swung0x48/MobileGL
1
0
Fork 0
mirror of https://github.com/MobileGL-Dev/MobileGL synced 2026-11-08 02:23:23 +09:00
Code Issues Packages Projects Releases Wiki Activity

[Perf] (ProgramState): serve a whole linked program from translation cache L1, skipping the link entirely

Browse Source
This commit is contained in:
Swung0x48
2026-08-20 12:00:01 -04:00
parent 14744f117c
commit 1eeeb44d94
9 changed files with 348 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files
MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.cpp
+33 -33
View File
@@ -12,6 +12,7 @@
#include <MG_Util/Async/ShaderCompilePool.h>
#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
#include <MG_Util/ShaderTranspiler/SpvcSession.h>
#include <MG_State/GLState/ProgramState/ProgramTranslationCache.h>
#include <MG_Util/ShaderTranspiler/TranslationCache.h>
#include <MG_Util/ShaderTranspiler/Types.h>
@@ -96,11 +97,25 @@ namespace MobileGL::MG_State::GLState {
// and `diagnostics`, and this node is the sole reader of the handoff.
ProgramLinkTask::SpirvHandoff& handoff = m_phaseA->spirvHandoff;
const Uint externalIndex = m_phaseA->in.externalIndex;
if (!handoff.ready || !handoff.reflection.program) {
if (!handoff.ready) {
// Phase A did not reach its tail (it failed the link, or was cancelled mid-body).
// Publish nothing; spirvStatus stays false.
return;
}
// A TProgram is required only to GENERATE. A link served from the L1 memo has none by
// construction - that is the entire point of the widened payload - and its SPIR-V and
// routing tables arrive ready-made in cachedSpirv.
if (!handoff.cachedSpirv && !handoff.reflection.program) return;
// An L1 hit already carries everything this phase would have produced. Publish it
// and stop: no GlslangToSpv, no spirv-opt, no routing pass.
if (handoff.cachedSpirv) {
artifacts = *handoff.cachedSpirv;
MGLOG_D("ProgramObject %u: L1 cache hit - %zu SPIR-V module(s) and the global-UBO "
"routing reused",
externalIndex, artifacts.generatedSpirv.size());
return;
}
MGLOG_D("ProgramObject %u: Starting SPIR-V generation", externalIndex);
const Bool deferOutputValidationForDirectVulkan =
@@ -138,6 +153,23 @@ namespace MobileGL::MG_State::GLState {
MGLOG_D("ProgramObject %u: Building global-UBO routing tables", externalIndex);
BuildGlobalUboRouting(handoff, externalIndex);
// The completed front end goes into the L1 memo HERE, where both halves exist: phase
// A's LinkArtifacts (carried in the handoff) and this phase's SpirvArtifacts.
//
// Only a clean run is memoized. A failed optimizer run leaves a module as whatever the
// chain got to before it gave up, and that is exactly the binary no other program
// should ever be handed.
if (artifacts.spirvStatus && handoff.spirvCacheKey.Valid() && handoff.linkArtifactsForCache) {
auto payload = MakeShared<ProgramTranslationResult>();
payload->link = *handoff.linkArtifactsForCache;
payload->link.program.reset(); // belt and braces: never memoize a glslang arena
payload->spirv = artifacts;
const SizeT payloadBytes = ProgramTranslationResultBytes(*payload);
GetProgramTranslationCache().Insert(handoff.spirvCacheKey,
ProgramTranslationResultPtr(Move(payload)),
payloadBytes);
}
MGLOG_D("ProgramObject %u: Binary generation finished (generatedSpirv size=%zu)", externalIndex,
artifacts.generatedSpirv.size());
}
@@ -153,28 +185,6 @@ namespace MobileGL::MG_State::GLState {
using namespace MG_Util::ShaderTranspiler;
MGLOG_D("ProgramObject %u: GenerateSpirv - start", externalIndex);
// L1 of the shader translation memo. The segment this short-circuits is the whole
// of GlslangToSpv plus the 11-pass SanitizeAndOptimizeBinary chain, for every stage
// of the program at once - ~136 us per stage on the RelWithDebInfo host measurement.
// The key was built at the tail of phase A (ProgramLinkTask::BuildSpirvCacheKey) and
// covers every input that can move these bytes; see TranslationCache.h.
//
// Note what a HIT does NOT skip: the glslang parse and link, which already happened
// in phase A because the frontend's whole GL query surface is built out of the
// TProgram they produce.
auto& spirvCache = GetSpirvTranslationCache();
const TranslationCacheKey& cacheKey = handoff.spirvCacheKey;
if (cacheKey.Valid()) {
if (const SpirvTranslationResultPtr hit = spirvCache.Find(cacheKey);
hit && hit->modules.size() == handoff.shaderTypes.size()) {
artifacts.generatedSpirv = hit->modules;
artifacts.spirvStatus = true;
MGLOG_D("ProgramObject %u: GenerateSpirv - L1 cache hit, %zu module(s) reused",
externalIndex, artifacts.generatedSpirv.size());
return;
}
}
// The shaders were parsed once, in the link-compatible (relaxed Vulkan-rules)
// configuration, and the handoff's program linked those parses - so it IS the program
// the backends consume. Generate SPIR-V straight from its intermediates, which the
@@ -218,16 +228,6 @@ namespace MobileGL::MG_State::GLState {
}
}
artifacts.spirvStatus = allOptimized;
// Only a clean run is memoized. A failed optimizer run leaves `spv` as whatever the
// chain got to before it gave up, and that is exactly the binary no other program
// should ever be handed.
if (allOptimized && cacheKey.Valid()) {
auto payload = MakeShared<SpirvTranslationResult>();
payload->modules = artifacts.generatedSpirv;
const SizeT payloadBytes = SpirvTranslationResultBytes(*payload);
spirvCache.Insert(cacheKey, SpirvTranslationResultPtr(Move(payload)), payloadBytes);
}
}
void ProgramSpirvTask::BuildGlobalUboRouting(const ProgramLinkTask::SpirvHandoff& handoff,