5 files changed, 62 insertions, 1 deletions

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index 2442c3c29..e61d9af80 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -33,6 +33,7 @@ set(SHADER_FILES
     opengl_fidelityfx_fsr.frag
     opengl_fidelityfx_fsr_easu.frag
     opengl_fidelityfx_fsr_rcas.frag
+    opengl_lmem_warmup.comp
     opengl_present.frag
     opengl_present.vert
     opengl_present_scaleforce.frag
diff --git a/src/video_core/host_shaders/opengl_lmem_warmup.comp b/src/video_core/host_shaders/opengl_lmem_warmup.comp
new file mode 100644
index 000000000..518268477
--- /dev/null
+++ b/src/video_core/host_shaders/opengl_lmem_warmup.comp
@@ -0,0 +1,47 @@
+// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+// This shader is a workaround for a quirk in NVIDIA OpenGL drivers
+// Shaders using local memory see a great performance benefit if a shader that was dispatched
+// before it had more local memory allocated.
+// This shader allocates the maximum local memory allowed on NVIDIA drivers to ensure that
+// subsequent shaders see the performance boost.
+
+// NOTE: This shader does no actual meaningful work and returns immediately,
+// it is simply a means to have the driver expect a shader using lots of local memory.
+
+#version 450
+
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+
+layout(location = 0) uniform uint uniform_data;
+
+layout(binding = 0, rgba8) uniform writeonly restrict image2DArray dest_image;
+
+#define MAX_LMEM_SIZE 4080 // Size chosen to avoid errors in Nvidia's GLSL compiler
+#define NUM_LMEM_CONSTANTS 1
+#define ARRAY_SIZE MAX_LMEM_SIZE - NUM_LMEM_CONSTANTS
+
+uint lmem_0[ARRAY_SIZE];
+const uvec4 constant_values[NUM_LMEM_CONSTANTS] = uvec4[](uvec4(0));
+
+void main() {
+    const uint global_id = gl_GlobalInvocationID.x;
+    if (global_id <= 128) {
+        // Since the shader is called with a dispatch of 1x1x1
+        // This should always be the case, and this shader will not actually execute
+        return;
+    }
+    for (uint t = 0; t < uniform_data; t++) {
+        const uint offset = (t * uniform_data);
+        lmem_0[offset] = t;
+    }
+    const uint offset = (gl_GlobalInvocationID.y * uniform_data + gl_GlobalInvocationID.x);
+    const uint value = lmem_0[offset];
+    const uint const_value = constant_values[offset / 4][offset % 4];
+    const uvec4 color = uvec4(value + const_value);
+
+    // A "side-effect" is needed so the variables don't get optimized out,
+    // but this should never execute so there should be no clobbering of previously bound state.
+    imageStore(dest_image, ivec3(gl_GlobalInvocationID), color);
+}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index fc711c44a..d03288516 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -222,6 +222,7 @@ void RasterizerOpenGL::PrepareDraw(bool is_indexed, Func&& draw_func) {
     gpu.TickWork();
 
     std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
+    program_manager.LocalMemoryWarmup();
     pipeline->SetEngine(maxwell3d, gpu_memory);
     pipeline->Configure(is_indexed);
 
@@ -371,6 +372,7 @@ void RasterizerOpenGL::DispatchCompute() {
     if (!pipeline) {
         return;
     }
+    program_manager.LocalMemoryWarmup();
     pipeline->SetEngine(kepler_compute, gpu_memory);
     pipeline->Configure();
     const auto& qmd{kepler_compute->launch_description};
diff --git a/src/video_core/renderer_opengl/gl_shader_manager.cpp b/src/video_core/renderer_opengl/gl_shader_manager.cpp
index 98841ae65..2f6ba6823 100644
--- a/src/video_core/renderer_opengl/gl_shader_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_manager.cpp
@@ -3,7 +3,9 @@
 
 #include <glad/glad.h>
 
+#include "video_core/host_shaders/opengl_lmem_warmup_comp.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
+#include "video_core/renderer_opengl/gl_shader_util.h"
 
 namespace OpenGL {
 
@@ -12,7 +14,8 @@ static constexpr std::array ASSEMBLY_PROGRAM_ENUMS{
     GL_GEOMETRY_PROGRAM_NV, GL_FRAGMENT_PROGRAM_NV,
 };
 
-ProgramManager::ProgramManager(const Device& device) {
+ProgramManager::ProgramManager(const Device& device)
+    : lmem_warmup_program(CreateProgram(HostShaders::OPENGL_LMEM_WARMUP_COMP, GL_COMPUTE_SHADER)) {
     glCreateProgramPipelines(1, &pipeline.handle);
     if (device.UseAssemblyShaders()) {
         glEnable(GL_COMPUTE_PROGRAM_NV);
@@ -98,6 +101,11 @@ void ProgramManager::BindAssemblyPrograms(std::span<const OGLAssemblyProgram, NU
 
 void ProgramManager::RestoreGuestCompute() {}
 
+void ProgramManager::LocalMemoryWarmup() {
+    BindComputeProgram(lmem_warmup_program.handle);
+    glDispatchCompute(1, 1, 1);
+}
+
 void ProgramManager::BindPipeline() {
     if (!is_pipeline_bound) {
         is_pipeline_bound = true;
diff --git a/src/video_core/renderer_opengl/gl_shader_manager.h b/src/video_core/renderer_opengl/gl_shader_manager.h
index 07ffab77f..852d8c88e 100644
--- a/src/video_core/renderer_opengl/gl_shader_manager.h
+++ b/src/video_core/renderer_opengl/gl_shader_manager.h
@@ -30,6 +30,8 @@ public:
 
     void RestoreGuestCompute();
 
+    void LocalMemoryWarmup();
+
 private:
     void BindPipeline();
 
@@ -44,6 +46,7 @@ private:
     u32 current_stage_mask = 0;
     std::array<GLuint, NUM_STAGES> current_programs{};
     GLuint current_assembly_compute_program = 0;
+    OGLProgram lmem_warmup_program;
 };
 
 } // namespace OpenGL