thirdparty/ppsspp
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

GPU: Allow gcc/clang/icc runtime SSE4 usage.

Browse source 
All our builds before were only using SSE4 in jit...
This commit is contained in:
Unknown W. Brackets 2022-01-08 17:00:00 -08:00
parent eee62849fe
commit 8a00c2d233
6 changed files with 104 additions and 58 deletions
Download patch file Download diff file Expand all files Collapse all files

61
Common/Data/Convert/ColorConv.cpp
View file

@ -25,9 +25,6 @@
#ifdef _M_SSE #ifdef _M_SSE
#include <emmintrin.h> #include <emmintrin.h>
#endif
#if _M_SSE >= 0x401
#include <smmintrin.h> #include <smmintrin.h>
#endif #endif
@ -181,19 +178,15 @@ void ConvertBGRA8888ToRGB888(u8 *dst, const u32 *src, u32 numPixels) {
} }
} }
void ConvertRGBA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) { #if defined(_M_SSE)
#if _M_SSE >= 0x401 #if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static inline void ConvertRGBA8888ToRGBA5551_SSE4(__m128i *dstp, const __m128i *srcp, u32 sseChunks) {
const __m128i maskAG = _mm_set1_epi32(0x8000F800); const __m128i maskAG = _mm_set1_epi32(0x8000F800);
const __m128i maskRB = _mm_set1_epi32(0x00F800F8); const __m128i maskRB = _mm_set1_epi32(0x00F800F8);
const __m128i mask = _mm_set1_epi32(0x0000FFFF); const __m128i mask = _mm_set1_epi32(0x0000FFFF);
const __m128i *srcp = (const __m128i *)src;
__m128i *dstp = (__m128i *)dst;
u32 sseChunks = (numPixels / 4) & ~1;
// SSE 4.1 required for _mm_packus_epi32.
if (((intptr_t)src & 0xF) || ((intptr_t)dst & 0xF) || !cpu_info.bSSE4_1) {
sseChunks = 0;
}
for (u32 i = 0; i < sseChunks; i += 2) { for (u32 i = 0; i < sseChunks; i += 2) {
__m128i c1 = _mm_load_si128(&srcp[i + 0]); __m128i c1 = _mm_load_si128(&srcp[i + 0]);
__m128i c2 = _mm_load_si128(&srcp[i + 1]); __m128i c2 = _mm_load_si128(&srcp[i + 1]);
@ -213,6 +206,21 @@ void ConvertRGBA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) {
_mm_store_si128(&dstp[i / 2], _mm_packus_epi32(c1, c2)); _mm_store_si128(&dstp[i / 2], _mm_packus_epi32(c1, c2));
} }
}
#endif
void ConvertRGBA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) {
#if defined(_M_SSE)
const __m128i *srcp = (const __m128i *)src;
__m128i *dstp = (__m128i *)dst;
u32 sseChunks = (numPixels / 4) & ~1;
// SSE 4.1 required for _mm_packus_epi32.
if (((intptr_t)src & 0xF) || ((intptr_t)dst & 0xF) || !cpu_info.bSSE4_1) {
sseChunks = 0;
} else {
ConvertRGBA8888ToRGBA5551_SSE4(dstp, srcp, sseChunks);
}
// The remainder starts right after those done via SSE. // The remainder starts right after those done via SSE.
u32 i = sseChunks * 4; u32 i = sseChunks * 4;
#else #else
@ -223,19 +231,15 @@ void ConvertRGBA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) {
} }
} }
void ConvertBGRA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) { #if defined(_M_SSE)
#if _M_SSE >= 0x401 #if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static inline void ConvertBGRA8888ToRGBA5551_SSE4(__m128i *dstp, const __m128i *srcp, u32 sseChunks) {
const __m128i maskAG = _mm_set1_epi32(0x8000F800); const __m128i maskAG = _mm_set1_epi32(0x8000F800);
const __m128i maskRB = _mm_set1_epi32(0x00F800F8); const __m128i maskRB = _mm_set1_epi32(0x00F800F8);
const __m128i mask = _mm_set1_epi32(0x0000FFFF); const __m128i mask = _mm_set1_epi32(0x0000FFFF);
const __m128i *srcp = (const __m128i *)src;
__m128i *dstp = (__m128i *)dst;
u32 sseChunks = (numPixels / 4) & ~1;
// SSE 4.1 required for _mm_packus_epi32.
if (((intptr_t)src & 0xF) || ((intptr_t)dst & 0xF) || !cpu_info.bSSE4_1) {
sseChunks = 0;
}
for (u32 i = 0; i < sseChunks; i += 2) { for (u32 i = 0; i < sseChunks; i += 2) {
__m128i c1 = _mm_load_si128(&srcp[i + 0]); __m128i c1 = _mm_load_si128(&srcp[i + 0]);
__m128i c2 = _mm_load_si128(&srcp[i + 1]); __m128i c2 = _mm_load_si128(&srcp[i + 1]);
@ -255,6 +259,21 @@ void ConvertBGRA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) {
_mm_store_si128(&dstp[i / 2], _mm_packus_epi32(c1, c2)); _mm_store_si128(&dstp[i / 2], _mm_packus_epi32(c1, c2));
} }
}
#endif
void ConvertBGRA8888ToRGBA5551(u16 *dst, const u32 *src, u32 numPixels) {
#if defined(_M_SSE)
const __m128i *srcp = (const __m128i *)src;
__m128i *dstp = (__m128i *)dst;
u32 sseChunks = (numPixels / 4) & ~1;
// SSE 4.1 required for _mm_packus_epi32.
if (((intptr_t)src & 0xF) || ((intptr_t)dst & 0xF) || !cpu_info.bSSE4_1) {
sseChunks = 0;
} else {
ConvertBGRA8888ToRGBA5551_SSE4(dstp, srcp, sseChunks);
}
// The remainder starts right after those done via SSE. // The remainder starts right after those done via SSE.
u32 i = sseChunks * 4; u32 i = sseChunks * 4;
#else #else

2
GPU/Common/TextureDecoder.cpp
View file

@ -29,9 +29,7 @@
#ifdef _M_SSE #ifdef _M_SSE
#include <emmintrin.h> #include <emmintrin.h>
#if _M_SSE >= 0x401
#include <smmintrin.h> #include <smmintrin.h>
#endif
u32 QuickTexHashSSE2(const void *checkp, u32 size) { u32 QuickTexHashSSE2(const void *checkp, u32 size) {
u32 check = 0; u32 check = 0;

18
GPU/Common/TextureScalerCommon.cpp
View file

@ -31,7 +31,8 @@
#include "Common/CPUDetect.h" #include "Common/CPUDetect.h"
#include "ext/xbrz/xbrz.h" #include "ext/xbrz/xbrz.h"
#if _M_SSE >= 0x401 #if defined(_M_SSE)
#include <emmintrin.h>
#include <smmintrin.h> #include <smmintrin.h>
#endif #endif
@ -281,9 +282,12 @@ void scaleBicubicT(u32* data, u32* out, int w, int h, int l, int u) {
} }
} }
} }
#if _M_SSE >= 0x401 #if defined(_M_SSE)
template<int f, int T> template<int f, int T>
void scaleBicubicTSSE41(u32* data, u32* out, int w, int h, int l, int u) { #if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static void scaleBicubicTSSE41(u32* data, u32* out, int w, int h, int l, int u) {
int outw = w*f; int outw = w*f;
for (int yb = 0; yb < (u - l)*f / BLOCK_SIZE + 1; ++yb) { for (int yb = 0; yb < (u - l)*f / BLOCK_SIZE + 1; ++yb) {
for (int xb = 0; xb < w*f / BLOCK_SIZE + 1; ++xb) { for (int xb = 0; xb < w*f / BLOCK_SIZE + 1; ++xb) {
@ -321,7 +325,7 @@ void scaleBicubicTSSE41(u32* data, u32* out, int w, int h, int l, int u) {
#endif #endif
void scaleBicubicBSpline(int factor, u32* data, u32* out, int w, int h, int l, int u) { void scaleBicubicBSpline(int factor, u32* data, u32* out, int w, int h, int l, int u) {
#if _M_SSE >= 0x401 #if defined(_M_SSE)
if (cpu_info.bSSE4_1) { if (cpu_info.bSSE4_1) {
switch (factor) { switch (factor) {
case 2: scaleBicubicTSSE41<2, 0>(data, out, w, h, l, u); break; // when I first tested this, case 2: scaleBicubicTSSE41<2, 0>(data, out, w, h, l, u); break; // when I first tested this,
@ -339,13 +343,13 @@ void scaleBicubicBSpline(int factor, u32* data, u32* out, int w, int h, int l, i
case 5: scaleBicubicT<5, 0>(data, out, w, h, l, u); break; // any of these break statements case 5: scaleBicubicT<5, 0>(data, out, w, h, l, u); break; // any of these break statements
default: ERROR_LOG(G3D, "Bicubic upsampling only implemented for factors 2 to 5"); default: ERROR_LOG(G3D, "Bicubic upsampling only implemented for factors 2 to 5");
} }
#if _M_SSE >= 0x401 #if defined(_M_SSE)
} }
#endif #endif
} }
void scaleBicubicMitchell(int factor, u32* data, u32* out, int w, int h, int l, int u) { void scaleBicubicMitchell(int factor, u32* data, u32* out, int w, int h, int l, int u) {
#if _M_SSE >= 0x401 #if defined(_M_SSE)
if (cpu_info.bSSE4_1) { if (cpu_info.bSSE4_1) {
switch (factor) { switch (factor) {
case 2: scaleBicubicTSSE41<2, 1>(data, out, w, h, l, u); break; case 2: scaleBicubicTSSE41<2, 1>(data, out, w, h, l, u); break;
@ -363,7 +367,7 @@ void scaleBicubicMitchell(int factor, u32* data, u32* out, int w, int h, int l,
case 5: scaleBicubicT<5, 1>(data, out, w, h, l, u); break; case 5: scaleBicubicT<5, 1>(data, out, w, h, l, u); break;
default: ERROR_LOG(G3D, "Bicubic upsampling only implemented for factors 2 to 5"); default: ERROR_LOG(G3D, "Bicubic upsampling only implemented for factors 2 to 5");
} }
#if _M_SSE >= 0x401 #if defined(_M_SSE)
} }
#endif #endif
} }

11
GPU/Math3D.cpp
View file

@ -114,7 +114,9 @@ __m128 SSENormalizeMultiplierSSE2(__m128 v)
return _mm_shuffle_ps(rt, rt, _MM_SHUFFLE(0, 0, 0, 0)); return _mm_shuffle_ps(rt, rt, _MM_SHUFFLE(0, 0, 0, 0));
} }
#if _M_SSE >= 0x401 #if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
__m128 SSENormalizeMultiplierSSE4(__m128 v) __m128 SSENormalizeMultiplierSSE4(__m128 v)
{ {
return _mm_rsqrt_ps(_mm_dp_ps(v, v, 0xFF)); return _mm_rsqrt_ps(_mm_dp_ps(v, v, 0xFF));
@ -126,12 +128,7 @@ __m128 SSENormalizeMultiplier(bool useSSE4, __m128 v)
return SSENormalizeMultiplierSSE4(v); return SSENormalizeMultiplierSSE4(v);
return SSENormalizeMultiplierSSE2(v); return SSENormalizeMultiplierSSE2(v);
} }
#else
__m128 SSENormalizeMultiplier(bool useSSE4, __m128 v)
{
return SSENormalizeMultiplierSSE2(v);
}
#endif
template<> template<>
Vec3<float> Vec3<float>::Normalized(bool useSSE4) const Vec3<float> Vec3<float>::Normalized(bool useSSE4) const
{ {

2
GPU/Math3D.h
View file

@ -25,10 +25,8 @@
#if defined(_M_SSE) #if defined(_M_SSE)
#include <emmintrin.h> #include <emmintrin.h>
#if _M_SSE >= 0x401
#include <smmintrin.h> #include <smmintrin.h>
#endif #endif
#endif
#if PPSSPP_ARCH(ARM_NEON) #if PPSSPP_ARCH(ARM_NEON)
#if defined(_MSC_VER) && PPSSPP_ARCH(ARM64) #if defined(_MSC_VER) && PPSSPP_ARCH(ARM64)

68
GPU/Software/Rasterizer.cpp
View file

@ -40,9 +40,6 @@
#if defined(_M_SSE) #if defined(_M_SSE)
#include <emmintrin.h> #include <emmintrin.h>
#endif
#if _M_SSE >= 0x401
#include <smmintrin.h> #include <smmintrin.h>
#endif #endif
@ -583,6 +580,17 @@ struct TriangleEdge {
Vec4<int> stepY; Vec4<int> stepY;
}; };
#if defined(_M_SSE) && !PPSSPP_ARCH(X86)
#if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static inline __m128i SOFTRAST_CALL TriangleEdgeStartSSE4(__m128i initX, __m128i initY, int xf, int yf, int c) {
initX = _mm_mullo_epi32(initX, _mm_set1_epi32(xf));
initY = _mm_mullo_epi32(initY, _mm_set1_epi32(yf));
return _mm_add_epi32(_mm_add_epi32(initX, initY), _mm_set1_epi32(c));
}
#endif
template <bool useSSE4> template <bool useSSE4>
Vec4<int> TriangleEdge<useSSE4>::Start(const ScreenCoords &v0, const ScreenCoords &v1, const ScreenCoords &origin) { Vec4<int> TriangleEdge<useSSE4>::Start(const ScreenCoords &v0, const ScreenCoords &v1, const ScreenCoords &origin) {
// Start at pixel centers. // Start at pixel centers.
@ -597,12 +605,9 @@ Vec4<int> TriangleEdge<useSSE4>::Start(const ScreenCoords &v0, const ScreenCoord
stepX = Vec4<int>::AssignToAll(xf * 16 * 2); stepX = Vec4<int>::AssignToAll(xf * 16 * 2);
stepY = Vec4<int>::AssignToAll(yf * 16 * 2); stepY = Vec4<int>::AssignToAll(yf * 16 * 2);
#if defined(_M_SSE) && !PPSSPP_ARCH(X86) && _M_SSE >= 0x401 #if defined(_M_SSE) && !PPSSPP_ARCH(X86)
if (useSSE4) { if (useSSE4)
initX.ivec = _mm_mullo_epi32(initX.ivec, _mm_set1_epi32(xf)); return TriangleEdgeStartSSE4(initX.ivec, initY.ivec, xf, yf, c);
initY.ivec = _mm_mullo_epi32(initY.ivec, _mm_set1_epi32(yf));
return _mm_add_epi32(_mm_add_epi32(initX.ivec, initY.ivec), _mm_set1_epi32(c));
}
#endif #endif
return Vec4<int>::AssignToAll(xf) * initX + Vec4<int>::AssignToAll(yf) * initY + Vec4<int>::AssignToAll(c); return Vec4<int>::AssignToAll(xf) * initX + Vec4<int>::AssignToAll(yf) * initY + Vec4<int>::AssignToAll(c);
} }
@ -625,14 +630,23 @@ inline Vec4<int> TriangleEdge<useSSE4>::StepY(const Vec4<int> &w) {
#endif #endif
} }
#if defined(_M_SSE) && !PPSSPP_ARCH(X86)
#if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static inline int SOFTRAST_CALL MaxWeightSSE4(__m128i w) {
__m128i max2 = _mm_max_epi32(w, _mm_shuffle_epi32(w, _MM_SHUFFLE(3, 2, 3, 2)));
__m128i max1 = _mm_max_epi32(max2, _mm_shuffle_epi32(max2, _MM_SHUFFLE(1, 1, 1, 1)));
return _mm_cvtsi128_si32(max1);
}
#endif
template <bool useSSE4> template <bool useSSE4>
void TriangleEdge<useSSE4>::NarrowMinMaxX(const Vec4<int> &w, int64_t minX, int64_t &rowMinX, int64_t &rowMaxX) { void TriangleEdge<useSSE4>::NarrowMinMaxX(const Vec4<int> &w, int64_t minX, int64_t &rowMinX, int64_t &rowMaxX) {
int wmax; int wmax;
#if defined(_M_SSE) && !PPSSPP_ARCH(X86) && _M_SSE >= 0x401 #if defined(_M_SSE) && !PPSSPP_ARCH(X86)
if (useSSE4) { if (useSSE4) {
__m128i max01 = _mm_max_epi32(w.ivec, _mm_shuffle_epi32(w.ivec, _MM_SHUFFLE(3, 2, 3, 2))); wmax = MaxWeightSSE4(w.ivec);
__m128i max0 = _mm_max_epi32(max01, _mm_shuffle_epi32(max01, _MM_SHUFFLE(1, 1, 1, 1)));
wmax = _mm_cvtsi128_si32(max0);
} else { } else {
wmax = std::max(std::max(w.x, w.y), std::max(w.z, w.w)); wmax = std::max(std::max(w.x, w.y), std::max(w.z, w.w));
} }
@ -654,11 +668,20 @@ void TriangleEdge<useSSE4>::NarrowMinMaxX(const Vec4<int> &w, int64_t minX, int6
} }
} }
#if defined(_M_SSE) && !PPSSPP_ARCH(X86)
#if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static inline __m128i SOFTRAST_CALL StepTimesSSE4(__m128i w, __m128i step, int c) {
return _mm_add_epi32(w, _mm_mullo_epi32(_mm_set1_epi32(c), step));
}
#endif
template <bool useSSE4> template <bool useSSE4>
inline Vec4<int> TriangleEdge<useSSE4>::StepXTimes(const Vec4<int> &w, int c) { inline Vec4<int> TriangleEdge<useSSE4>::StepXTimes(const Vec4<int> &w, int c) {
#if defined(_M_SSE) && !PPSSPP_ARCH(X86) && _M_SSE >= 0x401 #if defined(_M_SSE) && !PPSSPP_ARCH(X86)
if (useSSE4) if (useSSE4)
return _mm_add_epi32(w.ivec, _mm_mullo_epi32(_mm_set1_epi32(c), stepX.ivec)); return StepTimesSSE4(w.ivec, stepX.ivec, c);
#endif #endif
return w + stepX * c; return w + stepX * c;
} }
@ -675,15 +698,22 @@ static inline Vec4<int> MakeMask(const Vec4<int> &w0, const Vec4<int> &w1, const
#endif #endif
} }
#if defined(_M_SSE) && !PPSSPP_ARCH(X86)
#if defined(__GNUC__) || defined(__clang__) || defined(__INTEL_COMPILER)
[[gnu::target("sse4.1")]]
#endif
static inline bool SOFTRAST_CALL AnyMaskSSE4(__m128i mask) {
__m128i sig = _mm_srai_epi32(mask, 31);
return _mm_test_all_ones(sig) == 0;
}
#endif
template <bool useSSE4> template <bool useSSE4>
static inline bool AnyMask(const Vec4<int> &mask) { static inline bool AnyMask(const Vec4<int> &mask) {
#if defined(_M_SSE) && !PPSSPP_ARCH(X86) #if defined(_M_SSE) && !PPSSPP_ARCH(X86)
#if _M_SSE >= 0x401
if (useSSE4) { if (useSSE4) {
__m128i sig = _mm_srai_epi32(mask.ivec, 31); return AnyMaskSSE4(mask.ivec);
return _mm_test_all_ones(sig) == 0;
} }
#endif
// In other words: !(mask.x < 0 && mask.y < 0 && mask.z < 0 && mask.w < 0) // In other words: !(mask.x < 0 && mask.y < 0 && mask.z < 0 && mask.w < 0)
__m128i low2 = _mm_and_si128(mask.ivec, _mm_shuffle_epi32(mask.ivec, _MM_SHUFFLE(3, 2, 3, 2))); __m128i low2 = _mm_and_si128(mask.ivec, _mm_shuffle_epi32(mask.ivec, _MM_SHUFFLE(3, 2, 3, 2)));