1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/opts/opts_check_SSE2.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,309 @@
1.4 +/*
1.5 + * Copyright 2009 The Android Open Source Project
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license that can be
1.8 + * found in the LICENSE file.
1.9 + */
1.10 +
1.11 +#include "SkBitmapProcState_opts_SSE2.h"
1.12 +#include "SkBitmapProcState_opts_SSSE3.h"
1.13 +#include "SkBitmapFilter_opts_SSE2.h"
1.14 +#include "SkBlitMask.h"
1.15 +#include "SkBlitRow.h"
1.16 +#include "SkBlitRect_opts_SSE2.h"
1.17 +#include "SkBlitRow_opts_SSE2.h"
1.18 +#include "SkBlurImage_opts_SSE2.h"
1.19 +#include "SkUtils_opts_SSE2.h"
1.20 +#include "SkUtils.h"
1.21 +#include "SkMorphology_opts.h"
1.22 +#include "SkMorphology_opts_SSE2.h"
1.23 +
1.24 +#include "SkRTConf.h"
1.25 +
1.26 +#if defined(_MSC_VER) && defined(_WIN64)
1.27 +#include <intrin.h>
1.28 +#endif
1.29 +
1.30 +/* This file must *not* be compiled with -msse or -msse2, otherwise
1.31 + gcc may generate sse2 even for scalar ops (and thus give an invalid
1.32 + instruction on Pentium3 on the code below). Only files named *_SSE2.cpp
1.33 + in this directory should be compiled with -msse2. */
1.34 +
1.35 +
1.36 +#ifdef _MSC_VER
1.37 +static inline void getcpuid(int info_type, int info[4]) {
1.38 +#if defined(_WIN64)
1.39 + __cpuid(info, info_type);
1.40 +#else
1.41 + __asm {
1.42 + mov eax, [info_type]
1.43 + cpuid
1.44 + mov edi, [info]
1.45 + mov [edi], eax
1.46 + mov [edi+4], ebx
1.47 + mov [edi+8], ecx
1.48 + mov [edi+12], edx
1.49 + }
1.50 +#endif
1.51 +}
1.52 +#else
1.53 +#if defined(__x86_64__)
1.54 +static inline void getcpuid(int info_type, int info[4]) {
1.55 + asm volatile (
1.56 + "cpuid \n\t"
1.57 + : "=a"(info[0]), "=b"(info[1]), "=c"(info[2]), "=d"(info[3])
1.58 + : "a"(info_type)
1.59 + );
1.60 +}
1.61 +#else
1.62 +static inline void getcpuid(int info_type, int info[4]) {
1.63 + // We save and restore ebx, so this code can be compatible with -fPIC
1.64 + asm volatile (
1.65 + "pushl %%ebx \n\t"
1.66 + "cpuid \n\t"
1.67 + "movl %%ebx, %1 \n\t"
1.68 + "popl %%ebx \n\t"
1.69 + : "=a"(info[0]), "=r"(info[1]), "=c"(info[2]), "=d"(info[3])
1.70 + : "a"(info_type)
1.71 + );
1.72 +}
1.73 +#endif
1.74 +#endif
1.75 +
1.76 +#if defined(__x86_64__) || defined(_WIN64) || SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
1.77 +/* All x86_64 machines have SSE2, or we know it's supported at compile time, so don't even bother checking. */
1.78 +static inline bool hasSSE2() {
1.79 + return true;
1.80 +}
1.81 +#else
1.82 +
1.83 +static inline bool hasSSE2() {
1.84 + int cpu_info[4] = { 0 };
1.85 + getcpuid(1, cpu_info);
1.86 + return (cpu_info[3] & (1<<26)) != 0;
1.87 +}
1.88 +#endif
1.89 +
1.90 +#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
1.91 +/* If we know SSSE3 is supported at compile time, don't even bother checking. */
1.92 +static inline bool hasSSSE3() {
1.93 + return true;
1.94 +}
1.95 +#else
1.96 +
1.97 +static inline bool hasSSSE3() {
1.98 + int cpu_info[4] = { 0 };
1.99 + getcpuid(1, cpu_info);
1.100 + return (cpu_info[2] & 0x200) != 0;
1.101 +}
1.102 +#endif
1.103 +
1.104 +static bool cachedHasSSE2() {
1.105 + static bool gHasSSE2 = hasSSE2();
1.106 + return gHasSSE2;
1.107 +}
1.108 +
1.109 +static bool cachedHasSSSE3() {
1.110 + static bool gHasSSSE3 = hasSSSE3();
1.111 + return gHasSSSE3;
1.112 +}
1.113 +
1.114 +SK_CONF_DECLARE( bool, c_hqfilter_sse, "bitmap.filter.highQualitySSE", false, "Use SSE optimized version of high quality image filters");
1.115 +
1.116 +void SkBitmapProcState::platformConvolutionProcs(SkConvolutionProcs* procs) {
1.117 + if (cachedHasSSE2()) {
1.118 + procs->fExtraHorizontalReads = 3;
1.119 + procs->fConvolveVertically = &convolveVertically_SSE2;
1.120 + procs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_SSE2;
1.121 + procs->fConvolveHorizontally = &convolveHorizontally_SSE2;
1.122 + procs->fApplySIMDPadding = &applySIMDPadding_SSE2;
1.123 + }
1.124 +}
1.125 +
1.126 +void SkBitmapProcState::platformProcs() {
1.127 + if (cachedHasSSSE3()) {
1.128 + if (fSampleProc32 == S32_opaque_D32_filter_DX) {
1.129 + fSampleProc32 = S32_opaque_D32_filter_DX_SSSE3;
1.130 + } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
1.131 + fSampleProc32 = S32_alpha_D32_filter_DX_SSSE3;
1.132 + }
1.133 +
1.134 + if (fSampleProc32 == S32_opaque_D32_filter_DXDY) {
1.135 + fSampleProc32 = S32_opaque_D32_filter_DXDY_SSSE3;
1.136 + } else if (fSampleProc32 == S32_alpha_D32_filter_DXDY) {
1.137 + fSampleProc32 = S32_alpha_D32_filter_DXDY_SSSE3;
1.138 + }
1.139 + } else if (cachedHasSSE2()) {
1.140 + if (fSampleProc32 == S32_opaque_D32_filter_DX) {
1.141 + fSampleProc32 = S32_opaque_D32_filter_DX_SSE2;
1.142 + } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
1.143 + fSampleProc32 = S32_alpha_D32_filter_DX_SSE2;
1.144 + }
1.145 +
1.146 + if (fSampleProc16 == S32_D16_filter_DX) {
1.147 + fSampleProc16 = S32_D16_filter_DX_SSE2;
1.148 + }
1.149 + }
1.150 +
1.151 + if (cachedHasSSSE3() || cachedHasSSE2()) {
1.152 + if (fMatrixProc == ClampX_ClampY_filter_scale) {
1.153 + fMatrixProc = ClampX_ClampY_filter_scale_SSE2;
1.154 + } else if (fMatrixProc == ClampX_ClampY_nofilter_scale) {
1.155 + fMatrixProc = ClampX_ClampY_nofilter_scale_SSE2;
1.156 + }
1.157 +
1.158 + if (fMatrixProc == ClampX_ClampY_filter_affine) {
1.159 + fMatrixProc = ClampX_ClampY_filter_affine_SSE2;
1.160 + } else if (fMatrixProc == ClampX_ClampY_nofilter_affine) {
1.161 + fMatrixProc = ClampX_ClampY_nofilter_affine_SSE2;
1.162 + }
1.163 + if (c_hqfilter_sse) {
1.164 + if (fShaderProc32 == highQualityFilter32) {
1.165 + fShaderProc32 = highQualityFilter_SSE2;
1.166 + }
1.167 + }
1.168 + }
1.169 +}
1.170 +
1.171 +static SkBlitRow::Proc platform_16_procs[] = {
1.172 + S32_D565_Opaque_SSE2, // S32_D565_Opaque
1.173 + NULL, // S32_D565_Blend
1.174 + S32A_D565_Opaque_SSE2, // S32A_D565_Opaque
1.175 + NULL, // S32A_D565_Blend
1.176 + S32_D565_Opaque_Dither_SSE2, // S32_D565_Opaque_Dither
1.177 + NULL, // S32_D565_Blend_Dither
1.178 + S32A_D565_Opaque_Dither_SSE2, // S32A_D565_Opaque_Dither
1.179 + NULL, // S32A_D565_Blend_Dither
1.180 +};
1.181 +
1.182 +static SkBlitRow::Proc32 platform_32_procs[] = {
1.183 + NULL, // S32_Opaque,
1.184 + S32_Blend_BlitRow32_SSE2, // S32_Blend,
1.185 + S32A_Opaque_BlitRow32_SSE2, // S32A_Opaque
1.186 + S32A_Blend_BlitRow32_SSE2, // S32A_Blend,
1.187 +};
1.188 +
1.189 +SkBlitRow::Proc SkBlitRow::PlatformProcs565(unsigned flags) {
1.190 + if (cachedHasSSE2()) {
1.191 + return platform_16_procs[flags];
1.192 + } else {
1.193 + return NULL;
1.194 + }
1.195 +}
1.196 +
1.197 +SkBlitRow::ColorProc SkBlitRow::PlatformColorProc() {
1.198 + if (cachedHasSSE2()) {
1.199 + return Color32_SSE2;
1.200 + } else {
1.201 + return NULL;
1.202 + }
1.203 +}
1.204 +
1.205 +SkBlitRow::Proc32 SkBlitRow::PlatformProcs32(unsigned flags) {
1.206 + if (cachedHasSSE2()) {
1.207 + return platform_32_procs[flags];
1.208 + } else {
1.209 + return NULL;
1.210 + }
1.211 +}
1.212 +
1.213 +
1.214 +SkBlitMask::ColorProc SkBlitMask::PlatformColorProcs(SkBitmap::Config dstConfig,
1.215 + SkMask::Format maskFormat,
1.216 + SkColor color) {
1.217 + if (SkMask::kA8_Format != maskFormat) {
1.218 + return NULL;
1.219 + }
1.220 +
1.221 + ColorProc proc = NULL;
1.222 + if (cachedHasSSE2()) {
1.223 + switch (dstConfig) {
1.224 + case SkBitmap::kARGB_8888_Config:
1.225 + // The SSE2 version is not (yet) faster for black, so we check
1.226 + // for that.
1.227 + if (SK_ColorBLACK != color) {
1.228 + proc = SkARGB32_A8_BlitMask_SSE2;
1.229 + }
1.230 + break;
1.231 + default:
1.232 + break;
1.233 + }
1.234 + }
1.235 + return proc;
1.236 +}
1.237 +
1.238 +SkBlitMask::BlitLCD16RowProc SkBlitMask::PlatformBlitRowProcs16(bool isOpaque) {
1.239 + if (cachedHasSSE2()) {
1.240 + if (isOpaque) {
1.241 + return SkBlitLCD16OpaqueRow_SSE2;
1.242 + } else {
1.243 + return SkBlitLCD16Row_SSE2;
1.244 + }
1.245 + } else {
1.246 + return NULL;
1.247 + }
1.248 +
1.249 +}
1.250 +SkBlitMask::RowProc SkBlitMask::PlatformRowProcs(SkBitmap::Config dstConfig,
1.251 + SkMask::Format maskFormat,
1.252 + RowFlags flags) {
1.253 + return NULL;
1.254 +}
1.255 +
1.256 +SkMemset16Proc SkMemset16GetPlatformProc() {
1.257 + if (cachedHasSSE2()) {
1.258 + return sk_memset16_SSE2;
1.259 + } else {
1.260 + return NULL;
1.261 + }
1.262 +}
1.263 +
1.264 +SkMemset32Proc SkMemset32GetPlatformProc() {
1.265 + if (cachedHasSSE2()) {
1.266 + return sk_memset32_SSE2;
1.267 + } else {
1.268 + return NULL;
1.269 + }
1.270 +}
1.271 +
1.272 +SkMorphologyImageFilter::Proc SkMorphologyGetPlatformProc(SkMorphologyProcType type) {
1.273 + if (!cachedHasSSE2()) {
1.274 + return NULL;
1.275 + }
1.276 + switch (type) {
1.277 + case kDilateX_SkMorphologyProcType:
1.278 + return SkDilateX_SSE2;
1.279 + case kDilateY_SkMorphologyProcType:
1.280 + return SkDilateY_SSE2;
1.281 + case kErodeX_SkMorphologyProcType:
1.282 + return SkErodeX_SSE2;
1.283 + case kErodeY_SkMorphologyProcType:
1.284 + return SkErodeY_SSE2;
1.285 + default:
1.286 + return NULL;
1.287 + }
1.288 +}
1.289 +
1.290 +bool SkBoxBlurGetPlatformProcs(SkBoxBlurProc* boxBlurX,
1.291 + SkBoxBlurProc* boxBlurY,
1.292 + SkBoxBlurProc* boxBlurXY,
1.293 + SkBoxBlurProc* boxBlurYX) {
1.294 +#ifdef SK_DISABLE_BLUR_DIVISION_OPTIMIZATION
1.295 + return false;
1.296 +#else
1.297 + if (!cachedHasSSE2()) {
1.298 + return false;
1.299 + }
1.300 + return SkBoxBlurGetPlatformProcs_SSE2(boxBlurX, boxBlurY, boxBlurXY, boxBlurYX);
1.301 +#endif
1.302 +}
1.303 +
1.304 +SkBlitRow::ColorRectProc PlatformColorRectProcFactory(); // suppress warning
1.305 +
1.306 +SkBlitRow::ColorRectProc PlatformColorRectProcFactory() {
1.307 + if (cachedHasSSE2()) {
1.308 + return ColorRect32_SSE2;
1.309 + } else {
1.310 + return NULL;
1.311 + }
1.312 +}
