Mike's radial gradient CL with better float -> int.
patch from issue 1072303005 at patchset 40001 (http://crrev.com/1072303005#ps40001)
This looks quite launchable. radial_gradient3, min of 100 samples:
N5: 985µs -> 946µs
MBP: 395µs -> 279µs
On my MBP, most of the meat looks like it's now in reading the cache and writing to dst one color at a time. Is that something we could do in float math rather than with a lookup table?
BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia.compile:Build-Mac10.8-Clang-Arm7-Debug-Android-Trybot,Build-Ubuntu-GCC-Arm7-Release-Android_NoNeon-Trybot
Committed: https://skia.googlesource.com/skia/+/abf6c5cf95e921fae59efb487480e5b5081cf0ec
Review URL: https://codereview.chromium.org/1109643002
diff --git a/src/effects/gradients/SkRadialGradient.cpp b/src/effects/gradients/SkRadialGradient.cpp
index b25a875..bf3c821 100644
--- a/src/effects/gradients/SkRadialGradient.cpp
+++ b/src/effects/gradients/SkRadialGradient.cpp
@@ -8,6 +8,7 @@
#include "SkRadialGradient.h"
#include "SkRadialGradient_Table.h"
+#include "SkNx.h"
#define kSQRT_TABLE_BITS 11
#define kSQRT_TABLE_SIZE (1 << kSQRT_TABLE_BITS)
@@ -270,13 +271,16 @@
namespace {
inline bool radial_completely_pinned(int fx, int dx, int fy, int dy) {
- // fast, overly-conservative test: checks unit square instead
- // of unit circle
- bool xClamped = (fx >= SK_FixedHalf && dx >= 0) ||
- (fx <= -SK_FixedHalf && dx <= 0);
- bool yClamped = (fy >= SK_FixedHalf && dy >= 0) ||
- (fy <= -SK_FixedHalf && dy <= 0);
+ // fast, overly-conservative test: checks unit square instead of unit circle
+ bool xClamped = (fx >= SK_FixedHalf && dx >= 0) || (fx <= -SK_FixedHalf && dx <= 0);
+ bool yClamped = (fy >= SK_FixedHalf && dy >= 0) || (fy <= -SK_FixedHalf && dy <= 0);
+ return xClamped || yClamped;
+}
+inline bool radial_completely_pinned(SkScalar fx, SkScalar dx, SkScalar fy, SkScalar dy) {
+ // fast, overly-conservative test: checks unit square instead of unit circle
+ bool xClamped = (fx >= 1 && dx >= 0) || (fx <= -1 && dx <= 0);
+ bool yClamped = (fy >= 1 && dy >= 0) || (fy <= -1 && dy <= 0);
return xClamped || yClamped;
}
@@ -373,6 +377,70 @@
}
}
+// TODO: can we get away with 0th approximatino of inverse-sqrt (i.e. faster than rsqrt)?
+// seems like ~10bits is more than enough for our use, since we want a byte-index
+static inline Sk4f fast_sqrt(const Sk4f& R) {
+ return R * R.rsqrt();
+}
+
+static inline Sk4f sum_squares(const Sk4f& a, const Sk4f& b) {
+ return a * a + b * b;
+}
+
+void shadeSpan_radial_clamp2(SkScalar sfx, SkScalar sdx, SkScalar sfy, SkScalar sdy,
+ SkPMColor* SK_RESTRICT dstC, const SkPMColor* SK_RESTRICT cache,
+ int count, int toggle) {
+ if (radial_completely_pinned(sfx, sdx, sfy, sdy)) {
+ unsigned fi = SkGradientShaderBase::kCache32Count - 1;
+ sk_memset32_dither(dstC,
+ cache[toggle + fi],
+ cache[next_dither_toggle(toggle) + fi],
+ count);
+ } else {
+ const Sk4f max(255);
+ const float scale = 255;
+ sfx *= scale;
+ sfy *= scale;
+ sdx *= scale;
+ sdy *= scale;
+ const Sk4f fx4(sfx, sfx + sdx, sfx + 2*sdx, sfx + 3*sdx);
+ const Sk4f fy4(sfy, sfy + sdy, sfy + 2*sdy, sfy + 3*sdy);
+ const Sk4f dx4(sdx * 4);
+ const Sk4f dy4(sdy * 4);
+
+ Sk4f tmpxy = fx4 * dx4 + fy4 * dy4;
+ Sk4f tmpdxdy = sum_squares(dx4, dy4);
+ Sk4f R = sum_squares(fx4, fy4);
+ Sk4f dR = tmpxy + tmpxy + tmpdxdy;
+ const Sk4f ddR = tmpdxdy + tmpdxdy;
+
+ for (int i = 0; i < (count >> 2); ++i) {
+ Sk4f dist = Sk4f::Min(fast_sqrt(R), max);
+ R += dR;
+ dR += ddR;
+
+ int fi[4];
+ dist.castTrunc().store(fi);
+
+ for (int i = 0; i < 4; i++) {
+ *dstC++ = cache[toggle + fi[i]];
+ toggle = next_dither_toggle(toggle);
+ }
+ }
+ count &= 3;
+ if (count) {
+ Sk4f dist = Sk4f::Min(fast_sqrt(R), max);
+
+ int fi[4];
+ dist.castTrunc().store(fi);
+ for (int i = 0; i < count; i++) {
+ *dstC++ = cache[toggle + fi[i]];
+ toggle = next_dither_toggle(toggle);
+ }
+ }
+ }
+}
+
// Unrolling this loop doesn't seem to help (when float); we're stalling to
// get the results of the sqrt (?), and don't have enough extra registers to
// have many in flight.
@@ -407,6 +475,11 @@
void SkRadialGradient::RadialGradientContext::shadeSpan(int x, int y,
SkPMColor* SK_RESTRICT dstC, int count) {
+#ifdef SK_SUPPORT_LEGACY_RADIAL_GRADIENT_SQRT
+ const bool use_new_proc = false;
+#else
+ const bool use_new_proc = true;
+#endif
SkASSERT(count > 0);
const SkRadialGradient& radialGradient = static_cast<const SkRadialGradient&>(fShader);
@@ -435,7 +508,7 @@
RadialShadeProc shadeProc = shadeSpan_radial_repeat;
if (SkShader::kClamp_TileMode == radialGradient.fTileMode) {
- shadeProc = shadeSpan_radial_clamp;
+ shadeProc = use_new_proc ? shadeSpan_radial_clamp2 : shadeSpan_radial_clamp;
} else if (SkShader::kMirror_TileMode == radialGradient.fTileMode) {
shadeProc = shadeSpan_radial_mirror;
} else {