Reland "Adding a trace to nanobench for RP/VM comparison"

This reverts commit 39728eb980.

Reason for revert: Need to reland after I figure out what happened

Original change's description:
> Revert "Adding a trace to nanobench for RP/VM comparison"
>
> This reverts commit e855b1273f.
>
> Reason for revert: breaking the tree
>
> Original change's description:
> > Adding a trace to nanobench for RP/VM comparison
> >
> > Run nanobench with --compare key and collect data for comparison:
> > ./out/Release/nanobench --csv --config 8888 --skvm --compare --loops 100 --samples 1 --match $(ls skps | grep --invert-match svg ) 2>&1 | tee VM.data
> > ./out/Release/nanobench --csv --config 8888 --forceRasterPipeline --compare --loops 100 --samples 1 --match $(ls skps | grep --invert-match svg ) 2>&1 | tee RP.data
> > awk 'BEGIN {OFS=","; fileNum = 0} ($2 ~ /MB/) && fileNum == 0 {vmvmcycles[$3] = $6; vmvmscan[$3] = $8; vmvmpixels[$3] = $10; vmvminterp[$3] = $11;  vmrpcycle[$3] = $14; vmrpscan[$3] = $16; vmrppixels[$3] = $18} ($2 ~ /MB/) && fileNum == 1  {print $3, vmvmcycles[$3], vmvmscan[$3], vmvmpixels[$3], vmvminterp[$3], $6, $8, $10, $11, $14, $16, $18} ENDFILE {fileNum += 1}' VM.data RP.data > compare.csv
> >
> > You can see an example of comparison table here:
> > https://docs.google.com/spreadsheets/d/1Q57oz6Jn8JPQkPzUbtO0fcZh2VhwPhxwsLaHnSb9uR0/edit#gid=406063636
> >
> > Change-Id: I8fa35e3fb087bce00ab19355a3bc021334aa7a80
> > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/501337
> > Reviewed-by: Herb Derby <herb@google.com>
> > Commit-Queue: Julia Lavrova <jlavrova@google.com>
>
> Change-Id: I6799b90ffd8bc05c65d1185bfc07a06af05e2038
> No-Presubmit: true
> No-Tree-Checks: true
> No-Try: true
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/507182
> Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
> Commit-Queue: Julia Lavrova <jlavrova@google.com>

Change-Id: I3668b5a40f310598e23c922b1eb6f93db4dac13c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/507183
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Julia Lavrova <jlavrova@google.com>

bench/nanobench.cpp

@@ -36,6 +36,7 @@
 #include "src/core/SkOSFile.h"
 #include "src/core/SkTaskGroup.h"
 #include "src/core/SkTraceEvent.h"
+#include "src/utils/SkBlitterTrace.h"
 #include "src/utils/SkJSONWriter.h"
 #include "src/utils/SkOSPath.h"
 #include "src/utils/SkShaderUtils.h"
@@ -72,6 +73,9 @@ extern bool gUseSkVMBlitter;
 extern bool gSkVMAllowJIT;
 extern bool gSkVMJITViaDylib;
 
+extern SkBlitterTrace gSkVMBlitterTrace;
+extern SkBlitterTrace gSkRPBlitterTrace;
+
 #ifndef SK_BUILD_FOR_WIN
     #include <unistd.h>
 
@@ -191,6 +195,8 @@ static DEFINE_string(properties, "",
 static DEFINE_bool(purgeBetweenBenches, false,
                    "Call SkGraphics::PurgeAllCaches() between each benchmark?");
 
+static DEFINE_bool(compare, false, "Prepare results for Raster Pipeline/SkVM comparison.");
+
 static double now_ms() { return SkTime::GetNSecs() * 1e-6; }
 
 static SkString humanize(double ms) {
@@ -281,6 +287,9 @@ struct GPUTarget : public Target {
     }
 };
 
+static SkBlitterTrace gSkVMBlitterTraceCapture("VM");
+static SkBlitterTrace gSkRPBlitterTraceCapture("RP");
+
 static double time(int loops, Benchmark* bench, Target* target) {
     SkCanvas* canvas = target->getCanvas();
     if (canvas) {
@@ -289,7 +298,13 @@ static double time(int loops, Benchmark* bench, Target* target) {
     bench->preDraw(canvas);
     double start = now_ms();
     canvas = target->beginTiming(canvas);
+
+    gSkVMBlitterTrace.reset();
+    gSkRPBlitterTrace.reset();
     bench->draw(loops, canvas);
+    gSkVMBlitterTraceCapture = gSkVMBlitterTrace;
+    gSkRPBlitterTraceCapture = gSkRPBlitterTrace;
+
     target->endTiming();
     double elapsed = now_ms() - start;
     bench->postDraw(canvas);
@@ -1266,6 +1281,11 @@ int main(int argc, char** argv) {
     gSkVMAllowJIT = FLAGS_jit;
     gSkVMJITViaDylib = FLAGS_dylib;
 
+    if (FLAGS_compare) {
+        gSkVMBlitterTrace.turnTrace(true);
+        gSkRPBlitterTrace.turnTrace(true);
+    }
+
     int runs = 0;
     BenchmarkStream benchStream;
     log.beginObject("results");
@@ -1414,11 +1434,18 @@ int main(int argc, char** argv) {
                 if (configs.count() == 1) {
                     config = ""; // Only print the config if we run the same bench on more than one.
                 }
-                SkDebugf("%4d/%-4dMB\t%s\t%s\n"
+                SkDebugf("%4d/%-4dMB\t%s\t%s "
                          , sk_tools::getCurrResidentSetSizeMB()
                          , sk_tools::getMaxResidentSetSizeMB()
                          , bench->getUniqueName()
                          , config);
+                if (FLAGS_compare) {
+                    gSkVMBlitterTraceCapture.printCounts("Total");
+                    SkDebugf("0 ");
+                    gSkRPBlitterTraceCapture.printCounts("Total");
+                }
+
+                SkDebugf("\n");
             } else if (FLAGS_quiet) {
                 const char* mark = " ";
                 const double stddev_percent =

gn/utils.gni

@@ -32,6 +32,7 @@ skia_utils_sources = [
   "$_src/utils/SkAnimCodecPlayer.cpp",
   "$_src/utils/SkBase64.cpp",
   "$_src/utils/SkBitSet.h",
+  "$_src/utils/SkBlitterTrace.h",
   "$_src/utils/SkCallableTraits.h",
   "$_src/utils/SkCamera.cpp",
   "$_src/utils/SkCanvasStack.cpp",
@@ -42,6 +43,7 @@ skia_utils_sources = [
   "$_src/utils/SkClipStackUtils.cpp",
   "$_src/utils/SkClipStackUtils.h",
   "$_src/utils/SkCustomTypeface.cpp",
+  "$_src/utils/SkCycles.h",
   "$_src/utils/SkDashPath.cpp",
   "$_src/utils/SkDashPathPriv.h",
   "$_src/utils/SkEventTracer.cpp",

src/core/SkBlitter.cpp

@@ -632,7 +632,7 @@ SkBlitter* SkBlitterClipper::apply(SkBlitter* blitter, const SkRegion* clip,
 bool SkBlitter::UseLegacyBlitter(const SkPixmap& device,
                                  const SkPaint& paint,
                                  const SkMatrix& matrix) {
-    if (gSkForceRasterPipelineBlitter) {
+    if (gSkForceRasterPipelineBlitter || gUseSkVMBlitter) {
         return false;
     }
 #if defined(SK_FORCE_RASTER_PIPELINE_BLITTER)
@@ -734,20 +734,15 @@ SkBlitter* SkBlitter::Choose(const SkPixmap& device,
         paint.writable()->setDither(false);
     }
 
-    if (gUseSkVMBlitter) {
-        if (auto blitter = SkVMBlitter::Make(device, *paint, matrixProvider,
-                                             alloc, clipShader)) {
-            return blitter;
-        }
-    }
-
     // Same basic idea used a few times: try SkRP, then try SkVM, then give up with a null-blitter.
     // (Setting gUseSkVMBlitter is the only way we prefer SkVM over SkRP at the moment.)
     auto create_SkRP_or_SkVMBlitter = [&]() -> SkBlitter* {
+        if (!gUseSkVMBlitter) {
             if (auto blitter = SkCreateRasterPipelineBlitter(device, *paint, matrixProvider,
                                                              alloc, clipShader)) {
                 return blitter;
             }
+        }
         if (auto blitter = SkVMBlitter::Make(device, *paint, matrixProvider,
                                              alloc, clipShader)) {
             return blitter;

src/core/SkBlitter_Sprite.cpp

@@ -17,6 +17,7 @@
 #include "src/core/SkVMBlitter.h"
 
 extern bool gUseSkVMBlitter;
+extern bool gSkForceRasterPipelineBlitter;
 
 SkSpriteBlitter::SkSpriteBlitter(const SkPixmap& source)
     : fSource(source) {}
@@ -199,7 +200,9 @@ SkBlitter* SkBlitter::ChooseSprite(const SkPixmap& dst, const SkPaint& paint,
 
     SkSpriteBlitter* blitter = nullptr;
 
-    if (0 == SkColorSpaceXformSteps(source,dst).flags.mask() && !clipShader) {
+    if (gSkForceRasterPipelineBlitter) {
+        // Do not use any of these optimized memory blitters
+    } else if (0 == SkColorSpaceXformSteps(source,dst).flags.mask() && !clipShader) {
         if (!blitter && SkSpriteBlitter_Memcpy::Supports(dst, source, paint)) {
             blitter = alloc->make<SkSpriteBlitter_Memcpy>(source);
         }

src/core/SkRasterPipelineBlitter.cpp

@@ -20,6 +20,9 @@
 #include "src/core/SkRasterPipeline.h"
 #include "src/core/SkUtils.h"
 #include "src/shaders/SkShaderBase.h"
+#include "src/utils/SkBlitterTrace.h"
+
+SkBlitterTrace gSkRPBlitterTrace("RP", false);
 
 class SkRasterPipelineBlitter final : public SkBlitter {
 public:
@@ -47,6 +50,7 @@ public:
     void blitV     (int x, int y, int height, SkAlpha alpha)        override;
 
 private:
+    void blitRectWithTrace(int x, int y, int w, int h, bool trace);
     void append_load_dst      (SkRasterPipeline*) const;
     void append_store         (SkRasterPipeline*) const;
 
@@ -339,7 +343,15 @@ void SkRasterPipelineBlitter::blitH(int x, int y, int w) {
 }
 
 void SkRasterPipelineBlitter::blitRect(int x, int y, int w, int h) {
+    this->blitRectWithTrace(x, y, w, h, true);
+}
+
+void SkRasterPipelineBlitter::blitRectWithTrace(int x, int y, int w, int h, bool trace) {
     if (fMemset2D) {
+        SkBlitterTrace::Step trace1(trace ? &gSkRPBlitterTrace : nullptr,
+                                    "blitRectByMemset",
+                                    /*scanlines=*/h,
+                                    /*pixels=*/w * h);
         fMemset2D(&fDst, x,y, w,h, fMemsetColor);
         return;
     }
@@ -373,6 +385,10 @@ void SkRasterPipelineBlitter::blitRect(int x, int y, int w, int h) {
         fBlitRect = p.compile();
     }
 
+    SkBlitterTrace::Step trace2(trace ? &gSkRPBlitterTrace : nullptr,
+                                "blitRect",
+                                /*scanlines=*/h,
+                                /*pixels=*/w * h);
     fBlitRect(x,y,w,h);
 }
 
@@ -397,10 +413,12 @@ void SkRasterPipelineBlitter::blitAntiH(int x, int y, const SkAlpha aa[], const
         fBlitAntiH = p.compile();
     }
 
+    SkBlitterTrace::Step trace(&gSkRPBlitterTrace, "blitAntiH", /*scanlines=*/1ul, /*pixels=*/0ul);
     for (int16_t run = *runs; run > 0; run = *runs) {
+        trace.add(/*scanlines=*/0, /*pixels=*/run);
         switch (*aa) {
             case 0x00:                                break;
-            case 0xff: this->blitH(x,y,run); break;
+            case 0xff:this->blitRectWithTrace(x,y,run, 1, false); break;
             default:
                 fCurrentCoverage = *aa * (1/255.0f);
                 fBlitAntiH(x,y,run,1);
@@ -554,5 +572,9 @@ void SkRasterPipelineBlitter::blitMask(const SkMask& mask, const SkIRect& clip)
     }
 
     SkASSERT(blitter);
+    SkBlitterTrace::Step trace(&gSkRPBlitterTrace,
+                               "blitMask",
+                               /*scanlines=*/clip.height(),
+                               /*pixels=*/clip.width() * clip.height());
     (*blitter)(clip.left(),clip.top(), clip.width(),clip.height());
 }

src/core/SkVMBlitter.cpp

@@ -21,9 +21,12 @@
 #include "src/core/SkVM.h"
 #include "src/core/SkVMBlitter.h"
 #include "src/shaders/SkColorFilterShader.h"
+#include "src/utils/SkBlitterTrace.h"
 
 #include <cinttypes>
 
+SkBlitterTrace gSkVMBlitterTrace("VM", false);
+
 namespace {
 
     // Uniforms set by the Blitter itself,
@@ -667,8 +670,10 @@ void SkVMBlitter::blitH(int x, int y, int w) {
     skvm::Program* blit_h = this->buildProgram(Coverage::Full);
     this->updateUniforms(x+w, y);
     if (const void* sprite = this->isSprite(x,y)) {
+        SkBlitterTrace::Step trace(&gSkVMBlitterTrace, "blitH1", /*scanlines=*/1, /*pixels=*/w);
         blit_h->eval(w, fUniforms.buf.data(), fDevice.addr(x,y), sprite);
     } else {
+        SkBlitterTrace::Step trace(&gSkVMBlitterTrace, "blitH2", /*scanlines=*/1, /*pixels=*/w);
         blit_h->eval(w, fUniforms.buf.data(), fDevice.addr(x,y));
     }
 }
@@ -677,7 +682,9 @@ void SkVMBlitter::blitAntiH(int x, int y, const SkAlpha cov[], const int16_t run
     skvm::Program* blit_anti_h = this->buildProgram(Coverage::UniformF);
     skvm::Program* blit_h = this->buildProgram(Coverage::Full);
 
+    SkBlitterTrace::Step trace(&gSkVMBlitterTrace, "blitAntiH", /*scanlines=*/1ul, /*pixels=*/0ul);
     for (int16_t run = *runs; run > 0; run = *runs) {
+        trace.add(/*scanlines=*/0, /*pixels=*/run);
         const SkAlpha coverage = *cov;
         if (coverage != 0x00) {
             this->updateUniforms(x+run, y);
@@ -727,6 +734,11 @@ void SkVMBlitter::blitMask(const SkMask& mask, const SkIRect& clip) {
 
     SkASSERT(program);
     if (program) {
+        SkBlitterTrace::Step trace(&gSkVMBlitterTrace,
+                                   "blitMask",
+                                   /*scanlines=*/clip.height(),
+                                   /*pixels=*/clip.width() * clip.height());
+
         for (int y = clip.top(); y < clip.bottom(); y++) {
              int x = clip.left(),
                  w = clip.width();

src/core/SkVMBlitter.h

@@ -91,6 +91,8 @@ private:
 
     void blitH(int x, int y, int w) override;
     void blitAntiH(int x, int y, const SkAlpha cov[], const int16_t runs[]) override;
+
+private:
     void blitMask(const SkMask& mask, const SkIRect& clip) override;
 
     SkPixmap        fDevice;

src/utils/SkBlitterTrace.h

@@ -0,0 +1,140 @@
+/*
+ * Copyright 2022 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkBlitterTrace_DEFINED
+#define SkBlitterTrace_DEFINED
+
+#include <inttypes.h>
+#include <unordered_map>
+#include "src/utils/SkCycles.h"
+
+class SkBlitterTrace {
+/*
+ * This class collects information for RasterPipeLine vs SkVM
+ * performance comparison.
+ * How to get the comparison table:
+ * 1. Run nanobench for SkVM:
+ *    []/Release/nanobench
+ *    --csv --config 8888 --skvm --compare --loops 100 --samples 1
+ *    --match $(ls skps | grep --invert-match svg ) 2>&1 | tee VM.data
+ * 2. Run nanobench for RasterPipeLine:
+ *    []/Release/nanobench
+ *    --csv --config 8888 --forceRasterPipeline --compare --loops 100
+ *    --samples 1 --match $(ls skps | grep --invert-match svg )
+ *    2>&1 | tee RP.data
+ * 3. Extract the information side-by-side:
+ *    awk 'BEGIN {OFS=","; fileNum = 0} ($2 ~ /MB/) && fileNum == 0
+ *    {vmvmcycles[$3] = $6; vmvmscan[$3] = $8; vmvmpixels[$3] = $10;
+ *    vmvminterp[$3] = $11;  vmrpcycle[$3] = $14; vmrpscan[$3] = $16;
+ *    vmrppixels[$3] = $18} ($2 ~ /MB/) && fileNum == 1  {print $3,
+ *    vmvmcycles[$3], vmvmscan[$3], vmvmpixels[$3], vmvminterp[$3], $6, $8,
+ *    $10, $11, $14, $16, $18} ENDFILE {fileNum += 1}'
+ *    VM.data RP.data > compare.csv
+ * 4. Open the compare.csv table in Google Spreadsheets.
+ *     You will get columns [A:P]. Add 4 more columns with formulas:
+ *     Q: =B/M-1
+ *     R: =N-C
+ *     S: =O-D
+ *     T: =2*(S<>0)+(R<>0)
+ *     To be honest R, S, T columns are here for checking only (they all
+ *     supposed to have zero values in them)
+ *     Column Q shows the actual performance difference. Negative value means
+ *     that wins SkVM, positive - RasterPipeLine.
+ */
+public:
+    SkBlitterTrace(const char* header, bool traceSteps = false)
+        : fHeader(header), fTraceSteps(traceSteps) { }
+
+    SkBlitterTrace& operator= (const SkBlitterTrace&) = default;
+
+    void addTrace(const char* name, uint64_t cycles, uint64_t scanLines, uint64_t pixels) {
+        fCycles += cycles;
+        fScanlines += scanLines;
+        fPixels += pixels;
+        if (fTraceSteps) {
+            printIncrements(name, cycles, scanLines, pixels);
+        }
+    }
+
+    void reset() {
+        fCycles = 0ul;
+        fScanlines = 0ul;
+        fPixels = 0ul;
+    }
+
+    void printIncrements(const char* name,
+                         uint64_t cycles,
+                         uint64_t scanLines,
+                         uint64_t pixels) const {
+        SkDebugf("%s %s: cycles=%" PRIu64 "+%" PRIu64
+                       " scanlines=%" PRIu64 "+%" PRIu64 " pixels=%" PRIu64,
+                 fHeader, name,
+                 fCycles - cycles, cycles,
+                 fScanlines - scanLines, scanLines,
+                 fPixels);
+        SkDebugf("\n");
+    }
+
+    void printCounts(const char* name) const {
+        SkDebugf("%s cycles: %" PRIu64 " "
+                 "   scanlines: %" PRIu64 " pixels: %" PRIu64,
+                 fHeader,
+                 fCycles,
+                 fScanlines,
+                 fPixels);
+        SkDebugf(" ");
+    }
+
+    void turnTrace(bool value) { fTraceTime = value; }
+    uint64_t getCycles() const { return fCycles; }
+    uint64_t getScanlines() const { return fScanlines; }
+    uint64_t getPixels() const { return fPixels; }
+
+    class Step {
+    public:
+        Step(SkBlitterTrace* trace,
+             const char* name,
+             uint64_t scanlines,
+             uint64_t pixels)
+            : fTrace(trace)
+            , fName(name)
+            , fScanlines(scanlines)
+            , fPixels(pixels) {
+            fStartTime = SkCycles::Now();
+        }
+        void add(uint64_t scanlines, uint64_t pixels) {
+            fScanlines += scanlines;
+            fPixels += pixels;
+        }
+        ~Step() {
+            if (fTrace == nullptr || !fTrace->fTraceTime) {
+                return;
+            }
+            auto endTime = SkCycles::Now() - fStartTime;
+            fTrace->addTrace(/*name=*/fName,
+                             /*cycles=*/endTime,
+                             /*scanlines=*/fScanlines,
+                             /*pixels=*/fPixels);
+        }
+    private:
+        SkBlitterTrace* fTrace = nullptr;
+        const char* fName = "";
+        uint64_t fStartTime = 0ul;
+        uint64_t fScanlines = 0ul;
+        uint64_t fPixels = 0ul;
+    };
+
+private:
+    const char* fHeader = "";
+    bool fTraceSteps = false;
+    bool fTraceTime = false;
+    uint64_t fCycles = 0ul;
+    uint64_t fScanlines = 0ul;
+    uint64_t fPixels = 0ul;
+};
+
+#endif

src/utils/SkCycles.h

@@ -0,0 +1,39 @@
+/*
+ * Copyright 2021 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+#ifndef SkCycles_DEFINED
+#define SkCycles_DEFINED
+#include <cstdint>
+class SkCycles {
+public:
+    static uint64_t Now() {
+        #if defined(SK_BUILD_FOR_WIN)
+        {
+            return 0ul;
+        }
+        #elif defined(SK_BUILD_FOR_IOS)
+        {
+            return 0ul;
+        }
+        #elif defined(SK_CPU_X86)
+        {
+            uint32_t cpuInfo;
+            return __builtin_ia32_rdtscp(&cpuInfo);
+        }
+        #elif defined(SK_CPU_ARM64)
+        {
+            int64_t cycles;
+            asm volatile("mrs %0, cntvct_el0" : "=r"(cycles));
+            return cycles;
+        }
+        #else
+        {
+            return 0ul;
+        }
+        #endif
+    }
+};
+#endif  // SkCycles_DEFINED