swung0x48/MobileGL
1
0
Fork 0
mirror of https://github.com/MobileGL-Dev/MobileGL synced 2026-09-12 06:08:30 +09:00
Code Issues Packages Projects Releases Wiki Activity

[Bench] (ShaderTranspiler): bracket the translation-cache win with a CTS-sized and a heavy stage

Browse Source
This commit is contained in:
Swung0x48
2026-08-20 11:11:37 -04:00
parent d48e5d0053
commit 5fecfa42f6
1 changed files with 20 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
MobileGL/MG_Benchmark/ShaderCache/TranslationCacheBench.cpp
+20 -12
View File
@@ -61,7 +61,11 @@ void main() {
// TEXTURE_ACCESS, one CHANNEL, and an output whose BASIC_TYPE is the only thing that // TEXTURE_ACCESS, one CHANNEL, and an output whose BASIC_TYPE is the only thing that
// varies within a case. Padded with enough real arithmetic that the translation chain // varies within a case. Padded with enough real arithmetic that the translation chain
// is doing work rather than measuring fixed overheads. // is doing work rather than measuring fixed overheads.
String SwizzleLikeFragment(const String& prefix) { // `padLines` = 0 is the honest CTS size: gl3cTextureSwizzleTests' smoke template is a
// handful of lines, and that is the workload the memo exists for. The padded variant is
// kept alongside it because a shaderpack stage is orders of magnitude bigger, and the
// two bracket the ratio the cache is worth in practice.
String SwizzleLikeFragment(const String& prefix, const int padLines) {
String source = "#version 460\n"; String source = "#version 460\n";
source += "in vec3 vPos;\n"; source += "in vec3 vPos;\n";
source += "in vec2 vUv;\n"; source += "in vec2 vUv;\n";
@@ -73,7 +77,7 @@ void main() {
source += "void main() {\n"; source += "void main() {\n";
source += " vec4 s = texture(uTex, vUv);\n"; source += " vec4 s = texture(uTex, vUv);\n";
source += " float acc = s.r;\n"; source += " float acc = s.r;\n";
for (int i = 0; i < 120; ++i) { for (int i = 0; i < padLines; ++i) {
source += " acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0) * cos(acc);\n"; source += " acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0) * cos(acc);\n";
} }
source += " for (int i = 0; i < 8; ++i) acc += uArr[i];\n"; source += " for (int i = 0; i < 8; ++i) acc += uArr[i];\n";
@@ -195,41 +199,44 @@ void main() {
// --------------------------------------------------------------------------------------- // ---------------------------------------------------------------------------------------
// L1, in situ: the full glCompileShader + glLinkProgram path for a repeated program. // L1, in situ: the full glCompileShader + glLinkProgram path for a repeated program.
// --------------------------------------------------------------------------------------- // ---------------------------------------------------------------------------------------
// Arg(0) = the CTS smoke size; Arg(120) = a heavy stage, bracketing the ratio.
static void BM_ProgramLink_CacheOff(benchmark::State& state) { static void BM_ProgramLink_CacheOff(benchmark::State& state) {
MobileGL::Initialize(); MobileGL::Initialize();
const SyncCompileScope sync; const SyncCompileScope sync;
const CacheModeScope cache(false); const CacheModeScope cache(false);
const String vs = kVertexSource; const String vs = kVertexSource;
const String fs = SwizzleLikeFragment(""); const String fs = SwizzleLikeFragment("", static_cast<int>(state.range(0)));
for (auto _ : state) { for (auto _ : state) {
LinkOneProgram(vs, fs); LinkOneProgram(vs, fs);
} }
state.SetLabel("MOBILEGL_SHADER_CACHE=0"); state.SetLabel("MOBILEGL_SHADER_CACHE=0");
} }
BENCHMARK(BM_ProgramLink_CacheOff)->Unit(benchmark::kMicrosecond); BENCHMARK(BM_ProgramLink_CacheOff)->Arg(0)->Arg(120)->Unit(benchmark::kMicrosecond);
static void BM_ProgramLink_CacheOn(benchmark::State& state) { static void BM_ProgramLink_CacheOn(benchmark::State& state) {
MobileGL::Initialize(); MobileGL::Initialize();
const SyncCompileScope sync; const SyncCompileScope sync;
const CacheModeScope cache(true); const CacheModeScope cache(true);
const String vs = kVertexSource; const String vs = kVertexSource;
const String fs = SwizzleLikeFragment(""); const String fs = SwizzleLikeFragment("", static_cast<int>(state.range(0)));
LinkOneProgram(vs, fs); // prime, so the measured loop is the steady state LinkOneProgram(vs, fs); // prime, so the measured loop is the steady state
const TranslationCacheStats before = GetSpirvTranslationCache().Stats();
for (auto _ : state) { for (auto _ : state) {
LinkOneProgram(vs, fs); LinkOneProgram(vs, fs);
} }
const TranslationCacheStats stats = GetSpirvTranslationCache().Stats(); const TranslationCacheStats stats = GetSpirvTranslationCache().Stats();
state.counters["L1_hits"] = static_cast<double>(stats.hits); state.counters["L1_hits"] = static_cast<double>(stats.hits - before.hits);
state.counters["L1_misses"] = static_cast<double>(stats.misses); state.counters["L1_misses"] = static_cast<double>(stats.misses - before.misses);
} }
BENCHMARK(BM_ProgramLink_CacheOn)->Unit(benchmark::kMicrosecond); BENCHMARK(BM_ProgramLink_CacheOn)->Arg(0)->Arg(120)->Unit(benchmark::kMicrosecond);
// --------------------------------------------------------------------------------------- // ---------------------------------------------------------------------------------------
// L2, component: the DirectGLES SPIR-V pass chain plus SPIRV-Cross for one stage. // L2, component: the DirectGLES SPIR-V pass chain plus SPIRV-Cross for one stage.
// --------------------------------------------------------------------------------------- // ---------------------------------------------------------------------------------------
static void BM_EsslTranspile_CacheOff(benchmark::State& state) { static void BM_EsslTranspile_CacheOff(benchmark::State& state) {
MobileGL::Initialize(); MobileGL::Initialize();
const Vector<Uint32> spirv = BuildSanitizedFragmentSpirv(SwizzleLikeFragment("")); const Vector<Uint32> spirv =
BuildSanitizedFragmentSpirv(SwizzleLikeFragment("", static_cast<int>(state.range(0))));
if (spirv.empty()) { if (spirv.empty()) {
state.SkipWithError("could not build the fragment module"); state.SkipWithError("could not build the fragment module");
return; return;
@@ -244,11 +251,12 @@ static void BM_EsslTranspile_CacheOff(benchmark::State& state) {
} }
state.SetLabel("MOBILEGL_SHADER_CACHE=0"); state.SetLabel("MOBILEGL_SHADER_CACHE=0");
} }
BENCHMARK(BM_EsslTranspile_CacheOff)->Unit(benchmark::kMicrosecond); BENCHMARK(BM_EsslTranspile_CacheOff)->Arg(0)->Arg(120)->Unit(benchmark::kMicrosecond);
static void BM_EsslTranspile_CacheOn(benchmark::State& state) { static void BM_EsslTranspile_CacheOn(benchmark::State& state) {
MobileGL::Initialize(); MobileGL::Initialize();
const Vector<Uint32> spirv = BuildSanitizedFragmentSpirv(SwizzleLikeFragment("")); const Vector<Uint32> spirv =
BuildSanitizedFragmentSpirv(SwizzleLikeFragment("", static_cast<int>(state.range(0))));
if (spirv.empty()) { if (spirv.empty()) {
state.SkipWithError("could not build the fragment module"); state.SkipWithError("could not build the fragment module");
return; return;
@@ -273,6 +281,6 @@ static void BM_EsslTranspile_CacheOn(benchmark::State& state) {
state.counters["L2_hits"] = static_cast<double>(stats.hits); state.counters["L2_hits"] = static_cast<double>(stats.hits);
state.counters["L2_misses"] = static_cast<double>(stats.misses); state.counters["L2_misses"] = static_cast<double>(stats.misses);
} }
BENCHMARK(BM_EsslTranspile_CacheOn)->Unit(benchmark::kMicrosecond); BENCHMARK(BM_EsslTranspile_CacheOn)->Arg(0)->Arg(120)->Unit(benchmark::kMicrosecond);
BENCHMARK_MAIN(); BENCHMARK_MAIN();