Skip to content

A
angle
Commit 5695fc99 by Geoff Lang Committed by Commit Bot
Options

Clean up the SSE detection logic.

* Include the correct header right in the SSE check in platform.h. * Don't use separate SSE versions of the load functions, have them use SSE automatically. BUG=612205 Change-Id: I70f9a5513e144db4d16c1f3ad922debeb6c50268 Reviewed-on: https://chromium-review.googlesource.com/358108Reviewed-by: 's avatarCorentin Wallez <cwallez@chromium.org> Reviewed-by: 's avatarJamie Madill <jmadill@chromium.org> Commit-Queue: Geoff Lang <geofflang@chromium.org>
parent e92507bc
Showing with 123 additions and 200 deletions
src/common/mathutil.h
...@@ -135,7 +135,7 @@ inline unsigned int unorm(float x) ...@@ -135,7 +135,7 @@ inline unsigned int unorm(float x)
inline bool supportsSSE2() inline bool supportsSSE2()
{ {
#if defined(ANGLE_PLATFORM_WINDOWS) && !defined(_M_ARM) #if defined(ANGLE_USE_SSE)
static bool checked = false; static bool checked = false;
static bool supports = false; static bool supports = false;
...@@ -144,21 +144,22 @@ inline bool supportsSSE2() ...@@ -144,21 +144,22 @@ inline bool supportsSSE2()
return supports; return supports;
} }
int info[4]; #if defined(ANGLE_PLATFORM_WINDOWS) && !defined(_M_ARM)
__cpuid(info, 0);
if (info[0] >= 1)
{ {
__cpuid(info, 1); int info[4];
__cpuid(info, 0);
supports = (info[3] >> 26) & 1; if (info[0] >= 1)
} {
__cpuid(info, 1);
supports = (info[3] >> 26) & 1;
}
}
#endif // defined(ANGLE_PLATFORM_WINDOWS) && !defined(_M_ARM)
checked = true; checked = true;
return supports; return supports;
#else #else // defined(ANGLE_USE_SSE)
UNIMPLEMENTED();
return false; return false;
#endif #endif
} }
......
src/common/platform.h
...@@ -77,8 +77,12 @@ ...@@ -77,8 +77,12 @@
# undef far # undef far
#endif #endif
#if !defined(_M_ARM) && !defined(ANGLE_PLATFORM_ANDROID) #if defined(_MSC_VER) && !defined(_M_ARM)
# define ANGLE_USE_SSE #include <intrin.h>
#define ANGLE_USE_SSE
#elif defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
#include <x86intrin.h>
#define ANGLE_USE_SSE
#endif #endif
#endif // COMMON_PLATFORM_H_ #endif // COMMON_PLATFORM_H_
src/image_util/loadimage.cpp
...@@ -9,6 +9,7 @@ ...@@ -9,6 +9,7 @@
#include "image_util/loadimage.h" #include "image_util/loadimage.h"
#include "common/mathutil.h" #include "common/mathutil.h"
#include "common/platform.h"
#include "image_util/imageformats.h" #include "image_util/imageformats.h"
namespace angle namespace angle
...@@ -24,6 +25,54 @@ void LoadA8ToRGBA8(size_t width, ...@@ -24,6 +25,54 @@ void LoadA8ToRGBA8(size_t width,
size_t outputRowPitch, size_t outputRowPitch,
size_t outputDepthPitch) size_t outputDepthPitch)
{ {
#if defined(ANGLE_USE_SSE)
if (gl::supportsSSE2())
{
__m128i zeroWide = _mm_setzero_si128();
for (size_t z = 0; z < depth; z++)
{
for (size_t y = 0; y < height; y++)
{
const uint8_t *source =
priv::OffsetDataPointer<uint8_t>(input, y, z, inputRowPitch, inputDepthPitch);
uint32_t *dest = priv::OffsetDataPointer<uint32_t>(output, y, z, outputRowPitch,
outputDepthPitch);
size_t x = 0;
// Make output writes aligned
for (; ((reinterpret_cast<intptr_t>(&dest[x]) & 0xF) != 0 && x < width); x++)
{
dest[x] = static_cast<uint32_t>(source[x]) << 24;
}
for (; x + 7 < width; x += 8)
{
__m128i sourceData =
_mm_loadl_epi64(reinterpret_cast<const __m128i *>(&source[x]));
// Interleave each byte to 16bit, make the lower byte to zero
sourceData = _mm_unpacklo_epi8(zeroWide, sourceData);
// Interleave each 16bit to 32bit, make the lower 16bit to zero
__m128i lo = _mm_unpacklo_epi16(zeroWide, sourceData);
__m128i hi = _mm_unpackhi_epi16(zeroWide, sourceData);
_mm_store_si128(reinterpret_cast<__m128i *>(&dest[x]), lo);
_mm_store_si128(reinterpret_cast<__m128i *>(&dest[x + 4]), hi);
}
// Handle the remainder
for (; x < width; x++)
{
dest[x] = static_cast<uint32_t>(source[x]) << 24;
}
}
}
return;
}
#endif
for (size_t z = 0; z < depth; z++) for (size_t z = 0; z < depth; z++)
{ {
for (size_t y = 0; y < height; y++) for (size_t y = 0; y < height; y++)
...@@ -545,6 +594,58 @@ void LoadRGBA8ToBGRA8(size_t width, ...@@ -545,6 +594,58 @@ void LoadRGBA8ToBGRA8(size_t width,
size_t outputRowPitch, size_t outputRowPitch,
size_t outputDepthPitch) size_t outputDepthPitch)
{ {
#if defined(ANGLE_USE_SSE)
if (gl::supportsSSE2())
{
__m128i brMask = _mm_set1_epi32(0x00ff00ff);
for (size_t z = 0; z < depth; z++)
{
for (size_t y = 0; y < height; y++)
{
const uint32_t *source =
priv::OffsetDataPointer<uint32_t>(input, y, z, inputRowPitch, inputDepthPitch);
uint32_t *dest = priv::OffsetDataPointer<uint32_t>(output, y, z, outputRowPitch,
outputDepthPitch);
size_t x = 0;
// Make output writes aligned
for (; ((reinterpret_cast<intptr_t>(&dest[x]) & 15) != 0) && x < width; x++)
{
uint32_t rgba = source[x];
dest[x] = (ANGLE_ROTL(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
}
for (; x + 3 < width; x += 4)
{
__m128i sourceData =
_mm_loadu_si128(reinterpret_cast<const __m128i *>(&source[x]));
// Mask out g and a, which don't change
__m128i gaComponents = _mm_andnot_si128(brMask, sourceData);
// Mask out b and r
__m128i brComponents = _mm_and_si128(sourceData, brMask);
// Swap b and r
__m128i brSwapped = _mm_shufflehi_epi16(
_mm_shufflelo_epi16(brComponents, _MM_SHUFFLE(2, 3, 0, 1)),
_MM_SHUFFLE(2, 3, 0, 1));
__m128i result = _mm_or_si128(gaComponents, brSwapped);
_mm_store_si128(reinterpret_cast<__m128i *>(&dest[x]), result);
}
// Perform leftover writes
for (; x < width; x++)
{
uint32_t rgba = source[x];
dest[x] = (ANGLE_ROTL(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
}
}
}
return;
}
#endif
for (size_t z = 0; z < depth; z++) for (size_t z = 0; z < depth; z++)
{ {
for (size_t y = 0; y < height; y++) for (size_t y = 0; y < height; y++)
......
src/image_util/loadimage.h
...@@ -35,16 +35,6 @@ void LoadA8ToBGRA8(size_t width, ...@@ -35,16 +35,6 @@ void LoadA8ToBGRA8(size_t width,
size_t outputRowPitch, size_t outputRowPitch,
size_t outputDepthPitch); size_t outputDepthPitch);
void LoadA8ToBGRA8_SSE2(size_t width,
size_t height,
size_t depth,
const uint8_t *input,
size_t inputRowPitch,
size_t inputDepthPitch,
uint8_t *output,
size_t outputRowPitch,
size_t outputDepthPitch);
void LoadA32FToRGBA32F(size_t width, void LoadA32FToRGBA32F(size_t width,
size_t height, size_t height,
size_t depth, size_t depth,
...@@ -215,16 +205,6 @@ void LoadR5G6B5ToRGBA8(size_t width, ...@@ -215,16 +205,6 @@ void LoadR5G6B5ToRGBA8(size_t width,
size_t outputRowPitch, size_t outputRowPitch,
size_t outputDepthPitch); size_t outputDepthPitch);
void LoadRGBA8ToBGRA8_SSE2(size_t width,
size_t height,
size_t depth,
const uint8_t *input,
size_t inputRowPitch,
size_t inputDepthPitch,
uint8_t *output,
size_t outputRowPitch,
size_t outputDepthPitch);
void LoadRGBA8ToBGRA8(size_t width, void LoadRGBA8ToBGRA8(size_t width,
size_t height, size_t height,
size_t depth, size_t depth,
......
src/image_util/loadimage_sse2.cpp deleted 100644 → 0
//
// Copyright (c) 2002-2015 The ANGLE Project Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
//
// loadimageSSE2.cpp: Defines image loading functions. It's
// in a separated file for GCC, which can enable SSE usage only per-file,
// not for code blocks that use SSE2 explicitly.
#include "image_util/loadimage.h"
#include "common/mathutil.h"
#include "common/platform.h"
#ifdef ANGLE_USE_SSE
#include <emmintrin.h>
#endif
namespace angle
{
void LoadA8ToBGRA8_SSE2(size_t width,
size_t height,
size_t depth,
const uint8_t *input,
size_t inputRowPitch,
size_t inputDepthPitch,
uint8_t *output,
size_t outputRowPitch,
size_t outputDepthPitch)
{
#if defined(ANGLE_USE_SSE)
__m128i zeroWide = _mm_setzero_si128();
for (size_t z = 0; z < depth; z++)
{
for (size_t y = 0; y < height; y++)
{
const uint8_t *source =
priv::OffsetDataPointer<uint8_t>(input, y, z, inputRowPitch, inputDepthPitch);
uint32_t *dest =
priv::OffsetDataPointer<uint32_t>(output, y, z, outputRowPitch, outputDepthPitch);
size_t x = 0;
// Make output writes aligned
for (; ((reinterpret_cast<intptr_t>(&dest[x]) & 0xF) != 0 && x < width); x++)
{
dest[x] = static_cast<uint32_t>(source[x]) << 24;
}
for (; x + 7 < width; x += 8)
{
__m128i sourceData = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(&source[x]));
// Interleave each byte to 16bit, make the lower byte to zero
sourceData = _mm_unpacklo_epi8(zeroWide, sourceData);
// Interleave each 16bit to 32bit, make the lower 16bit to zero
__m128i lo = _mm_unpacklo_epi16(zeroWide, sourceData);
__m128i hi = _mm_unpackhi_epi16(zeroWide, sourceData);
_mm_store_si128(reinterpret_cast<__m128i *>(&dest[x]), lo);
_mm_store_si128(reinterpret_cast<__m128i *>(&dest[x + 4]), hi);
}
// Handle the remainder
for (; x < width; x++)
{
dest[x] = static_cast<uint32_t>(source[x]) << 24;
}
}
}
#else
// Ensure that this function is reported as not implemented for ARM builds because
// the instructions below are not present for that architecture.
UNIMPLEMENTED();
return;
#endif
}
void LoadRGBA8ToBGRA8_SSE2(size_t width,
size_t height,
size_t depth,
const uint8_t *input,
size_t inputRowPitch,
size_t inputDepthPitch,
uint8_t *output,
size_t outputRowPitch,
size_t outputDepthPitch)
{
#if defined(ANGLE_USE_SSE)
__m128i brMask = _mm_set1_epi32(0x00ff00ff);
for (size_t z = 0; z < depth; z++)
{
for (size_t y = 0; y < height; y++)
{
const uint32_t *source =
priv::OffsetDataPointer<uint32_t>(input, y, z, inputRowPitch, inputDepthPitch);
uint32_t *dest =
priv::OffsetDataPointer<uint32_t>(output, y, z, outputRowPitch, outputDepthPitch);
size_t x = 0;
// Make output writes aligned
for (; ((reinterpret_cast<intptr_t>(&dest[x]) & 15) != 0) && x < width; x++)
{
uint32_t rgba = source[x];
dest[x] = (ANGLE_ROTL(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
}
for (; x + 3 < width; x += 4)
{
__m128i sourceData = _mm_loadu_si128(reinterpret_cast<const __m128i *>(&source[x]));
// Mask out g and a, which don't change
__m128i gaComponents = _mm_andnot_si128(brMask, sourceData);
// Mask out b and r
__m128i brComponents = _mm_and_si128(sourceData, brMask);
// Swap b and r
__m128i brSwapped =
_mm_shufflehi_epi16(_mm_shufflelo_epi16(brComponents, _MM_SHUFFLE(2, 3, 0, 1)),
_MM_SHUFFLE(2, 3, 0, 1));
__m128i result = _mm_or_si128(gaComponents, brSwapped);
_mm_store_si128(reinterpret_cast<__m128i *>(&dest[x]), result);
}
// Perform leftover writes
for (; x < width; x++)
{
uint32_t rgba = source[x];
dest[x] = (ANGLE_ROTL(rgba, 16) & 0x00ff00ff) | (rgba & 0xff00ff00);
}
}
}
#else
// Ensure that this function is reported as not implemented for ARM builds because
// the instructions below are not present for that architecture.
UNIMPLEMENTED();
return;
#endif
}
}
src/libANGLE/renderer/d3d/d3d9/formatutils9.cpp
...@@ -179,28 +179,6 @@ static InternalFormatInitialzerMap BuildInternalFormatInitialzerMap() ...@@ -179,28 +179,6 @@ static InternalFormatInitialzerMap BuildInternalFormatInitialzerMap()
return map; return map;
} }
// Each GL internal format corresponds to one D3D format and data loading function.
// Due to not all formats being available all the time, some of the function/format types are wrapped
// in templates that perform format support queries on a Renderer9 object which is supplied
// when requesting the function or format.
typedef bool(*FallbackPredicateFunction)();
template <FallbackPredicateFunction pred, LoadImageFunction prefered, LoadImageFunction fallback>
static void FallbackLoad(size_t width, size_t height, size_t depth,
const uint8_t *input, size_t inputRowPitch, size_t inputDepthPitch,
uint8_t *output, size_t outputRowPitch, size_t outputDepthPitch)
{
if (pred())
{
prefered(width, height, depth, input, inputRowPitch, inputDepthPitch, output, outputRowPitch, outputDepthPitch);
}
else
{
fallback(width, height, depth, input, inputRowPitch, inputDepthPitch, output, outputRowPitch, outputDepthPitch);
}
}
static void UnreachableLoad(size_t width, size_t height, size_t depth, static void UnreachableLoad(size_t width, size_t height, size_t depth,
const uint8_t *input, size_t inputRowPitch, size_t inputDepthPitch, const uint8_t *input, size_t inputRowPitch, size_t inputDepthPitch,
uint8_t *output, size_t outputRowPitch, size_t outputDepthPitch) uint8_t *output, size_t outputRowPitch, size_t outputDepthPitch)
...@@ -241,6 +219,7 @@ static D3D9FormatMap BuildD3D9FormatMap() ...@@ -241,6 +219,7 @@ static D3D9FormatMap BuildD3D9FormatMap()
D3D9FormatMap map; D3D9FormatMap map;
// clang-format off
// | Internal format | Texture format | Render format | Load function | // | Internal format | Texture format | Render format | Load function |
InsertD3D9FormatInfo(&map, GL_NONE, D3DFMT_NULL, D3DFMT_NULL, UnreachableLoad ); InsertD3D9FormatInfo(&map, GL_NONE, D3DFMT_NULL, D3DFMT_NULL, UnreachableLoad );
...@@ -274,11 +253,11 @@ static D3D9FormatMap BuildD3D9FormatMap() ...@@ -274,11 +253,11 @@ static D3D9FormatMap BuildD3D9FormatMap()
InsertD3D9FormatInfo(&map, GL_LUMINANCE16F_EXT, D3DFMT_A16B16G16R16F, D3DFMT_UNKNOWN, LoadL16FToRGBA16F ); InsertD3D9FormatInfo(&map, GL_LUMINANCE16F_EXT, D3DFMT_A16B16G16R16F, D3DFMT_UNKNOWN, LoadL16FToRGBA16F );
InsertD3D9FormatInfo(&map, GL_LUMINANCE_ALPHA16F_EXT, D3DFMT_A16B16G16R16F, D3DFMT_UNKNOWN, LoadLA16FToRGBA16F ); InsertD3D9FormatInfo(&map, GL_LUMINANCE_ALPHA16F_EXT, D3DFMT_A16B16G16R16F, D3DFMT_UNKNOWN, LoadLA16FToRGBA16F );
InsertD3D9FormatInfo(&map, GL_ALPHA8_EXT, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, FallbackLoad<gl::supportsSSE2, LoadA8ToBGRA8_SSE2, LoadA8ToBGRA8>); InsertD3D9FormatInfo(&map, GL_ALPHA8_EXT, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, LoadA8ToBGRA8 );
InsertD3D9FormatInfo(&map, GL_RGB8_OES, D3DFMT_X8R8G8B8, D3DFMT_X8R8G8B8, LoadRGB8ToBGRX8 ); InsertD3D9FormatInfo(&map, GL_RGB8_OES, D3DFMT_X8R8G8B8, D3DFMT_X8R8G8B8, LoadRGB8ToBGRX8 );
InsertD3D9FormatInfo(&map, GL_RGB565, D3DFMT_X8R8G8B8, D3DFMT_X8R8G8B8, LoadR5G6B5ToBGRA8 ); InsertD3D9FormatInfo(&map, GL_RGB565, D3DFMT_X8R8G8B8, D3DFMT_X8R8G8B8, LoadR5G6B5ToBGRA8 );
InsertD3D9FormatInfo(&map, GL_RGBA8_OES, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, FallbackLoad<gl::supportsSSE2, LoadRGBA8ToBGRA8_SSE2, LoadRGBA8ToBGRA8>); InsertD3D9FormatInfo(&map, GL_RGBA8_OES, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, LoadRGBA8ToBGRA8 );
InsertD3D9FormatInfo(&map, GL_RGBA4, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, LoadRGBA4ToBGRA8 ); InsertD3D9FormatInfo(&map, GL_RGBA4, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, LoadRGBA4ToBGRA8 );
InsertD3D9FormatInfo(&map, GL_RGB5_A1, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, LoadRGB5A1ToBGRA8 ); InsertD3D9FormatInfo(&map, GL_RGB5_A1, D3DFMT_A8R8G8B8, D3DFMT_A8R8G8B8, LoadRGB5A1ToBGRA8 );
InsertD3D9FormatInfo(&map, GL_R8_EXT, D3DFMT_X8R8G8B8, D3DFMT_X8R8G8B8, LoadR8ToBGRX8 ); InsertD3D9FormatInfo(&map, GL_R8_EXT, D3DFMT_X8R8G8B8, D3DFMT_X8R8G8B8, LoadR8ToBGRX8 );
...@@ -297,6 +276,7 @@ static D3D9FormatMap BuildD3D9FormatMap() ...@@ -297,6 +276,7 @@ static D3D9FormatMap BuildD3D9FormatMap()
// then changing the format and loading function appropriately. // then changing the format and loading function appropriately.
InsertD3D9FormatInfo(&map, GL_LUMINANCE8_EXT, D3DFMT_L8, D3DFMT_UNKNOWN, LoadToNative<GLubyte, 1> ); InsertD3D9FormatInfo(&map, GL_LUMINANCE8_EXT, D3DFMT_L8, D3DFMT_UNKNOWN, LoadToNative<GLubyte, 1> );
InsertD3D9FormatInfo(&map, GL_LUMINANCE8_ALPHA8_EXT, D3DFMT_A8L8, D3DFMT_UNKNOWN, LoadToNative<GLubyte, 2> ); InsertD3D9FormatInfo(&map, GL_LUMINANCE8_ALPHA8_EXT, D3DFMT_A8L8, D3DFMT_UNKNOWN, LoadToNative<GLubyte, 2> );
// clang-format on
return map; return map;
} }
......
src/libGLESv2.gypi
...@@ -51,7 +51,6 @@ ...@@ -51,7 +51,6 @@
'image_util/loadimage.h', 'image_util/loadimage.h',
'image_util/loadimage.inl', 'image_util/loadimage.inl',
'image_util/loadimage_etc.cpp', 'image_util/loadimage_etc.cpp',
'image_util/loadimage_sse2.cpp',
], ],
'libangle_includes': 'libangle_includes':
[ [
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment