michael@0: /*
michael@0:  * Copyright 2009 The Android Open Source Project
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: #include "SkBitmapProcState_opts_SSE2.h"
michael@0: #include "SkBitmapProcState_opts_SSSE3.h"
michael@0: #include "SkBitmapFilter_opts_SSE2.h"
michael@0: #include "SkBlitMask.h"
michael@0: #include "SkBlitRow.h"
michael@0: #include "SkBlitRect_opts_SSE2.h"
michael@0: #include "SkBlitRow_opts_SSE2.h"
michael@0: #include "SkBlurImage_opts_SSE2.h"
michael@0: #include "SkUtils_opts_SSE2.h"
michael@0: #include "SkUtils.h"
michael@0: #include "SkMorphology_opts.h"
michael@0: #include "SkMorphology_opts_SSE2.h"
michael@0: 
michael@0: #include "SkRTConf.h"
michael@0: 
michael@0: #if defined(_MSC_VER) && defined(_WIN64)
michael@0: #include <intrin.h>
michael@0: #endif
michael@0: 
michael@0: /* This file must *not* be compiled with -msse or -msse2, otherwise
michael@0:    gcc may generate sse2 even for scalar ops (and thus give an invalid
michael@0:    instruction on Pentium3 on the code below).  Only files named *_SSE2.cpp
michael@0:    in this directory should be compiled with -msse2. */
michael@0: 
michael@0: 
michael@0: #ifdef _MSC_VER
michael@0: static inline void getcpuid(int info_type, int info[4]) {
michael@0: #if defined(_WIN64)
michael@0:     __cpuid(info, info_type);
michael@0: #else
michael@0:     __asm {
michael@0:         mov    eax, [info_type]
michael@0:         cpuid
michael@0:         mov    edi, [info]
michael@0:         mov    [edi], eax
michael@0:         mov    [edi+4], ebx
michael@0:         mov    [edi+8], ecx
michael@0:         mov    [edi+12], edx
michael@0:     }
michael@0: #endif
michael@0: }
michael@0: #else
michael@0: #if defined(__x86_64__)
michael@0: static inline void getcpuid(int info_type, int info[4]) {
michael@0:     asm volatile (
michael@0:         "cpuid \n\t"
michael@0:         : "=a"(info[0]), "=b"(info[1]), "=c"(info[2]), "=d"(info[3])
michael@0:         : "a"(info_type)
michael@0:     );
michael@0: }
michael@0: #else
michael@0: static inline void getcpuid(int info_type, int info[4]) {
michael@0:     // We save and restore ebx, so this code can be compatible with -fPIC
michael@0:     asm volatile (
michael@0:         "pushl %%ebx      \n\t"
michael@0:         "cpuid            \n\t"
michael@0:         "movl %%ebx, %1   \n\t"
michael@0:         "popl %%ebx       \n\t"
michael@0:         : "=a"(info[0]), "=r"(info[1]), "=c"(info[2]), "=d"(info[3])
michael@0:         : "a"(info_type)
michael@0:     );
michael@0: }
michael@0: #endif
michael@0: #endif
michael@0: 
michael@0: #if defined(__x86_64__) || defined(_WIN64) || SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
michael@0: /* All x86_64 machines have SSE2, or we know it's supported at compile time,  so don't even bother checking. */
michael@0: static inline bool hasSSE2() {
michael@0:     return true;
michael@0: }
michael@0: #else
michael@0: 
michael@0: static inline bool hasSSE2() {
michael@0:     int cpu_info[4] = { 0 };
michael@0:     getcpuid(1, cpu_info);
michael@0:     return (cpu_info[3] & (1<<26)) != 0;
michael@0: }
michael@0: #endif
michael@0: 
michael@0: #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
michael@0: /* If we know SSSE3 is supported at compile time, don't even bother checking. */
michael@0: static inline bool hasSSSE3() {
michael@0:     return true;
michael@0: }
michael@0: #else
michael@0: 
michael@0: static inline bool hasSSSE3() {
michael@0:     int cpu_info[4] = { 0 };
michael@0:     getcpuid(1, cpu_info);
michael@0:     return (cpu_info[2] & 0x200) != 0;
michael@0: }
michael@0: #endif
michael@0: 
michael@0: static bool cachedHasSSE2() {
michael@0:     static bool gHasSSE2 = hasSSE2();
michael@0:     return gHasSSE2;
michael@0: }
michael@0: 
michael@0: static bool cachedHasSSSE3() {
michael@0:     static bool gHasSSSE3 = hasSSSE3();
michael@0:     return gHasSSSE3;
michael@0: }
michael@0: 
michael@0: SK_CONF_DECLARE( bool, c_hqfilter_sse, "bitmap.filter.highQualitySSE", false, "Use SSE optimized version of high quality image filters");
michael@0: 
michael@0: void SkBitmapProcState::platformConvolutionProcs(SkConvolutionProcs* procs) {
michael@0:     if (cachedHasSSE2()) {
michael@0:         procs->fExtraHorizontalReads = 3;
michael@0:         procs->fConvolveVertically = &convolveVertically_SSE2;
michael@0:         procs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_SSE2;
michael@0:         procs->fConvolveHorizontally = &convolveHorizontally_SSE2;
michael@0:         procs->fApplySIMDPadding = &applySIMDPadding_SSE2;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void SkBitmapProcState::platformProcs() {
michael@0:     if (cachedHasSSSE3()) {
michael@0:         if (fSampleProc32 == S32_opaque_D32_filter_DX) {
michael@0:             fSampleProc32 = S32_opaque_D32_filter_DX_SSSE3;
michael@0:         } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
michael@0:             fSampleProc32 = S32_alpha_D32_filter_DX_SSSE3;
michael@0:         }
michael@0: 
michael@0:         if (fSampleProc32 == S32_opaque_D32_filter_DXDY) {
michael@0:             fSampleProc32 = S32_opaque_D32_filter_DXDY_SSSE3;
michael@0:         } else if (fSampleProc32 == S32_alpha_D32_filter_DXDY) {
michael@0:             fSampleProc32 = S32_alpha_D32_filter_DXDY_SSSE3;
michael@0:         }
michael@0:     } else if (cachedHasSSE2()) {
michael@0:         if (fSampleProc32 == S32_opaque_D32_filter_DX) {
michael@0:             fSampleProc32 = S32_opaque_D32_filter_DX_SSE2;
michael@0:         } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
michael@0:             fSampleProc32 = S32_alpha_D32_filter_DX_SSE2;
michael@0:         }
michael@0: 
michael@0:         if (fSampleProc16 == S32_D16_filter_DX) {
michael@0:             fSampleProc16 = S32_D16_filter_DX_SSE2;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     if (cachedHasSSSE3() || cachedHasSSE2()) {
michael@0:         if (fMatrixProc == ClampX_ClampY_filter_scale) {
michael@0:             fMatrixProc = ClampX_ClampY_filter_scale_SSE2;
michael@0:         } else if (fMatrixProc == ClampX_ClampY_nofilter_scale) {
michael@0:             fMatrixProc = ClampX_ClampY_nofilter_scale_SSE2;
michael@0:         }
michael@0: 
michael@0:         if (fMatrixProc == ClampX_ClampY_filter_affine) {
michael@0:             fMatrixProc = ClampX_ClampY_filter_affine_SSE2;
michael@0:         } else if (fMatrixProc == ClampX_ClampY_nofilter_affine) {
michael@0:             fMatrixProc = ClampX_ClampY_nofilter_affine_SSE2;
michael@0:         }
michael@0:         if (c_hqfilter_sse) {
michael@0:             if (fShaderProc32 == highQualityFilter32) {
michael@0:                 fShaderProc32 = highQualityFilter_SSE2;
michael@0:             }
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: static SkBlitRow::Proc platform_16_procs[] = {
michael@0:     S32_D565_Opaque_SSE2,               // S32_D565_Opaque
michael@0:     NULL,                               // S32_D565_Blend
michael@0:     S32A_D565_Opaque_SSE2,              // S32A_D565_Opaque
michael@0:     NULL,                               // S32A_D565_Blend
michael@0:     S32_D565_Opaque_Dither_SSE2,        // S32_D565_Opaque_Dither
michael@0:     NULL,                               // S32_D565_Blend_Dither
michael@0:     S32A_D565_Opaque_Dither_SSE2,       // S32A_D565_Opaque_Dither
michael@0:     NULL,                               // S32A_D565_Blend_Dither
michael@0: };
michael@0: 
michael@0: static SkBlitRow::Proc32 platform_32_procs[] = {
michael@0:     NULL,                               // S32_Opaque,
michael@0:     S32_Blend_BlitRow32_SSE2,           // S32_Blend,
michael@0:     S32A_Opaque_BlitRow32_SSE2,         // S32A_Opaque
michael@0:     S32A_Blend_BlitRow32_SSE2,          // S32A_Blend,
michael@0: };
michael@0: 
michael@0: SkBlitRow::Proc SkBlitRow::PlatformProcs565(unsigned flags) {
michael@0:     if (cachedHasSSE2()) {
michael@0:         return platform_16_procs[flags];
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: }
michael@0: 
michael@0: SkBlitRow::ColorProc SkBlitRow::PlatformColorProc() {
michael@0:     if (cachedHasSSE2()) {
michael@0:         return Color32_SSE2;
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: }
michael@0: 
michael@0: SkBlitRow::Proc32 SkBlitRow::PlatformProcs32(unsigned flags) {
michael@0:     if (cachedHasSSE2()) {
michael@0:         return platform_32_procs[flags];
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: }
michael@0: 
michael@0: 
michael@0: SkBlitMask::ColorProc SkBlitMask::PlatformColorProcs(SkBitmap::Config dstConfig,
michael@0:                                                      SkMask::Format maskFormat,
michael@0:                                                      SkColor color) {
michael@0:     if (SkMask::kA8_Format != maskFormat) {
michael@0:         return NULL;
michael@0:     }
michael@0: 
michael@0:     ColorProc proc = NULL;
michael@0:     if (cachedHasSSE2()) {
michael@0:         switch (dstConfig) {
michael@0:             case SkBitmap::kARGB_8888_Config:
michael@0:                 // The SSE2 version is not (yet) faster for black, so we check
michael@0:                 // for that.
michael@0:                 if (SK_ColorBLACK != color) {
michael@0:                     proc = SkARGB32_A8_BlitMask_SSE2;
michael@0:                 }
michael@0:                 break;
michael@0:             default:
michael@0:                 break;
michael@0:         }
michael@0:     }
michael@0:     return proc;
michael@0: }
michael@0: 
michael@0: SkBlitMask::BlitLCD16RowProc SkBlitMask::PlatformBlitRowProcs16(bool isOpaque) {
michael@0:     if (cachedHasSSE2()) {
michael@0:         if (isOpaque) {
michael@0:             return SkBlitLCD16OpaqueRow_SSE2;
michael@0:         } else {
michael@0:             return SkBlitLCD16Row_SSE2;
michael@0:         }
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: 
michael@0: }
michael@0: SkBlitMask::RowProc SkBlitMask::PlatformRowProcs(SkBitmap::Config dstConfig,
michael@0:                                                  SkMask::Format maskFormat,
michael@0:                                                  RowFlags flags) {
michael@0:     return NULL;
michael@0: }
michael@0: 
michael@0: SkMemset16Proc SkMemset16GetPlatformProc() {
michael@0:     if (cachedHasSSE2()) {
michael@0:         return sk_memset16_SSE2;
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: }
michael@0: 
michael@0: SkMemset32Proc SkMemset32GetPlatformProc() {
michael@0:     if (cachedHasSSE2()) {
michael@0:         return sk_memset32_SSE2;
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: }
michael@0: 
michael@0: SkMorphologyImageFilter::Proc SkMorphologyGetPlatformProc(SkMorphologyProcType type) {
michael@0:     if (!cachedHasSSE2()) {
michael@0:         return NULL;
michael@0:     }
michael@0:     switch (type) {
michael@0:         case kDilateX_SkMorphologyProcType:
michael@0:             return SkDilateX_SSE2;
michael@0:         case kDilateY_SkMorphologyProcType:
michael@0:             return SkDilateY_SSE2;
michael@0:         case kErodeX_SkMorphologyProcType:
michael@0:             return SkErodeX_SSE2;
michael@0:         case kErodeY_SkMorphologyProcType:
michael@0:             return SkErodeY_SSE2;
michael@0:         default:
michael@0:             return NULL;
michael@0:     }
michael@0: }
michael@0: 
michael@0: bool SkBoxBlurGetPlatformProcs(SkBoxBlurProc* boxBlurX,
michael@0:                                SkBoxBlurProc* boxBlurY,
michael@0:                                SkBoxBlurProc* boxBlurXY,
michael@0:                                SkBoxBlurProc* boxBlurYX) {
michael@0: #ifdef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
michael@0:     return false;
michael@0: #else
michael@0:     if (!cachedHasSSE2()) {
michael@0:         return false;
michael@0:     }
michael@0:     return SkBoxBlurGetPlatformProcs_SSE2(boxBlurX, boxBlurY, boxBlurXY, boxBlurYX);
michael@0: #endif
michael@0: }
michael@0: 
michael@0: SkBlitRow::ColorRectProc PlatformColorRectProcFactory(); // suppress warning
michael@0: 
michael@0: SkBlitRow::ColorRectProc PlatformColorRectProcFactory() {
michael@0:     if (cachedHasSSE2()) {
michael@0:         return ColorRect32_SSE2;
michael@0:     } else {
michael@0:         return NULL;
michael@0:     }
michael@0: }