SkPixelInfo::CopyPixels() refactor part 0: only one memcpy() case
I envision this eventually looking like the following:
Fast Path 1: memcpy
Fast Path 2: Legacy RGBA/BGRA unpremul/premul
Fast Path 3: SkColorSpaceXform
Fallback: Raster pipeline
BUG=skia:6021
Change-Id: I3f0e36b08b051ab4ba3949f04474f20d0c20a963
Reviewed-on: https://skia-review.googlesource.com/7435
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Matt Sarett <msarett@google.com>
diff --git a/src/core/SkConfig8888.cpp b/src/core/SkConfig8888.cpp
index cdecf05..10bed27 100644
--- a/src/core/SkConfig8888.cpp
+++ b/src/core/SkConfig8888.cpp
@@ -18,6 +18,24 @@
#include "SkRasterPipeline.h"
#include "SkUnPreMultiply.h"
+// Fast Path 1: The memcpy() case.
+static inline bool can_memcpy(const SkImageInfo& dstInfo, const SkImageInfo& srcInfo) {
+ if (dstInfo.colorType() != srcInfo.colorType()) {
+ return false;
+ }
+
+ if (dstInfo.alphaType() != srcInfo.alphaType() &&
+ kOpaque_SkAlphaType != dstInfo.alphaType() &&
+ kOpaque_SkAlphaType != srcInfo.alphaType())
+ {
+ // We need to premultiply or unpremultiply.
+ return false;
+ }
+
+ return !dstInfo.colorSpace() || !srcInfo.colorSpace() ||
+ SkColorSpace::Equals(dstInfo.colorSpace(), srcInfo.colorSpace());
+}
+
// For now disable 565 in the pipeline. Its (higher) quality is so different its too much to
// rebase (for now)
//
@@ -169,10 +187,6 @@
}
}
-static void memcpy32_row(uint32_t* dst, const uint32_t* src, int count) {
- memcpy(dst, src, count * 4);
-}
-
bool SkSrcPixelInfo::convertPixelsTo(SkDstPixelInfo* dst, int width, int height) const {
SkASSERT(width > 0 && height > 0);
@@ -186,14 +200,8 @@
switch (doAlpha) {
case kNothing_AlphaVerb:
- if (doSwapRB) {
- proc = convert32_row<true, kNothing_AlphaVerb>;
- } else {
- if (fPixels == dst->fPixels) {
- return true;
- }
- proc = memcpy32_row;
- }
+ SkASSERT(doSwapRB);
+ proc = convert32_row<true, kNothing_AlphaVerb>;
break;
case kPremul_AlphaVerb:
if (doSwapRB) {
@@ -370,14 +378,9 @@
const int width = srcInfo.width();
const int height = srcInfo.height();
- // Do the easiest one first : both configs are equal
- if (srcInfo == dstInfo) {
- size_t bytes = width * srcInfo.bytesPerPixel();
- for (int y = 0; y < height; ++y) {
- memcpy(dstPixels, srcPixels, bytes);
- srcPixels = (const char*)srcPixels + srcRB;
- dstPixels = (char*)dstPixels + dstRB;
- }
+ // Fast Path 1: The memcpy() case.
+ if (can_memcpy(dstInfo, srcInfo)) {
+ SkRectMemcpy(dstPixels, dstRB, srcPixels, srcRB, dstInfo.minRowBytes(), dstInfo.height());
return true;
}
@@ -405,30 +408,6 @@
return true;
}
- // If they agree on colorType and the alphaTypes are compatible, then we just memcpy.
- // Note: we've already taken care of 32bit colortypes above.
- if (srcInfo.colorType() == dstInfo.colorType()) {
- switch (srcInfo.colorType()) {
- case kRGBA_F16_SkColorType:
- if (!SkColorSpace::Equals(srcInfo.colorSpace(), dstInfo.colorSpace())) {
- break;
- }
- case kIndex_8_SkColorType:
- case kARGB_4444_SkColorType:
- if (srcInfo.alphaType() != dstInfo.alphaType()) {
- break;
- }
- case kRGB_565_SkColorType:
- case kAlpha_8_SkColorType:
- case kGray_8_SkColorType:
- SkRectMemcpy(dstPixels, dstRB, srcPixels, srcRB,
- width * srcInfo.bytesPerPixel(), height);
- return true;
- default:
- break;
- }
- }
-
/*
* Begin section where we try to change colorTypes along the way. Not all combinations
* are supported.
diff --git a/src/core/SkConfig8888.h b/src/core/SkConfig8888.h
index ff28726..3544cbe 100644
--- a/src/core/SkConfig8888.h
+++ b/src/core/SkConfig8888.h
@@ -9,6 +9,7 @@
#define SkPixelInfo_DEFINED
#include "SkImageInfo.h"
+#include "SkTemplates.h"
class SkColorTable;
@@ -36,12 +37,17 @@
static inline void SkRectMemcpy(void* dst, size_t dstRB, const void* src, size_t srcRB,
size_t bytesPerRow, int rowCount) {
- SkASSERT(bytesPerRow <= srcRB);
SkASSERT(bytesPerRow <= dstRB);
+ SkASSERT(bytesPerRow <= srcRB);
+ if (bytesPerRow == dstRB && bytesPerRow == srcRB) {
+ memcpy(dst, src, bytesPerRow * rowCount);
+ return;
+ }
+
for (int i = 0; i < rowCount; ++i) {
memcpy(dst, src, bytesPerRow);
- dst = (char*)dst + dstRB;
- src = (const char*)src + srcRB;
+ dst = SkTAddOffset<void>(dst, dstRB);
+ src = SkTAddOffset<const void>(src, srcRB);
}
}
diff --git a/src/gpu/vk/GrVkGpu.cpp b/src/gpu/vk/GrVkGpu.cpp
index 9bb4c57..b6b8a5d 100644
--- a/src/gpu/vk/GrVkGpu.cpp
+++ b/src/gpu/vk/GrVkGpu.cpp
@@ -477,13 +477,8 @@
dstRow -= layout.rowPitch;
}
} else {
- // If there is no padding on the src (rowBytes) or dst (layout.rowPitch) we can memcpy
- if (trimRowBytes == rowBytes && trimRowBytes == layout.rowPitch) {
- memcpy(mapPtr, data, trimRowBytes * height);
- } else {
- SkRectMemcpy(mapPtr, static_cast<size_t>(layout.rowPitch), data, rowBytes, trimRowBytes,
- height);
- }
+ SkRectMemcpy(mapPtr, static_cast<size_t>(layout.rowPitch), data, rowBytes, trimRowBytes,
+ height);
}
GrVkMemory::FlushMappedAlloc(this, alloc);
@@ -580,8 +575,6 @@
src -= rowBytes;
dst += trimRowBytes;
}
- } else if (trimRowBytes == rowBytes) {
- memcpy(dst, src, trimRowBytes * currentHeight);
} else {
SkRectMemcpy(dst, trimRowBytes, src, rowBytes, trimRowBytes, currentHeight);
}
@@ -966,12 +959,7 @@
// If there is no padding on dst we can do a single memcopy.
// This assumes the srcData comes in with no padding.
- if (srcRowBytes == dstRowBytes) {
- memcpy(mapPtr, srcData, srcRowBytes * h);
- } else {
- SkRectMemcpy(mapPtr, static_cast<size_t>(dstRowBytes), srcData, srcRowBytes,
- srcRowBytes, h);
- }
+ SkRectMemcpy(mapPtr, static_cast<size_t>(dstRowBytes), srcData, srcRowBytes, srcRowBytes, h);
GrVkMemory::FlushMappedAlloc(gpu, alloc);
GR_VK_CALL(gpu->vkInterface(), UnmapMemory(gpu->device(), alloc.fMemory));
return true;
@@ -1756,12 +1744,7 @@
dstRow -= rowBytes;
}
} else {
- if (transBufferRowBytes == rowBytes) {
- memcpy(buffer, mappedMemory, rowBytes*height);
- } else {
- SkRectMemcpy(buffer, rowBytes, mappedMemory, transBufferRowBytes, tightRowBytes,
- height);
- }
+ SkRectMemcpy(buffer, rowBytes, mappedMemory, transBufferRowBytes, tightRowBytes, height);
}
transferBuffer->unmap();