Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(461)

Issues Search
    My Issues | Starred     Open | Closed | All

Unified Diff: bench/IndexBench.cpp

Issue 1656543002: NEON for table lookups? (Closed) Base URL: https://skia.googlesource.com/skia.git@index
Patch Set: Inline assembly Created 4 years, 11 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « no previous file | bench/nanobench.cpp » ('j') | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: bench/IndexBench.cpp
diff --git a/bench/IndexBench.cpp b/bench/IndexBench.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..1d76ac82e28e62242407d3dfea325d425b3491dc
--- /dev/null
+++ b/bench/IndexBench.cpp
@@ -0,0 +1,411 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "Benchmark.h"
+#include "SkColor.h"
+
+static void swizzle_index_to_n32(
+ void* dst, const uint8_t* src, int width,
+ int bpp, int deltaSrc, int offset, const SkPMColor ctable[]) {
+
+#if defined(SK_ARM_HAS_NEON)
+ src += offset;
+ const SkPMColor* table = ctable;
+
+ while (width >= 16) {
+ ctable = table;
+ asm volatile (
+ "movi v13.8b, #0xe0 \t\n"
+
+ "ldr q14, [%[src]], #16 \t\n"
+ "mov v15.d[0], v14.d[1] \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]], #64 \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "ld4 {v4.16b-v7.16b}, [%[ctable]], #64 \t\n"
+ "ld4 {v8.16b-v11.16b}, [%[ctable]] \t\n"
+ "mov v12.16b, v5.16b \t\n"
+ "mov v5.16b, v8.16b \t\n"
+ "mov v8.16b, v12.16b \t\n"
+ "mov v12.16b, v7.16b \t\n"
+ "mov v7.16b, v10.16b \t\n"
+ "mov v10.16b, v12.16b \t\n"
+
+ "tbx v0.8b, {v4.16b, v5.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v4.16b, v5.16b}, v15.8b \t\n"
+ "mov v0.d[1], v12.d[0] \t\n"
+
+ "tbx v1.8b, {v8.16b, v9.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v8.16b, v9.16b}, v15.8b \t\n"
+ "mov v1.d[1], v12.d[0] \t\n"
+
+ "tbx v2.8b, {v6.16b, v7.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v6.16b, v7.16b}, v15.8b \t\n"
+ "mov v2.d[1], v12.d[0] \t\n"
+
+ "tbx v3.8b, {v10.16b, v11.16b}, v14.8b \t\n"
+ "tbx v12.8b, {v10.16b, v11.16b}, v15.8b \t\n"
+ "mov v3.d[1], v12.d[0] \t\n"
+
+ "add v14.8b, v14.8b, v13.8b \t\n"
+ "add v15.8b, v15.8b, v13.8b \t\n"
+
+ "st4 {v0.16b-v3.16b}, [%[dst]], #64 \t\n"
+ : [dst] "+&r" (dst), [src] "+&r" (src), [ctable] "+&r" (ctable)
+ :
+ : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
+ );
+ width -= 16;
+ }
+
+#elif 0
+ uint32_t* dst = (uint32_t*) dstRow;
+ src += offset;
+
+ while (dstWidth >= 16) {
+ // Table registers
+ uint8%[src]6x4_t t0, t1;
+ uint8x8x4_t tr, tg, tb, ta;
+
+ // Indices into table
+ uint8x16_t indices = vld1q_u8(src);
+
+ // Pixel output registers
+ uint8x16x4_t rgba;
+ rgba.val[0] = vdupq_n_u8(0);
+ rgba.val[1] = vdupq_n_u8(0);
+ rgba.val[2] = vdupq_n_u8(0);
+ rgba.val[3] = vdupq_n_u8(0);
+
+ const uint32_t* table = ctable;
+ const int numColors = 256;
+ const int numColorsPerLoop = 32;
+ for (int j = 0; j < numColors / numColorsPerLoop; j++) {
+ // Load a separate color table for each of r, g, b, a
+ t0 = vld4q_u8((const uint8_t*) (table + 0)); // rgba
+ t1 = vld4q_u8((const uint8_t*) (table + 16)); // RGBA
+ SkTSwap(t0.val[1], t1.val[0]); // rRba, gGBA
+ SkTSwap(t0.val[3], t1.val[2]); // rRbB, gGaA
+ tr = *(((uint8x8x4_t*) &t0) + 0); // rR
+ tb = *(((uint8x8x4_t*) &t0) + 1); // bB
+ tg = *(((uint8x8x4_t*) &t1) + 0); // gG
+ ta = *(((uint8x8x4_t*) &t1) + 1); // aA
+
+ // Use VTBL, then OR the results together.
+ rgba.val[0] = vorrq_u8(rgba.val[0],
+ vcombine_u8(vtbl4_u8(tr, *(((uint8x8_t*) &indices) + 0)),
+ vtbl4_u8(tr, *(((uint8x8_t*) &indices) + 1))));
+ rgba.val[1] = vorrq_u8(rgba.val[1],
+ vcombine_u8(vtbl4_u8(tg, *(((uint8x8_t*) &indices) + 0)),
+ vtbl4_u8(tg, *(((uint8x8_t*) &indices) + 1))));
+ rgba.val[2] = vorrq_u8(rgba.val[2],
+ vcombine_u8(vtbl4_u8(tb, *(((uint8x8_t*) &indices) + 0)),
+ vtbl4_u8(tb, *(((uint8x8_t*) &indices) + 1))));
+ rgba.val[3] = vorrq_u8(rgba.val[3],
+ vcombine_u8(vtbl4_u8(ta, *(((uint8x8_t*) &indices) + 0)),
+ vtbl4_u8(ta, *(((uint8x8_t*) &indices) + 1))));
+
+ // Move the next set of indices into the range of the color table. Indices
+ // that are currently in range should go out of range due to underflow.
+ indices = vsubq_u8(indices, vdupq_n_u8(32));
+ table += numColorsPerLoop;
+ }
+
+ // Store output pixel values.
+ vst4q_u8((uint8_t*) dst, rgba);
+
+ src += 16;
+ dst += 16;
+ dstWidth -= 16;
+ }
+#else
+ src += offset;
+ SkPMColor* dst = (SkPMColor*)dstRow;
+ for (int x = 0; x < dstWidth; x++) {
+ SkPMColor c = ctable[*src];
+ dst[x] = c;
+ src += deltaSrc;
+ }
+#endif
+}
+
+class IndexBench : public Benchmark {
+public:
+ IndexBench() {}
+
+ bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }
+ const char* onGetName() override { return "IndexYay"; }
+
+ void onDelayedSetup() override {
+ fColorTable.reset(256);
+ uint32_t* ptr = fColorTable.get();
+ for (int i = 0; i < 256; i++) {
+ ptr[i] = (i * 37 - 52) * 49;
+ }
+
+ fSrc.reset(1024);
+ uint8_t* p = fSrc.get();
+ for (int i = 0; i < 1024; i++) {
+ p[i] = i % 256;
+ }
+
+ fDst.reset(1024);
+ }
+
+ void onDraw(int loops, SkCanvas*) override {
+ while (loops --> 0) {
+ swizzle_index_to_n32(fDst, fSrc, 1024, 1, 1, 0, fColorTable);
+ }
+ }
+private:
+ SkAutoTMalloc<uint32_t> fDst;
+ SkAutoTMalloc<uint8_t> fSrc;
+ SkAutoTMalloc<uint32_t> fColorTable;
+};
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+DEF_BENCH(return new IndexBench());
« no previous file with comments | « no previous file | bench/nanobench.cpp » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698