Use SSE some in framebuffers to memory.

Since people are using it (even though it seems dangerous), might as well
make it faster.  6% better in Tales of Phantasia X.
This commit is contained in:
Unknown W. Brackets committed 2014-05-04 01:03:48 -07:00
1 parent 4475c99426
commit e8d5866e5a
3 files changed
+184 -23

No files matched your search

+52 -23
View File
@@ -35,6 +35,7 @@
#include "GPU/GPUState.h"
#include "GPU/Common/PostShader.h"
#include "GPU/Common/TextureDecoder.h"
#include "GPU/GLES/Framebuffer.h"
#include "GPU/GLES/TextureCache.h"
#include "GPU/GLES/ShaderManager.h"
@@ -111,11 +112,15 @@ inline u16 RGBA8888toRGBA4444(u32 px) {
return ((px >> 4) & 0x000F) | ((px >> 8) & 0x00F0) | ((px >> 12) & 0x0F00) | ((px >> 16) & 0xF000);
}
inline u16 RGBA8888toRGBA5551(u32 px) {
return ((px >> 3) & 0x001F) | ((px >> 6) & 0x03E0) | ((px >> 9) & 0x7C00) | ((px >> 16) & 0x8000);
inline u16 BGRA8888toRGB565(u32 px) {
return ((px >> 19) & 0x001F) | ((px >> 5) & 0x07E0) | ((px << 8) & 0xF800);
}
void ConvertFromRGBA8888(u8 *dst, u8 *src, u32 stride, u32 height, GEBufferFormat format);
inline u16 BGRA8888toRGBA4444(u32 px) {
return ((px >> 20) & 0x000F) | ((px >> 8) & 0x00F0) | ((px << 4) & 0x0F00) | ((px >> 16) & 0xF000);
}
void ConvertFromRGBA8888(u8 *dst, const u8 *src, u32 stride, u32 height, GEBufferFormat format);
void CenterRect(float *x, float *y, float *w, float *h,
float origW, float origH, float frameW, float frameH) {
@@ -1256,12 +1261,23 @@ void FramebufferManager::BlitFramebuffer_(VirtualFramebuffer *src, VirtualFrameb
fbo_unbind();
}
static inline bool UseBGRA8888() {
// TODO: Other platforms? May depend on vendor which is faster?
#ifdef _WIN32
return gl_extensions.EXT_bgra;
#endif
return false;
}
// TODO: SSE/NEON
// Could also make C fake-simd for 64-bit, two 8888 pixels fit in a register :)
void ConvertFromRGBA8888(u8 *dst, u8 *src, u32 stride, u32 height, GEBufferFormat format) {
void ConvertFromRGBA8888(u8 *dst, const u8 *src, u32 stride, u32 height, GEBufferFormat format) {
if (format == GE_FORMAT_8888) {
if (src == dst) {
return;
} else if (UseBGRA8888()) {
u32 numPixels = height * stride;
ConvertBGRA8888ToRGBA8888((u32 *)dst, (const u32 *)src, numPixels);
} else { // Here lets assume they don't intersect
memcpy(dst, src, stride * height * 4);
}
@@ -1271,25 +1287,38 @@ void ConvertFromRGBA8888(u8 *dst, u8 *src, u32 stride, u32 height, GEBufferForma
u16 *dst16 = (u16 *)dst;
switch (format) {
case GE_FORMAT_565: // BGR 565
for (int i = 0; i < size; i++) {
dst16[i] = RGBA8888toRGB565(src32[i]);
if (UseBGRA8888()) {
for (int i = 0; i < size; i++) {
dst16[i] = BGRA8888toRGB565(src32[i]);
}
} else {
for (int i = 0; i < size; i++) {
dst16[i] = RGBA8888toRGB565(src32[i]);
}
}
break;
case GE_FORMAT_5551: // ABGR 1555
for (int i = 0; i < size; i++) {
dst16[i] = RGBA8888toRGBA5551(src32[i]);
if (UseBGRA8888()) {
ConvertBGRA8888ToRGBA5551(dst16, src32, size);
} else {
ConvertRGBA8888ToRGBA5551(dst16, src32, size);
}
break;
case GE_FORMAT_4444: // ABGR 4444
for (int i = 0; i < size; i++) {
dst16[i] = RGBA8888toRGBA4444(src32[i]);
if (UseBGRA8888()) {
for (int i = 0; i < size; i++) {
dst16[i] = BGRA8888toRGBA4444(src32[i]);
}
} else {
for (int i = 0; i < size; i++) {
dst16[i] = RGBA8888toRGBA4444(src32[i]);
}
}
break;
case GE_FORMAT_8888:
case GE_FORMAT_INVALID:
// Not possible.
break;
default:
break;
}
}
}
@@ -1317,24 +1346,24 @@ void FramebufferManager::PackFramebufferAsync_(VirtualFramebuffer *vfb) {
}
// Receive previously requested data from a PBO
if (pixelBufObj_[nextPBO].reading) {
glBindBuffer(GL_PIXEL_PACK_BUFFER, pixelBufObj_[nextPBO].handle);
AsyncPBO &pbo = pixelBufObj_[nextPBO];
if (pbo.reading) {
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo.handle);
packed = (GLubyte *)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
if (packed) {
DEBUG_LOG(SCEGE, "Reading PBO to memory , bufSize = %u, packed = %p, fb_address = %08x, stride = %u, pbo = %u",
pixelBufObj_[nextPBO].size, packed, pixelBufObj_[nextPBO].fb_address, pixelBufObj_[nextPBO].stride, nextPBO);
pbo.size, packed, pbo.fb_address, pbo.stride, nextPBO);
if (useCPU) {
ConvertFromRGBA8888(Memory::GetPointer(pixelBufObj_[nextPBO].fb_address), packed,
pixelBufObj_[nextPBO].stride, pixelBufObj_[nextPBO].height,
pixelBufObj_[nextPBO].format);
if (useCPU || (UseBGRA8888() && pbo.format == GE_FORMAT_8888)) {
u8 *dst = Memory::GetPointer(pbo.fb_address);
ConvertFromRGBA8888(dst, packed, pbo.stride, pbo.height, pbo.format);
} else {
// We don't need to convert, GPU already did (or should have)
Memory::Memcpy(pixelBufObj_[nextPBO].fb_address, packed, pixelBufObj_[nextPBO].size);
Memory::Memcpy(pbo.fb_address, packed, pbo.size);
}
pixelBufObj_[nextPBO].reading = false;
pbo.reading = false;
}
glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
@@ -1371,7 +1400,7 @@ void FramebufferManager::PackFramebufferAsync_(VirtualFramebuffer *vfb) {
case GE_FORMAT_8888: // 32 bit RGBA
default:
pixelType = GL_UNSIGNED_BYTE;
pixelFormat = GL_RGBA;
pixelFormat = UseBGRA8888() ? GL_BGRA_EXT : GL_RGBA;
pixelSize = 4;
align = 4;
break;
@@ -1412,7 +1441,7 @@ void FramebufferManager::PackFramebufferAsync_(VirtualFramebuffer *vfb) {
if (useCPU) {
// If converting pixel formats on the CPU we'll always request RGBA8888
glPixelStorei(GL_PACK_ALIGNMENT, 4);
glReadPixels(0, 0, vfb->fb_stride, vfb->height, GL_RGBA, GL_UNSIGNED_BYTE, 0);
glReadPixels(0, 0, vfb->fb_stride, vfb->height, UseBGRA8888() ? GL_BGRA_EXT : GL_RGBA, GL_UNSIGNED_BYTE, 0);
} else {
// Otherwise we'll directly request the format we need and let the GPU sort it out
glPixelStorei(GL_PACK_ALIGNMENT, align);