Revert "In SkImageBlurFilter, use ScaledDividerU32"
This reverts commit 8021f21b132ce5b90ed1d885311ec7f936466ede.
Reason for revert: Missing change
Original change's description:
> In SkImageBlurFilter, use ScaledDividerU32
>
> Bug: skia:12522
>
> Change-Id: I6608e66b44710f3633635d3621bdae2a5523f28b
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458456
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Herb Derby <herb@google.com>
Bug: skia:12522
Change-Id: I5c677ce435c9964be648acbda17d694c12894b3b
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/458719
Auto-Submit: Herb Derby <herb@google.com>
Commit-Queue: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
diff --git a/src/effects/imagefilters/SkBlurImageFilter.cpp b/src/effects/imagefilters/SkBlurImageFilter.cpp
index 12c38e5..059a076 100644
--- a/src/effects/imagefilters/SkBlurImageFilter.cpp
+++ b/src/effects/imagefilters/SkBlurImageFilter.cpp
@@ -306,7 +306,22 @@
int window2 = window * window;
int window3 = window2 * window;
int divisor = (window & 1) == 1 ? window3 : window3 + window2;
- return alloc->make<GaussPass>(buffer0, buffer1, buffer2, buffersEnd, border, divisor);
+
+ // NB the sums in the blur code use the following technique to avoid
+ // adding 1/2 to round the divide.
+ //
+ // Sum/d + 1/2 == (Sum + h) / d
+ // Sum + d(1/2) == Sum + h
+ // h == (1/2)d
+ //
+ // But the d/2 it self should be rounded.
+ // h == d/2 + 1/2 == (d + 1) / 2
+ //
+ // divisorFactor = (1 / d) * 2 ^ 32
+ auto divisorFactor = static_cast<uint32_t>(round((1.0 / divisor) * (1ull << 32)));
+ auto half = static_cast<uint32_t>((divisor + 1) / 2);
+ return alloc->make<GaussPass>(
+ buffer0, buffer1, buffer2, buffersEnd, border, divisorFactor, half);
}
GaussPass(skvx::Vec<4, uint32_t>* buffer0,
@@ -314,21 +329,22 @@
skvx::Vec<4, uint32_t>* buffer2,
skvx::Vec<4, uint32_t>* buffersEnd,
int border,
- int divisor)
+ uint32_t divisorFactor,
+ uint32_t half)
: Pass{border}
, fBuffer0{buffer0}
, fBuffer1{buffer1}
, fBuffer2{buffer2}
, fBuffersEnd{buffersEnd}
- , fDivider(divisor) {}
+ , fDivisorFactor{divisorFactor}
+ , fHalf{half} {}
private:
void startBlur() override {
skvx::Vec<4, uint32_t> zero = {0u, 0u, 0u, 0u};
zero.store(fSum0);
zero.store(fSum1);
- auto half = fDivider.half();
- skvx::Vec<4, uint32_t>{half, half, half, half}.store(fSum2);
+ skvx::Vec<4, uint32_t>{fHalf, fHalf, fHalf, fHalf}.store(fSum2);
sk_bzero(fBuffer0, (fBuffersEnd - fBuffer0) * sizeof(skvx::Vec<4, uint32_t>));
fBuffer0Cursor = fBuffer0;
@@ -389,7 +405,8 @@
sum1 += sum0;
sum2 += sum1;
- skvx::Vec<4, uint32_t> blurred = fDivider.divide(sum2);
+ skvx::Vec<4, uint64_t> w = skvx::cast<uint64_t>(sum2) * fDivisorFactor;
+ skvx::Vec<4, uint32_t> value = skvx::cast<uint32_t>(w >> 32);
sum2 -= *buffer2Cursor;
*buffer2Cursor = sum1;
@@ -401,7 +418,7 @@
*buffer0Cursor = leadingEdge;
buffer0Cursor = (buffer0Cursor + 1) < fBuffer1 ? buffer0Cursor + 1 : fBuffer0;
- return skvx::cast<uint8_t>(blurred);
+ return skvx::cast<uint8_t>(value);
};
auto loadEdge = [&](const uint32_t* srcCursor) {
@@ -444,7 +461,8 @@
skvx::Vec<4, uint32_t>* const fBuffer1;
skvx::Vec<4, uint32_t>* const fBuffer2;
skvx::Vec<4, uint32_t>* const fBuffersEnd;
- const skvx::ScaledDividerU32 fDivider;
+ const uint32_t fDivisorFactor;
+ const uint32_t fHalf;
// blur state
char fSum0[sizeof(skvx::Vec<4, uint32_t>)];
@@ -559,25 +577,40 @@
int border = window - 1;
int divisor = window * window;
- return alloc->make<TentPass>(buffer0, buffer1, buffersEnd, border, divisor);
+
+ // NB the sums in the blur code use the following technique to avoid
+ // adding 1/2 to round the divide.
+ //
+ // Sum/d + 1/2 == (Sum + h) / d
+ // Sum + d(1/2) == Sum + h
+ // h == (1/2)d
+ //
+ // But the d/2 it self should be rounded.
+ // h == d/2 + 1/2 == (d + 1) / 2
+ //
+ // divisorFactor = (1 / d) * 2 ^ 32
+ auto divisorFactor = static_cast<uint32_t>(round((1.0 / divisor) * (1ull << 32)));
+ auto half = static_cast<uint32_t>((divisor + 1) / 2);
+ return alloc->make<TentPass>(buffer0, buffer1, buffersEnd, border, divisorFactor, half);
}
TentPass(skvx::Vec<4, uint32_t>* buffer0,
skvx::Vec<4, uint32_t>* buffer1,
skvx::Vec<4, uint32_t>* buffersEnd,
int border,
- int divisor)
+ uint32_t divisorFactor,
+ uint32_t half)
: Pass{border}
, fBuffer0{buffer0}
, fBuffer1{buffer1}
, fBuffersEnd{buffersEnd}
- , fDivider(divisor) {}
+ , fDivisorFactor{divisorFactor}
+ , fHalf{half} {}
private:
void startBlur() override {
skvx::Vec<4, uint32_t>{0u, 0u, 0u, 0u}.store(fSum0);
- auto half = fDivider.half();
- skvx::Vec<4, uint32_t>{half, half, half, half}.store(fSum1);
+ skvx::Vec<4, uint32_t>{fHalf, fHalf, fHalf, fHalf}.store(fSum1);
sk_bzero(fBuffer0, (fBuffersEnd - fBuffer0) * sizeof(skvx::Vec<4, uint32_t>));
fBuffer0Cursor = fBuffer0;
@@ -630,7 +663,8 @@
sum0 += leadingEdge;
sum1 += sum0;
- skvx::Vec<4, uint32_t> blurred = fDivider.divide(sum1);
+ skvx::Vec<4, uint64_t> w = skvx::cast<uint64_t>(sum1) * fDivisorFactor;
+ skvx::Vec<4, uint32_t> value = skvx::cast<uint32_t>(w >> 32);
sum1 -= *buffer1Cursor;
*buffer1Cursor = sum0;
@@ -639,7 +673,7 @@
*buffer0Cursor = leadingEdge;
buffer0Cursor = (buffer0Cursor + 1) < fBuffer1 ? buffer0Cursor + 1 : fBuffer0;
- return skvx::cast<uint8_t>(blurred);
+ return skvx::cast<uint8_t>(value);
};
auto loadEdge = [&](const uint32_t* srcCursor) {
@@ -678,7 +712,8 @@
skvx::Vec<4, uint32_t>* const fBuffer0;
skvx::Vec<4, uint32_t>* const fBuffer1;
skvx::Vec<4, uint32_t>* const fBuffersEnd;
- const skvx::ScaledDividerU32 fDivider;
+ const uint32_t fDivisorFactor;
+ const uint32_t fHalf;
// blur state
char fSum0[sizeof(skvx::Vec<4, uint32_t>)];
