From 59235d9cb0de5c27dcba2160a75188da9b585bf3 Mon Sep 17 00:00:00 2001 From: LeandroLP23 Date: Tue, 13 Oct 2026 15:29:43 +0200 Subject: [PATCH] test(agc): compile each v_cmp denormals program once for its eight input groups The test recompiled and rebuilt the same program for every group of 32 input pairs. Only the input buffer changes between groups, so compile once per compare, EXEC/VCC capture and float mode (96 programs instead of 768) and dispatch it for each group. On lavapipe the test drops from about 11.5 s to 1.3 s; it is the slowest test of the Linux CI run (14.5 s). Co-Authored-By: Claude Opus 5.5 --- .../tests/execution/VCmpF32Denormals.cpp | 28 +++++++++++-------- 1 file changed, 16 insertions(+), 12 deletions(-) diff --git a/core/libs/prx/libSceAgcDriver/tests/execution/VCmpF32Denormals.cpp b/core/libs/prx/libSceAgcDriver/tests/execution/VCmpF32Denormals.cpp index 54e132b6c6..22568a2f9a 100644 --- a/core/libs/prx/libSceAgcDriver/tests/execution/VCmpF32Denormals.cpp +++ b/core/libs/prx/libSceAgcDriver/tests/execution/VCmpF32Denormals.cpp @@ -95,17 +95,9 @@ std::string Hex(std::uint32_t value) { return text; } -std::uint32_t Run(AgcDriver::VulkanDevice& device, ShaderRecompiler::ShaderFloatMode mode, std::uint32_t compare, std::uint32_t capture, std::uint32_t group) { +ShaderRecompiler::RecompileResult Compile(const AgcDriver::VulkanDevice& device, ShaderRecompiler::ShaderFloatMode mode, std::uint32_t compare, std::uint32_t capture) { Code[6] = compare; Code[7] = capture; - for (std::uint32_t lane = 0; lane < Threads; ++lane) { - const std::uint32_t pair = group * Threads + lane; - Input[lane * Inputs] = Values[pair / ValueCount]; - Input[lane * Inputs + 1u] = Values[pair % ValueCount]; - Input[lane * Inputs + 2u] = 0u; - Input[lane * Inputs + 3u] = 0u; - } - Output.fill(0xdeadbeefu); std::vector userData(8, 0u); const auto input = BufferDescriptor(Input.data(), static_cast(Input.size() * 4u)); const auto output = BufferDescriptor(Output.data(), static_cast(Output.size() * 4u)); @@ -122,8 +114,19 @@ std::uint32_t Run(AgcDriver::VulkanDevice& device, ShaderRecompiler::ShaderFloat }; request.useCache = false; request.context.floatMode = mode; - const auto result = ShaderRecompiler::Recompile(request); - device.Dispatch(result, 1, 1, 1, {}, reinterpret_cast(code.data())); + return ShaderRecompiler::Recompile(request); +} + +std::uint32_t Run(AgcDriver::VulkanDevice& device, const ShaderRecompiler::RecompileResult& program, std::uint32_t group) { + for (std::uint32_t lane = 0; lane < Threads; ++lane) { + const std::uint32_t pair = group * Threads + lane; + Input[lane * Inputs] = Values[pair / ValueCount]; + Input[lane * Inputs + 1u] = Values[pair % ValueCount]; + Input[lane * Inputs + 2u] = 0u; + Input[lane * Inputs + 3u] = 0u; + } + Output.fill(0xdeadbeefu); + device.Dispatch(program, 1, 1, 1, {}, reinterpret_cast(Code.data())); device.WaitIdle(); std::uint32_t mask = 0u; for (std::uint32_t lane = 0; lane < Threads; ++lane) { @@ -135,8 +138,9 @@ std::uint32_t Run(AgcDriver::VulkanDevice& device, ShaderRecompiler::ShaderFloat void Check(AgcDriver::VulkanDevice& device, const char* modeName, ShaderRecompiler::ShaderFloatMode mode, const Row& row, bool exec, const std::uint32_t* expected) { const std::string name = std::string(exec ? "v_cmpx_" : "v_cmp_") + row.condition + "_f32"; + const auto program = Compile(device, mode, exec ? row.cmpx : row.cmp, exec ? CaptureExec : CaptureVcc); for (std::uint32_t group = 0; group < Groups; ++group) { - const std::uint32_t actual = Run(device, mode, exec ? row.cmpx : row.cmp, exec ? CaptureExec : CaptureVcc, group); + const std::uint32_t actual = Run(device, program, group); if (actual == expected[group]) continue; const std::uint32_t diff = actual ^ expected[group]; for (std::uint32_t lane = 0; lane < Threads; ++lane) {