Merge pull request #2512 from ReinUsesLisp/comp-indexing

gl_shader_decompiler: Pessimize uniform buffer access on AMD's prorpietary driver

3 files changed, 80 insertions, 3 deletions

diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp
index 1d1581f49..65a88b06c 100644
--- a/src/video_core/renderer_opengl/gl_device.cpp
+++ b/src/video_core/renderer_opengl/gl_device.cpp
@@ -2,11 +2,14 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <array>
 #include <cstddef>
 #include <glad/glad.h>
 
 #include "common/logging/log.h"
+#include "common/scope_exit.h"
 #include "video_core/renderer_opengl/gl_device.h"
+#include "video_core/renderer_opengl/gl_resource_manager.h"
 
 namespace OpenGL {
 
@@ -24,6 +27,7 @@ Device::Device() {
     max_vertex_attributes = GetInteger<u32>(GL_MAX_VERTEX_ATTRIBS);
     max_varyings = GetInteger<u32>(GL_MAX_VARYING_VECTORS);
     has_variable_aoffi = TestVariableAoffi();
+    has_component_indexing_bug = TestComponentIndexingBug();
 }
 
 Device::Device(std::nullptr_t) {
@@ -31,6 +35,7 @@ Device::Device(std::nullptr_t) {
     max_vertex_attributes = 16;
     max_varyings = 15;
     has_variable_aoffi = true;
+    has_component_indexing_bug = false;
 }
 
 bool Device::TestVariableAoffi() {
@@ -52,4 +57,53 @@ void main() {
     return supported;
 }
 
+bool Device::TestComponentIndexingBug() {
+    constexpr char log_message[] = "Renderer_ComponentIndexingBug: {}";
+    const GLchar* COMPONENT_TEST = R"(#version 430 core
+layout (std430, binding = 0) buffer OutputBuffer {
+    uint output_value;
+};
+layout (std140, binding = 0) uniform InputBuffer {
+    uvec4 input_value[4096];
+};
+layout (location = 0) uniform uint idx;
+void main() {
+    output_value = input_value[idx >> 2][idx & 3];
+})";
+    const GLuint shader{glCreateShaderProgramv(GL_VERTEX_SHADER, 1, &COMPONENT_TEST)};
+    SCOPE_EXIT({ glDeleteProgram(shader); });
+    glUseProgram(shader);
+
+    OGLVertexArray vao;
+    vao.Create();
+    glBindVertexArray(vao.handle);
+
+    constexpr std::array<GLuint, 8> values{0, 0, 0, 0, 0x1236327, 0x985482, 0x872753, 0x2378432};
+    OGLBuffer ubo;
+    ubo.Create();
+    glNamedBufferData(ubo.handle, sizeof(values), values.data(), GL_STATIC_DRAW);
+    glBindBufferBase(GL_UNIFORM_BUFFER, 0, ubo.handle);
+
+    OGLBuffer ssbo;
+    ssbo.Create();
+    glNamedBufferStorage(ssbo.handle, sizeof(GLuint), nullptr, GL_CLIENT_STORAGE_BIT);
+
+    for (GLuint index = 4; index < 8; ++index) {
+        glInvalidateBufferData(ssbo.handle);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, ssbo.handle);
+
+        glProgramUniform1ui(shader, 0, index);
+        glDrawArrays(GL_POINTS, 0, 1);
+
+        GLuint result;
+        glGetNamedBufferSubData(ssbo.handle, 0, sizeof(result), &result);
+        if (result != values.at(index)) {
+            LOG_INFO(Render_OpenGL, log_message, true);
+            return true;
+        }
+    }
+    LOG_INFO(Render_OpenGL, log_message, false);
+    return false;
+}
+
 } // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_device.h b/src/video_core/renderer_opengl/gl_device.h
index de8490682..8c8c93760 100644
--- a/src/video_core/renderer_opengl/gl_device.h
+++ b/src/video_core/renderer_opengl/gl_device.h
@@ -30,13 +30,19 @@ public:
         return has_variable_aoffi;
     }
 
+    bool HasComponentIndexingBug() const {
+        return has_component_indexing_bug;
+    }
+
 private:
     static bool TestVariableAoffi();
+    static bool TestComponentIndexingBug();
 
     std::size_t uniform_buffer_alignment{};
     u32 max_vertex_attributes{};
     u32 max_varyings{};
     bool has_variable_aoffi{};
+    bool has_component_indexing_bug{};
 };
 
 } // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index e9f8d40db..3b61bf77f 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -577,9 +577,26 @@ private:
             if (std::holds_alternative<OperationNode>(*offset)) {
                 // Indirect access
                 const std::string final_offset = code.GenerateTemporary();
-                code.AddLine("uint {} = (ftou({}) / 4);", final_offset, Visit(offset));
-                return fmt::format("{}[{} / 4][{} % 4]", GetConstBuffer(cbuf->GetIndex()),
-                                   final_offset, final_offset);
+                code.AddLine("uint {} = ftou({}) >> 2;", final_offset, Visit(offset));
+
+                if (!device.HasComponentIndexingBug()) {
+                    return fmt::format("{}[{} >> 2][{} & 3]", GetConstBuffer(cbuf->GetIndex()),
+                                       final_offset, final_offset);
+                }
+
+                // AMD's proprietary GLSL compiler emits ill code for variable component access.
+                // To bypass this driver bug generate 4 ifs, one per each component.
+                const std::string pack = code.GenerateTemporary();
+                code.AddLine("vec4 {} = {}[{} >> 2];", pack, GetConstBuffer(cbuf->GetIndex()),
+                             final_offset);
+
+                const std::string result = code.GenerateTemporary();
+                code.AddLine("float {};", result);
+                for (u32 swizzle = 0; swizzle < 4; ++swizzle) {
+                    code.AddLine("if (({} & 3) == {}) {} = {}{};", final_offset, swizzle, result,
+                                 pack, GetSwizzle(swizzle));
+                }
+                return result;
             }
 
             UNREACHABLE_MSG("Unmanaged offset node type");