GL: Respect minimum UBO alignment requirement (#794)

* GL: Respect minimum UBO alignment requirement

* Use glBindBufferRange for UBOs

* Fix fragment shadergen UBO bindings

* Nit

include/renderer_gl/gl_driver.hpp

@@ -1,4 +1,5 @@
 #pragma once
+#include "opengl.hpp"
 
 // Information about our OpenGL/OpenGL ES driver that we should keep track of
 // Stuff like whether specific extensions are supported, and potentially things like OpenGL context information
@@ -8,6 +9,9 @@ namespace OpenGL {
 		bool supportsExtFbFetch = false;
 		bool supportsArmFbFetch = false;
 
+		// Minimum alignment for UBO offsets. Fetched by the OpenGL renderer using glGetIntegerV.
+		GLuint uboAlignment = 16;
+
 		bool supportFbFetch() const { return supportsExtFbFetch || supportsArmFbFetch; }
 	};
 }  // namespace OpenGL

include/renderer_gl/renderer_gl.hpp

@@ -101,6 +101,9 @@ class RendererGL final : public Renderer {
 	std::unique_ptr<StreamBuffer> hwVertexBuffer;
 	std::unique_ptr<StreamBuffer> hwIndexBuffer;
 
+	// Current offset for our hw shader uniform UBO
+	u32 hwShaderUniformUBOOffset = 0;
+
 	// Cache of fixed attribute values so that we don't do any duplicate updates
 	std::array<std::array<float, 4>, 16> fixedAttrValues;
 

src/core/renderer_gl/renderer_gl.cpp

@@ -2,6 +2,7 @@
 
 #include <stb_image_write.h>
 
+#include <algorithm>
 #include <bit>
 #include <cmrc/cmrc.hpp>
 
@@ -179,6 +180,10 @@ void RendererGL::initGraphicsContextInternal() {
 	driverInfo.supportsExtFbFetch = (GLAD_GL_EXT_shader_framebuffer_fetch != 0);
 	driverInfo.supportsArmFbFetch = (GLAD_GL_ARM_shader_framebuffer_fetch != 0);
 
+	// UBOs have an alignment requirement we have to respect
+	glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, reinterpret_cast<GLint*>(&driverInfo.uboAlignment));
+	driverInfo.uboAlignment = std::max<GLuint>(driverInfo.uboAlignment, 16);
+
 	// Initialize the default vertex shader used with shadergen
 	std::string defaultShadergenVSSource = fragShaderGen.getDefaultVertexShader();
 	defaultShadergenVs.create({defaultShadergenVSSource.c_str(), defaultShadergenVSSource.size()}, OpenGL::Vertex);
@@ -930,11 +935,6 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 			glUniformBlockBinding(program.handle(), vertexUBOIndex, vsUBOBlockBinding);
 		}
 	}
-	glBindBufferBase(GL_UNIFORM_BUFFER, fsUBOBlockBinding, shadergenFragmentUBO->GetGLBufferId());
-
-	if (usingAcceleratedShader) {
-		glBindBufferBase(GL_UNIFORM_BUFFER, vsUBOBlockBinding, hwShaderUniformUBO->GetGLBufferId());
-	}
 
 	// Upload uniform data to our shader's UBO
 	PICA::FragmentUniforms uniforms;
@@ -1020,10 +1020,21 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 
 	// Upload fragment uniforms to UBO
 	shadergenFragmentUBO->Bind();
-	auto mapRes = shadergenFragmentUBO->Map(4, sizeof(PICA::FragmentUniforms));
-	std::memcpy(mapRes.pointer, &uniforms, sizeof(PICA::FragmentUniforms));
+	auto uboRes = shadergenFragmentUBO->Map(driverInfo.uboAlignment, sizeof(PICA::FragmentUniforms));
+	std::memcpy(uboRes.pointer, &uniforms, sizeof(PICA::FragmentUniforms));
 	shadergenFragmentUBO->Unmap(sizeof(PICA::FragmentUniforms));
 
+	// Bind our UBOs
+	glBindBufferRange(
+		GL_UNIFORM_BUFFER, fsUBOBlockBinding, shadergenFragmentUBO->GetGLBufferId(), uboRes.buffer_offset, sizeof(PICA::FragmentUniforms)
+	);
+
+	if (usingAcceleratedShader) {
+		glBindBufferRange(
+			GL_UNIFORM_BUFFER, vsUBOBlockBinding, hwShaderUniformUBO->GetGLBufferId(), hwShaderUniformUBOOffset, PICAShader::totalUniformSize()
+		);
+	}
+
 	return program;
 }
 
@@ -1074,11 +1085,14 @@ bool RendererGL::prepareForDraw(ShaderUnit& shaderUnit, PICA::DrawAcceleration*
 			generatedVertexShader = &(*shader);
 			hwShaderUniformUBO->Bind();
 
+			// Upload shader uniforms to our UBO
 			if (shaderUnit.vs.uniformsDirty) {
 				shaderUnit.vs.uniformsDirty = false;
-				auto mapRes = hwShaderUniformUBO->Map(4, PICAShader::totalUniformSize());
-				std::memcpy(mapRes.pointer, shaderUnit.vs.getUniformPointer(), PICAShader::totalUniformSize());
+				auto uboRes = hwShaderUniformUBO->Map(driverInfo.uboAlignment, PICAShader::totalUniformSize());
+				std::memcpy(uboRes.pointer, shaderUnit.vs.getUniformPointer(), PICAShader::totalUniformSize());
 				hwShaderUniformUBO->Unmap(PICAShader::totalUniformSize());
+
+				hwShaderUniformUBOOffset = uboRes.buffer_offset;
 			}
 
 			performIndexedRender = accel->indexed;