Avoid recompiling identical SPIR-V code
We were creating SpirvShader objects for every shader stage of the
pipeline, each with their own unique serial ID. This caused us to
compile the same SPIR-V code over an over again when multiple pipelines
are created from the same shader module(s).
This change essentially moves the serial ID to the shader module. Things
that still require us to recompile code from the same shader module are
the entry point specification, and specialization constants. The former
is taken into account by using a 64-bit ID consisting of the module ID
and entry point ID. For the latter we assume any use of specialization
constants will result in a unique SPIR-V binary. This is conservative
and may still lead to unnecessary recompiles.
This change also minimizes the state passed to SpirvShader, to prevent
specialization on state not taken into account by the routine caches.
Bug: b/135609394
Tests: dEQP-VK.pipeline.render_to_image.core.*.huge.*
Change-Id: I204e812265067462f8019af9f6b7b3067ef5dc7f
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33109
Presubmit-Ready: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
diff --git a/src/Vulkan/VkShaderModule.cpp b/src/Vulkan/VkShaderModule.cpp
index b48c909..ce5e831 100644
--- a/src/Vulkan/VkShaderModule.cpp
+++ b/src/Vulkan/VkShaderModule.cpp
@@ -19,7 +19,10 @@
namespace vk
{
-ShaderModule::ShaderModule(const VkShaderModuleCreateInfo* pCreateInfo, void* mem) : code(reinterpret_cast<uint32_t*>(mem))
+std::atomic<uint32_t> ShaderModule::serialCounter(1); // Start at 1, 0 is invalid shader.
+
+ShaderModule::ShaderModule(const VkShaderModuleCreateInfo* pCreateInfo, void* mem)
+ : serialID(nextSerialID()), code(reinterpret_cast<uint32_t*>(mem))
{
memcpy(code, pCreateInfo->pCode, pCreateInfo->codeSize);
wordCount = static_cast<uint32_t>(pCreateInfo->codeSize / sizeof(uint32_t));
