } - if (CF_BLEND && srcMSB) + should_blend = MSB_PRESERVED ? uSrc & 0x8000 : srcMSB; + if (CF_BLEND && should_blend) uSrc = gpuBlending(uSrc, uDst); } - if (CF_MASKSET) { *pDst = uSrc | 0x8000; } - else if (CF_BLEND || CF_LIGHT) { *pDst = uSrc | srcMSB; } - else { *pDst = uSrc; } + if (CF_MASKSET) { *pDst = uSrc | 0x8000; } + else if (!MSB_PRESERVED && (CF_BLEND || CF_LIGHT)) { *pDst = uSrc | srcMSB; } + else { *pDst = uSrc; } endpolytext: pDst++; l_u = (l_u + l_u_inc) & l_u_msk; diff --git a/plugins/gpu_unai/gpu_inner_blend.h b/plugins/gpu_unai/gpu_inner_blend.h index a4695418..febc7ede 100644 --- a/plugins/gpu_unai/gpu_inner_blend.h +++ b/plugins/gpu_unai/gpu_inner_blend.h @@ -37,14 +37,14 @@ // Where '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// template -GPU_INLINE u16 gpuBlendingGeneric(u16 uSrc, u16 uDst) +GPU_INLINE uint_fast16_t gpuBlendingGeneric(uint_fast16_t uSrc, uint_fast16_t uDst) { // These use Blargg's bitwise modulo-clamping: // http://blargg.8bitalley.com/info/rgb_mixing.html // http://blargg.8bitalley.com/info/rgb_clamped_add.html // http://blargg.8bitalley.com/info/rgb_clamped_sub.html - u16 mix; + uint_fast16_t mix; // 0.5 x Back + 0.5 x Forward if (BLENDMODE==0) { @@ -113,7 +113,7 @@ GPU_INLINE u16 gpuBlendingGeneric(u16 uSrc, u16 uDst) // ^ bit 31 // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u32 gpuGetRGB24(u16 uSrc) +GPU_INLINE u32 gpuGetRGB24(uint_fast16_t uSrc) { return ((uSrc & 0x7C00)<<14) | ((uSrc & 0x03E0)<< 9) @@ -137,7 +137,7 @@ GPU_INLINE u32 gpuGetRGB24(u16 uSrc) // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// template -GPU_INLINE u32 gpuBlending24(u32 uSrc24, u16 uDst) +GPU_INLINE u32 gpuBlending24(u32 uSrc24, uint_fast16_t uDst) { // These use techniques adapted from Blargg's techniques mentioned in // in gpuBlending() comments above. Not as much bitwise trickery is diff --git a/plugins/gpu_unai/gpu_inner_blend_arm.h b/plugins/gpu_unai/gpu_inner_blend_arm.h index 4db105a7..5ddbdbb5 100644 --- a/plugins/gpu_unai/gpu_inner_blend_arm.h +++ b/plugins/gpu_unai/gpu_inner_blend_arm.h @@ -15,7 +15,7 @@ // Where '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// template -GPU_INLINE u16 gpuBlendingARM(u16 uSrc, u16 uDst) +GPU_INLINE uint_fast16_t gpuBlendingARM(uint_fast16_t uSrc, uint_fast16_t uDst) { // These use Blargg's bitwise modulo-clamping: // http://blargg.8bitalley.com/info/rgb_mixing.html @@ -23,7 +23,7 @@ GPU_INLINE u16 gpuBlendingARM(u16 uSrc, u16 uDst) // http://blargg.8bitalley.com/info/rgb_clamped_sub.html - u16 mix; + uint_fast16_t mix; asm ("bic %[uDst], %[uDst], #0x8000" : [uDst] "+r" (uDst)); @@ -89,6 +89,12 @@ GPU_INLINE u16 gpuBlendingARM(u16 uSrc, u16 uDst) : [diff] "=&r" (diff), [mix] "=&r" (mix) : [uSrc] "r" (uSrc), [uDst] "r" (uDst), [mask] "r" (0x8420)); } + + // There's not a case where we can get into this function, + // SKIP_USRC_MSB_MASK is false, and the msb of uSrc is unset. + if (!SKIP_USRC_MSB_MASK) { + asm ("orr %[mix], %[mix], #0x8000" : [mix] "+r" (mix)); + } return mix; } diff --git a/plugins/gpu_unai/gpu_inner_light.h b/plugins/gpu_unai/gpu_inner_light.h index 71d85b14..f90e8ec1 100644 --- a/plugins/gpu_unai/gpu_inner_light.h +++ b/plugins/gpu_unai/gpu_inner_light.h @@ -127,7 +127,7 @@ GPU_INLINE u32 gpuPackGouraudColInc(s32 dr, s32 dg, s32 db) // ^ bit 16 // Where 'r,g,b' are integer bits of colors, 'X' fixed-pt, and '0' zero //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u16 gpuLightingRGBGeneric(u32 gCol) +GPU_INLINE uint_fast16_t gpuLightingRGBGeneric(u32 gCol) { return ((gCol<< 5)&0x7C00) | ((gCol>>11)&0x03E0) | @@ -168,7 +168,7 @@ GPU_INLINE u32 gpuLightingRGB24(u32 gCol) // u16 output: 0bbbbbgggggrrrrr // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u16 gpuLightingTXTGeneric(u16 uSrc, u8 r5, u8 g5, u8 b5) +GPU_INLINE uint_fast16_t gpuLightingTXTGeneric(uint_fast16_t uSrc, u8 r5, u8 g5, u8 b5) { return (gpu_unai.LightLUT[((uSrc&0x7C00)>>5) | b5] << 10) | (gpu_unai.LightLUT[ (uSrc&0x03E0) | g5] << 5) | @@ -190,7 +190,7 @@ GPU_INLINE u16 gpuLightingTXTGeneric(u16 uSrc, u8 r5, u8 g5, u8 b5) // u16 output: 0bbbbbgggggrrrrr // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u16 gpuLightingTXTGouraudGeneric(u16 uSrc, u32 gCol) +GPU_INLINE uint_fast16_t gpuLightingTXTGouraudGeneric(uint_fast16_t uSrc, u32 gCol) { return (gpu_unai.LightLUT[((uSrc&0x7C00)>>5) | ((gCol>> 5)&0x1F)]<<10) | (gpu_unai.LightLUT[ (uSrc&0x03E0) | ((gCol>>16)&0x1F)]<< 5) | @@ -214,15 +214,15 @@ GPU_INLINE u16 gpuLightingTXTGouraudGeneric(u16 uSrc, u32 gCol) // ^ bit 31 // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u32 gpuLightingTXT24(u16 uSrc, u8 r8, u8 g8, u8 b8) +GPU_INLINE u32 gpuLightingTXT24(uint_fast16_t uSrc, u8 r8, u8 g8, u8 b8) { - u16 r1 = uSrc&0x001F; - u16 g1 = uSrc&0x03E0; - u16 b1 = uSrc&0x7C00; + uint_fast16_t r1 = uSrc&0x001F; + uint_fast16_t g1 = uSrc&0x03E0; + uint_fast16_t b1 = uSrc&0x7C00; - u16 r2 = r8; - u16 g2 = g8; - u16 b2 = b8; + uint_fast16_t r2 = r8; + uint_fast16_t g2 = g8; + uint_fast16_t b2 = b8; u32 r3 = r1 * r2; if (r3 & 0xFFFFF000) r3 = ~0xFFFFF000; u32 g3 = g1 * g2; if (g3 & 0xFFFE0000) g3 = ~0xFFFE0000; @@ -249,15 +249,15 @@ GPU_INLINE u32 gpuLightingTXT24(u16 uSrc, u8 r8, u8 g8, u8 b8) // ^ bit 31 // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u32 gpuLightingTXT24Gouraud(u16 uSrc, u32 gCol) +GPU_INLINE u32 gpuLightingTXT24Gouraud(uint_fast16_t uSrc, u32 gCol) { - u16 r1 = uSrc&0x001F; - u16 g1 = uSrc&0x03E0; - u16 b1 = uSrc&0x7C00; + uint_fast16_t r1 = uSrc&0x001F; + uint_fast16_t g1 = uSrc&0x03E0; + uint_fast16_t b1 = uSrc&0x7C00; - u16 r2 = (gCol>>24) & 0xFF; - u16 g2 = (gCol>>13) & 0xFF; - u16 b2 = (gCol>> 2) & 0xFF; + uint_fast16_t r2 = (gCol>>24) & 0xFF; + uint_fast16_t g2 = (gCol>>13) & 0xFF; + uint_fast16_t b2 = (gCol>> 2) & 0xFF; u32 r3 = r1 * r2; if (r3 & 0xFFFFF000) r3 = ~0xFFFFF000; u32 g3 = g1 * g2; if (g3 & 0xFFFE0000) g3 = ~0xFFFE0000; diff --git a/plugins/gpu_unai/gpu_inner_light_arm.h b/plugins/gpu_unai/gpu_inner_light_arm.h index ac99cf32..a1c36286 100644 --- a/plugins/gpu_unai/gpu_inner_light_arm.h +++ b/plugins/gpu_unai/gpu_inner_light_arm.h @@ -12,9 +12,9 @@ // ^ bit 16 // Where 'r,g,b' are integer bits of colors, 'X' fixed-pt, and '0' zero //////////////////////////////////////////////////////////////////////////////// -GPU_INLINE u16 gpuLightingRGBARM(u32 gCol) +GPU_INLINE uint_fast16_t gpuLightingRGBARM(u32 gCol) { - u16 out = 0x03E0; // don't need the mask after starting to write output + uint_fast16_t out = 0x03E0; // don't need the mask after starting to write output u32 tmp; asm ("and %[tmp], %[gCol], %[out]\n\t" // tmp holds 0x000000bbbbb00000 @@ -29,9 +29,9 @@ GPU_INLINE u16 gpuLightingRGBARM(u32 gCol) } -GPU_INLINE u16 gpuLightingTXTARM(u16 uSrc, u8 r5, u8 g5, u8 b5) +GPU_INLINE uint_fast16_t gpuLightingTXTARM(uint_fast16_t uSrc, u8 r5, u8 g5, u8 b5) { - u16 out = 0x03E0; + uint_fast16_t out = 0x03E0; u32 db, dg; asm ("and %[dg], %[out], %[src] \n\t" "orr %[dg], %[dg], %[g5] \n\t" @@ -42,7 +42,9 @@ GPU_INLINE u16 gpuLightingTXTARM(u16 uSrc, u8 r5, u8 g5, u8 b5) "orr %[db], %[db], %[b5] \n\t" "ldrb %[out], [%[lut], %[out]] \n\t" "ldrb %[db], [%[lut], %[db]] \n\t" + "tst %[src], #0x8000\n\t" "orr %[out], %[out], %[dg], lsl #0x05 \n\t" + "orrne %[out], %[out], #0x8000\n\t" "orr %[out], %[out], %[db], lsl #0x0A \n\t" : [out] "=&r" (out), [db] "=&r" (db), [dg] "=&r" (dg) : [r5] "r" (r5), [g5] "r" (g5), [b5] "r" (b5), @@ -51,9 +53,9 @@ GPU_INLINE u16 gpuLightingTXTARM(u16 uSrc, u8 r5, u8 g5, u8 b5) return out; } -GPU_INLINE u16 gpuLightingTXTGouraudARM(u16 uSrc, u32 gCol) +GPU_INLINE uint_fast16_t gpuLightingTXTGouraudARM(uint_fast16_t uSrc, u32 gCol) { - u16 out = 0x03E0; // don't need the mask after starting to write output + uint_fast16_t out = 0x03E0; // don't need the mask after starting to write output u32 db,dg,gtmp; asm ("and %[dg], %[out], %[src] \n\t" "and %[gtmp],%[out], %[gCol], lsr #0x0B \n\t" @@ -66,7 +68,9 @@ GPU_INLINE u16 gpuLightingTXTGouraudARM(u16 uSrc, u32 gCol) "orr %[db], %[db], %[gtmp], lsr #0x05 \n\t" "ldrb %[out], [%[lut], %[out]] \n\t" "ldrb %[db], [%[lut], %[db]] \n\t" + "tst %[src], #0x8000\n\t" "orr %[out], %[out], %[dg], lsl #0x05 \n\t" + "orrne %[out], %[out], #0x8000\n\t" "orr %[out], %[out], %[db], lsl #0x0A \n\t" : [out] "=&r" (out), [db] "=&r" (db), [dg] "=&r" (dg), [gtmp] "=&r" (gtmp) \ -- 2.39.2