From 3d99b449d3c285a50ffb68ae077882cc2998d62f Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Wed, 3 Jun 2020 02:48:41 -0300 Subject: [PATCH] gl_rasterizer: Use NV_transform_feedback for XFB on assembly shaders NV_transform_feedback, NV_transform_feedback2 and ARB_transform_feedback3 with NV_transform_feedback interactions allows implementing transform feedbacks as dynamic state. Maxwell implements transform feedbacks as dynamic state, so using these extensions with TransformFeedbackStreamAttribsNV allows us to properly emulate transform feedbacks without having to recompile shaders when the state changes. --- src/video_core/renderer_opengl/gl_device.cpp | 3 +- .../renderer_opengl/gl_rasterizer.cpp | 90 +++++++++++++++++++ .../renderer_opengl/gl_rasterizer.h | 4 + 3 files changed, 96 insertions(+), 1 deletion(-) diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp index b772c37d9..8b424e2cb 100644 --- a/src/video_core/renderer_opengl/gl_device.cpp +++ b/src/video_core/renderer_opengl/gl_device.cpp @@ -206,7 +206,8 @@ Device::Device() has_precise_bug = TestPreciseBug(); has_fast_buffer_sub_data = is_nvidia; use_assembly_shaders = Settings::values.use_assembly_shaders && GLAD_GL_NV_gpu_program5 && - GLAD_GL_NV_compute_program5; + GLAD_GL_NV_compute_program5 && GLAD_GL_NV_transform_feedback && + GLAD_GL_NV_transform_feedback2; LOG_INFO(Render_OpenGL, "Renderer_VariableAOFFI: {}", has_variable_aoffi); LOG_INFO(Render_OpenGL, "Renderer_ComponentIndexingBug: {}", has_component_indexing_bug); diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp index 55e79aaf6..f802fd384 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp @@ -93,6 +93,34 @@ std::size_t GetConstBufferSize(const Tegra::Engines::ConstBufferInfo& buffer, return buffer.size; } +/// Translates hardware transform feedback indices +/// @param location Hardware location +/// @return Pair of ARB_transform_feedback3 token stream first and third arguments +/// @note Read https://www.khronos.org/registry/OpenGL/extensions/ARB/ARB_transform_feedback3.txt +std::pair TransformFeedbackEnum(u8 location) { + const u8 index = location / 4; + if (index >= 8 && index <= 39) { + return {GL_GENERIC_ATTRIB_NV, index - 8}; + } + if (index >= 48 && index <= 55) { + return {GL_TEXTURE_COORD_NV, index - 48}; + } + switch (index) { + case 7: + return {GL_POSITION, 0}; + case 40: + return {GL_PRIMARY_COLOR_NV, 0}; + case 41: + return {GL_SECONDARY_COLOR_NV, 0}; + case 42: + return {GL_BACK_PRIMARY_COLOR_NV, 0}; + case 43: + return {GL_BACK_SECONDARY_COLOR_NV, 0}; + } + UNIMPLEMENTED_MSG("index={}", static_cast(index)); + return {GL_POSITION, 0}; +} + void oglEnable(GLenum cap, bool state) { (state ? glEnable : glDisable)(cap); } @@ -1547,12 +1575,70 @@ void RasterizerOpenGL::SyncFramebufferSRGB() { oglEnable(GL_FRAMEBUFFER_SRGB, gpu.regs.framebuffer_srgb); } +void RasterizerOpenGL::SyncTransformFeedback() { + // TODO(Rodrigo): Inject SKIP_COMPONENTS*_NV when required. An unimplemented message will signal + // when this is required. + const auto& regs = system.GPU().Maxwell3D().regs; + + static constexpr std::size_t STRIDE = 3; + std::array attribs; + std::array streams; + + GLint* cursor = attribs.data(); + GLint* current_stream = streams.data(); + + for (std::size_t feedback = 0; feedback < Maxwell::NumTransformFeedbackBuffers; ++feedback) { + const auto& layout = regs.tfb_layouts[feedback]; + UNIMPLEMENTED_IF_MSG(layout.stride != layout.varying_count * 4, "Stride padding"); + if (layout.varying_count == 0) { + continue; + } + + *current_stream = static_cast(feedback); + if (current_stream != streams.data()) { + // When stepping one stream, push the expected token + cursor[0] = GL_NEXT_BUFFER_NV; + cursor[1] = 0; + cursor[2] = 0; + cursor += STRIDE; + } + ++current_stream; + + const auto& locations = regs.tfb_varying_locs[feedback]; + std::optional current_index; + for (u32 offset = 0; offset < layout.varying_count; ++offset) { + const u8 location = locations[offset]; + const u8 index = location / 4; + + if (current_index == index) { + // Increase number of components of the previous attachment + ++cursor[-2]; + continue; + } + current_index = index; + + std::tie(cursor[0], cursor[2]) = TransformFeedbackEnum(location); + cursor[1] = 1; + cursor += STRIDE; + } + } + + const GLsizei num_attribs = static_cast((cursor - attribs.data()) / STRIDE); + const GLsizei num_strides = static_cast(current_stream - streams.data()); + glTransformFeedbackStreamAttribsNV(num_attribs, attribs.data(), num_strides, streams.data(), + GL_INTERLEAVED_ATTRIBS); +} + void RasterizerOpenGL::BeginTransformFeedback(GLenum primitive_mode) { const auto& regs = system.GPU().Maxwell3D().regs; if (regs.tfb_enabled == 0) { return; } + if (device.UseAssemblyShaders()) { + SyncTransformFeedback(); + } + UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationControl) || regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationEval) || regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::Geometry)); @@ -1579,6 +1665,10 @@ void RasterizerOpenGL::BeginTransformFeedback(GLenum primitive_mode) { static_cast(size)); } + // We may have to call BeginTransformFeedbackNV here since they seem to call different + // implementations on Nvidia's driver (the pointer is different) but we are using + // ARB_transform_feedback3 features with NV_transform_feedback interactions and the ARB + // extension doesn't define BeginTransformFeedback (without NV) interactions. It just works. glBeginTransformFeedback(GL_POINTS); } diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h index f5dc56a0e..7abc8fdbd 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.h +++ b/src/video_core/renderer_opengl/gl_rasterizer.h @@ -202,6 +202,10 @@ private: /// Syncs the framebuffer sRGB state to match the guest state void SyncFramebufferSRGB(); + /// Syncs transform feedback state to match guest state + /// @note Only valid on assembly shaders + void SyncTransformFeedback(); + /// Begin a transform feedback void BeginTransformFeedback(GLenum primitive_mode);