diff options
29 files changed, 259 insertions, 8 deletions
diff --git a/src/shader_recompiler/CMakeLists.txt b/src/shader_recompiler/CMakeLists.txt index af8e51fe8..189e5cb17 100644 --- a/src/shader_recompiler/CMakeLists.txt +++ b/src/shader_recompiler/CMakeLists.txt @@ -224,6 +224,7 @@ add_library(shader_recompiler STATIC ir_opt/lower_fp16_to_fp32.cpp ir_opt/lower_int64_to_int32.cpp ir_opt/passes.h + ir_opt/position_pass.cpp ir_opt/rescaling_pass.cpp ir_opt/ssa_rewrite_pass.cpp ir_opt/texture_pass.cpp diff --git a/src/shader_recompiler/backend/glasm/emit_glasm.cpp b/src/shader_recompiler/backend/glasm/emit_glasm.cpp index 97a6b383b..1419207e8 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm.cpp @@ -450,6 +450,9 @@ std::string EmitGLASM(const Profile& profile, const RuntimeInfo& runtime_info, I if (program.info.uses_rescaling_uniform) { header += "PARAM scaling[1]={program.local[0..0]};"; } + if (program.info.uses_render_area) { + header += "PARAM render_area[1]={program.local[1..1]};"; + } header += "TEMP "; for (size_t index = 0; index < ctx.reg_alloc.NumUsedRegisters(); ++index) { header += fmt::format("R{},", index); diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_context_get_set.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_context_get_set.cpp index b5c08d611..ee0b0f53d 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_context_get_set.cpp +++ b/src/shader_recompiler/backend/glasm/emit_glasm_context_get_set.cpp @@ -379,6 +379,10 @@ void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst) { ctx.Add("MOV.F {}.x,scaling[0].z;", inst); } +void EmitRenderArea(EmitContext& ctx, IR::Inst& inst) { + ctx.Add("MOV.F {},render_area[0];", inst); +} + void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, ScalarU32 word_offset) { ctx.Add("MOV.U {},lmem[{}].x;", inst, word_offset); } diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h index 8b0ac3031..c08f48ed9 100644 --- a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h +++ b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h @@ -73,6 +73,7 @@ void EmitSampleId(EmitContext& ctx, IR::Inst& inst); void EmitIsHelperInvocation(EmitContext& ctx, IR::Inst& inst); void EmitYDirection(EmitContext& ctx, IR::Inst& inst); void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst); +void EmitRenderArea(EmitContext& ctx, IR::Inst& inst); void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, ScalarU32 word_offset); void EmitWriteLocal(EmitContext& ctx, ScalarU32 word_offset, ScalarU32 value); void EmitUndefU1(EmitContext& ctx, IR::Inst& inst); diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp index fad8d1e30..d7c845469 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp +++ b/src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp @@ -416,6 +416,10 @@ void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst) { ctx.AddF32("{}=scaling.z;", inst); } +void EmitRenderArea(EmitContext& ctx, IR::Inst& inst) { + ctx.AddF32x4("{}=render_area;", inst); +} + void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset) { ctx.AddU32("{}=lmem[{}];", inst, word_offset); } diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h index 639691ba6..3c8bcb7e9 100644 --- a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h +++ b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h @@ -87,6 +87,7 @@ void EmitSampleId(EmitContext& ctx, IR::Inst& inst); void EmitIsHelperInvocation(EmitContext& ctx, IR::Inst& inst); void EmitYDirection(EmitContext& ctx, IR::Inst& inst); void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst); +void EmitRenderArea(EmitContext& ctx, IR::Inst& inst); void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset); void EmitWriteLocal(EmitContext& ctx, std::string_view word_offset, std::string_view value); void EmitUndefU1(EmitContext& ctx, IR::Inst& inst); diff --git a/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp b/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp index c767a9dc3..5d01ec0cd 100644 --- a/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp +++ b/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp @@ -358,6 +358,9 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile if (info.uses_rescaling_uniform) { header += "layout(location=0) uniform vec4 scaling;"; } + if (info.uses_render_area) { + header += "layout(location=1) uniform vec4 render_area;"; + } DefineConstantBuffers(bindings); DefineConstantBufferIndirect(); DefineStorageBuffers(bindings); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv.h b/src/shader_recompiler/backend/spirv/emit_spirv.h index 7567b6fc9..937881484 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv.h +++ b/src/shader_recompiler/backend/spirv/emit_spirv.h @@ -23,8 +23,12 @@ struct RescalingLayout { alignas(16) std::array<u32, NUM_IMAGE_SCALING_WORDS> rescaling_images; u32 down_factor; }; +struct RenderAreaLayout { + std::array<f32, 4> render_area; +}; constexpr u32 RESCALING_LAYOUT_WORDS_OFFSET = offsetof(RescalingLayout, rescaling_textures); constexpr u32 RESCALING_LAYOUT_DOWN_FACTOR_OFFSET = offsetof(RescalingLayout, down_factor); +constexpr u32 RENDERAREA_LAYOUT_OFFSET = offsetof(RenderAreaLayout, render_area); [[nodiscard]] std::vector<u32> EmitSPIRV(const Profile& profile, const RuntimeInfo& runtime_info, IR::Program& program, Bindings& bindings); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp index 2c68aba39..a4751b42d 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp @@ -353,7 +353,6 @@ Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, Id vertex) { case IR::Attribute::TessellationEvaluationPointV: return ctx.OpLoad(ctx.F32[1], ctx.OpAccessChain(ctx.input_f32, ctx.tess_coord, ctx.Const(1U))); - default: throw NotImplementedException("Read attribute {}", attr); } @@ -537,6 +536,17 @@ Id EmitResolutionDownFactor(EmitContext& ctx) { } } +Id EmitRenderArea(EmitContext& ctx) { + if (ctx.profile.unified_descriptor_binding) { + const Id pointer_type{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.F32[4])}; + const Id index{ctx.Const(ctx.render_are_member_index)}; + const Id pointer{ctx.OpAccessChain(pointer_type, ctx.render_area_push_constant, index)}; + return ctx.OpLoad(ctx.F32[4], pointer); + } else { + throw NotImplementedException("SPIR-V Instruction"); + } +} + Id EmitLoadLocal(EmitContext& ctx, Id word_offset) { const Id pointer{ctx.OpAccessChain(ctx.private_u32, ctx.local_memory, word_offset)}; return ctx.OpLoad(ctx.U32[1], pointer); diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h index 984d072b4..86dd3c4f3 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h +++ b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h @@ -76,6 +76,7 @@ Id EmitSampleId(EmitContext& ctx); Id EmitIsHelperInvocation(EmitContext& ctx); Id EmitYDirection(EmitContext& ctx); Id EmitResolutionDownFactor(EmitContext& ctx); +Id EmitRenderArea(EmitContext& ctx); Id EmitLoadLocal(EmitContext& ctx, Id word_offset); void EmitWriteLocal(EmitContext& ctx, Id word_offset, Id value); Id EmitUndefU1(EmitContext& ctx); diff --git a/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp b/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp index aecc4c612..c26ad8f93 100644 --- a/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp +++ b/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp @@ -473,6 +473,7 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf DefineAttributeMemAccess(program.info); DefineGlobalMemoryFunctions(program.info); DefineRescalingInput(program.info); + DefineRenderArea(program.info); } EmitContext::~EmitContext() = default; @@ -982,6 +983,36 @@ void EmitContext::DefineRescalingInputUniformConstant() { } } +void EmitContext::DefineRenderArea(const Info& info) { + if (!info.uses_render_area) { + return; + } + + if (profile.unified_descriptor_binding) { + boost::container::static_vector<Id, 1> members{}; + u32 member_index{0}; + + members.push_back(F32[4]); + render_are_member_index = member_index++; + + const Id push_constant_struct{TypeStruct(std::span(members.data(), members.size()))}; + Decorate(push_constant_struct, spv::Decoration::Block); + Name(push_constant_struct, "RenderAreaInfo"); + + MemberDecorate(push_constant_struct, render_are_member_index, spv::Decoration::Offset, 0); + MemberName(push_constant_struct, render_are_member_index, "render_area"); + + const Id pointer_type{TypePointer(spv::StorageClass::PushConstant, push_constant_struct)}; + render_area_push_constant = + AddGlobalVariable(pointer_type, spv::StorageClass::PushConstant); + Name(render_area_push_constant, "render_area_push_constants"); + + if (profile.supported_spirv >= 0x00010400) { + interfaces.push_back(render_area_push_constant); + } + } +} + void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) { if (info.constant_buffer_descriptors.empty()) { return; diff --git a/src/shader_recompiler/backend/spirv/spirv_emit_context.h b/src/shader_recompiler/backend/spirv/spirv_emit_context.h index bc25b8b84..c86e50911 100644 --- a/src/shader_recompiler/backend/spirv/spirv_emit_context.h +++ b/src/shader_recompiler/backend/spirv/spirv_emit_context.h @@ -243,6 +243,9 @@ public: u32 texture_rescaling_index{}; u32 image_rescaling_index{}; + Id render_area_push_constant{}; + u32 render_are_member_index{}; + Id local_memory{}; Id shared_memory_u8{}; @@ -318,6 +321,7 @@ private: void DefineRescalingInput(const Info& info); void DefineRescalingInputPushConstant(); void DefineRescalingInputUniformConstant(); + void DefineRenderArea(const Info& info); void DefineInputs(const IR::Program& program); void DefineOutputs(const IR::Program& program); diff --git a/src/shader_recompiler/environment.h b/src/shader_recompiler/environment.h index 9729d48c6..980e0e54c 100644 --- a/src/shader_recompiler/environment.h +++ b/src/shader_recompiler/environment.h @@ -22,6 +22,8 @@ public: [[nodiscard]] virtual TextureType ReadTextureType(u32 raw_handle) = 0; + [[nodiscard]] virtual u32 ReadViewportTransformState() = 0; + [[nodiscard]] virtual u32 TextureBoundBuffer() const = 0; [[nodiscard]] virtual u32 LocalMemorySize() const = 0; diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp index 11086ed8c..de1ce90a3 100644 --- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp +++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp @@ -378,6 +378,14 @@ F32 IREmitter::ResolutionDownFactor() { return Inst<F32>(Opcode::ResolutionDownFactor); } +F32 IREmitter::RenderAreaWidth() { + return F32(CompositeExtract(Inst<Value>(Opcode::RenderArea), 0)); +} + +F32 IREmitter::RenderAreaHeight() { + return F32(CompositeExtract(Inst<Value>(Opcode::RenderArea), 1)); +} + U32 IREmitter::LaneId() { return Inst<U32>(Opcode::LaneId); } diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h index 25839a371..f163c18d9 100644 --- a/src/shader_recompiler/frontend/ir/ir_emitter.h +++ b/src/shader_recompiler/frontend/ir/ir_emitter.h @@ -103,6 +103,9 @@ public: [[nodiscard]] F32 ResolutionDownFactor(); + [[nodiscard]] F32 RenderAreaWidth(); + [[nodiscard]] F32 RenderAreaHeight(); + [[nodiscard]] U32 LaneId(); [[nodiscard]] U32 LoadGlobalU8(const U64& address); diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc index 86410ddfc..f3038f298 100644 --- a/src/shader_recompiler/frontend/ir/opcodes.inc +++ b/src/shader_recompiler/frontend/ir/opcodes.inc @@ -63,6 +63,7 @@ OPCODE(SampleId, U32, OPCODE(IsHelperInvocation, U1, ) OPCODE(YDirection, F32, ) OPCODE(ResolutionDownFactor, F32, ) +OPCODE(RenderArea, F32x4, ) // Undefined OPCODE(UndefU1, U1, ) diff --git a/src/shader_recompiler/frontend/maxwell/translate_program.cpp b/src/shader_recompiler/frontend/maxwell/translate_program.cpp index 77efb4f57..dafb9deee 100644 --- a/src/shader_recompiler/frontend/maxwell/translate_program.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate_program.cpp @@ -213,6 +213,8 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo Optimization::ConstantPropagationPass(program); + Optimization::PositionPass(env, program); + Optimization::GlobalMemoryToStorageBufferPass(program); Optimization::TexturePass(env, program); diff --git a/src/shader_recompiler/ir_opt/passes.h b/src/shader_recompiler/ir_opt/passes.h index 6ff8e4266..24f609d69 100644 --- a/src/shader_recompiler/ir_opt/passes.h +++ b/src/shader_recompiler/ir_opt/passes.h @@ -17,6 +17,7 @@ void LowerFp16ToFp32(IR::Program& program); void LowerInt64ToInt32(IR::Program& program); void RescalingPass(IR::Program& program); void SsaRewritePass(IR::Program& program); +void PositionPass(Environment& env, IR::Program& program); void TexturePass(Environment& env, IR::Program& program); void VerificationPass(const IR::Program& program); diff --git a/src/shader_recompiler/ir_opt/position_pass.cpp b/src/shader_recompiler/ir_opt/position_pass.cpp new file mode 100644 index 000000000..3c20b7189 --- /dev/null +++ b/src/shader_recompiler/ir_opt/position_pass.cpp @@ -0,0 +1,77 @@ +// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project +// SPDX-License-Identifier: GPL-2.0-or-later + +#include <boost/container/small_vector.hpp> + +#include "shader_recompiler/frontend/ir/basic_block.h" +#include "shader_recompiler/frontend/ir/ir_emitter.h" +#include "shader_recompiler/frontend/ir/value.h" +#include "shader_recompiler/ir_opt/passes.h" + +namespace Shader::Optimization { + +namespace { +struct PositionInst { + IR::Inst* inst; + IR::Block* block; + IR::Attribute attr; +}; +using PositionInstVector = boost::container::small_vector<PositionInst, 24>; +} // Anonymous namespace + +void PositionPass(Environment& env, IR::Program& program) { + if (env.ShaderStage() != Stage::VertexB || env.ReadViewportTransformState()) { + return; + } + + Info& info{program.info}; + info.uses_render_area = true; + + PositionInstVector to_replace; + for (IR::Block* const block : program.post_order_blocks) { + for (IR::Inst& inst : block->Instructions()) { + switch (inst.GetOpcode()) { + case IR::Opcode::SetAttribute: { + const IR::Attribute attr{inst.Arg(0).Attribute()}; + switch (attr) { + case IR::Attribute::PositionX: + case IR::Attribute::PositionY: { + to_replace.push_back(PositionInst{.inst = &inst, .block = block, .attr = attr}); + break; + } + default: + break; + } + break; + } + default: + break; + } + } + } + + for (PositionInst& position_inst : to_replace) { + IR::IREmitter ir{*position_inst.block, + IR::Block::InstructionList::s_iterator_to(*position_inst.inst)}; + const IR::F32 value(position_inst.inst->Arg(1)); + const IR::F32F64 scale(ir.Imm32(2.f)); + const IR::F32 negative_one{ir.Imm32(-1.f)}; + switch (position_inst.attr) { + case IR::Attribute::PositionX: { + position_inst.inst->SetArg( + 1, + ir.FPFma(value, ir.FPMul(ir.FPRecip(ir.RenderAreaWidth()), scale), negative_one)); + break; + } + case IR::Attribute::PositionY: { + position_inst.inst->SetArg( + 1, + ir.FPFma(value, ir.FPMul(ir.FPRecip(ir.RenderAreaHeight()), scale), negative_one)); + break; + } + default: + break; + } + } +} +} // namespace Shader::Optimization diff --git a/src/shader_recompiler/shader_info.h b/src/shader_recompiler/shader_info.h index f5690805c..e4b5ba567 100644 --- a/src/shader_recompiler/shader_info.h +++ b/src/shader_recompiler/shader_info.h @@ -175,6 +175,7 @@ struct Info { bool uses_shadow_lod{}; bool uses_rescaling_uniform{}; bool uses_cbuf_indirect{}; + bool uses_render_area{}; IR::Type used_constant_buffer_types{}; IR::Type used_storage_buffer_types{}; diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp index 67eae369d..c297f7121 100644 --- a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp +++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp @@ -502,6 +502,17 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) { float_image_scaling_mask, down_factor, 0.0f); } } + if (info.uses_render_area) { + const auto render_area_width(static_cast<GLfloat>(regs.render_area.width)); + const auto render_area_height(static_cast<GLfloat>(regs.render_area.height)); + if (use_assembly) { + glProgramLocalParameter4fARB(AssemblyStage(stage), 1, render_area_width, + render_area_height, 0.0f, 0.0f); + } else { + glProgramUniform4f(source_programs[stage].handle, 1, render_area_width, + render_area_height, 0.0f, 0.0f); + } + } }}; if constexpr (Spec::enabled_stages[0]) { prepare_stage(0); diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp index a0d048b0b..f018333af 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp @@ -622,6 +622,16 @@ void RasterizerOpenGL::SyncViewport() { } flags[Dirty::Viewport0 + index] = false; + if (!regs.viewport_transform_enabled) { + const auto x = static_cast<GLfloat>(regs.render_area.x); + const auto y = static_cast<GLfloat>(regs.render_area.y); + const auto width = static_cast<GLfloat>(regs.render_area.width); + const auto height = static_cast<GLfloat>(regs.render_area.height); + glViewportIndexedf(static_cast<GLuint>(index), x, y, width != 0.0f ? width : 1.0f, + height != 0.0f ? height : 1.0f); + continue; + } + const auto& src = regs.viewport_transform[index]; GLfloat x = conv(src.translate_x - src.scale_x); GLfloat y = conv(src.translate_y - src.scale_y); diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp index 1ad56d9e7..7e76f679f 100644 --- a/src/video_core/renderer_opengl/gl_shader_cache.cpp +++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp @@ -49,7 +49,7 @@ using VideoCommon::LoadPipelines; using VideoCommon::SerializePipeline; using Context = ShaderContext::Context; -constexpr u32 CACHE_VERSION = 5; +constexpr u32 CACHE_VERSION = 7; template <typename Container> auto MakeSpan(Container& container) { diff --git a/src/video_core/renderer_vulkan/pipeline_helper.h b/src/video_core/renderer_vulkan/pipeline_helper.h index b24f3424a..b7843e995 100644 --- a/src/video_core/renderer_vulkan/pipeline_helper.h +++ b/src/video_core/renderer_vulkan/pipeline_helper.h @@ -68,13 +68,15 @@ public: } vk::PipelineLayout CreatePipelineLayout(VkDescriptorSetLayout descriptor_set_layout) const { + using Shader::Backend::SPIRV::RenderAreaLayout; using Shader::Backend::SPIRV::RescalingLayout; const u32 size_offset = is_compute ? sizeof(RescalingLayout::down_factor) : 0u; const VkPushConstantRange range{ .stageFlags = static_cast<VkShaderStageFlags>( is_compute ? VK_SHADER_STAGE_COMPUTE_BIT : VK_SHADER_STAGE_ALL_GRAPHICS), .offset = 0, - .size = static_cast<u32>(sizeof(RescalingLayout)) - size_offset, + .size = static_cast<u32>(sizeof(RescalingLayout)) - size_offset + + static_cast<u32>(sizeof(RenderAreaLayout)), }; return device->GetLogical().CreatePipelineLayout({ .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO, @@ -167,6 +169,12 @@ private: u32 image_bit{1u}; }; +class RenderAreaPushConstant { +public: + bool uses_render_area{}; + std::array<f32, 4> words{}; +}; + inline void PushImageDescriptors(TextureCache& texture_cache, UpdateDescriptorQueue& update_descriptor_queue, const Shader::Info& info, RescalingPushConstant& rescaling, diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp index 5aca8f038..7eb623a7c 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp @@ -31,6 +31,7 @@ namespace { using boost::container::small_vector; using boost::container::static_vector; using Shader::ImageBufferDescriptor; +using Shader::Backend::SPIRV::RENDERAREA_LAYOUT_OFFSET; using Shader::Backend::SPIRV::RESCALING_LAYOUT_DOWN_FACTOR_OFFSET; using Shader::Backend::SPIRV::RESCALING_LAYOUT_WORDS_OFFSET; using Tegra::Texture::TexturePair; @@ -433,12 +434,19 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) { update_descriptor_queue.Acquire(); RescalingPushConstant rescaling; + RenderAreaPushConstant render_area; const VkSampler* samplers_it{samplers.data()}; const VideoCommon::ImageViewInOut* views_it{views.data()}; const auto prepare_stage{[&](size_t stage) LAMBDA_FORCEINLINE { buffer_cache.BindHostStageBuffers(stage); PushImageDescriptors(texture_cache, update_descriptor_queue, stage_infos[stage], rescaling, samplers_it, views_it); + const auto& info{stage_infos[0]}; + if (info.uses_render_area) { + render_area.uses_render_area = true; + render_area.words = {static_cast<float>(regs.render_area.width), + static_cast<float>(regs.render_area.height)}; + } }}; if constexpr (Spec::enabled_stages[0]) { prepare_stage(0); @@ -455,10 +463,11 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) { if constexpr (Spec::enabled_stages[4]) { prepare_stage(4); } - ConfigureDraw(rescaling); + ConfigureDraw(rescaling, render_area); } -void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling) { +void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling, + const RenderAreaPushConstant& render_are) { texture_cache.UpdateRenderTargets(false); scheduler.RequestRenderpass(texture_cache.GetFramebuffer()); @@ -474,7 +483,9 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling) { const bool bind_pipeline{scheduler.UpdateGraphicsPipeline(this)}; const void* const descriptor_data{update_descriptor_queue.UpdateData()}; scheduler.Record([this, descriptor_data, bind_pipeline, rescaling_data = rescaling.Data(), - is_rescaling, update_rescaling](vk::CommandBuffer cmdbuf) { + is_rescaling, update_rescaling, + uses_render_area = render_are.uses_render_area, + render_area_data = render_are.words](vk::CommandBuffer cmdbuf) { if (bind_pipeline) { cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline); } @@ -483,11 +494,16 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling) { rescaling_data.data()); if (update_rescaling) { const f32 config_down_factor{Settings::values.resolution_info.down_factor}; - const f32 scale_down_factor{is_rescaling ? config_down_factor : 1.0f}; + const f32 scale_down_factor{is_rescaling ? config_down_factor : 2.0f}; cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_ALL_GRAPHICS, RESCALING_LAYOUT_DOWN_FACTOR_OFFSET, sizeof(scale_down_factor), &scale_down_factor); } + if (uses_render_area) { + cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_ALL_GRAPHICS, + RENDERAREA_LAYOUT_OFFSET, sizeof(render_area_data), + &render_area_data); + } if (!descriptor_set_layout) { return; } diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h index e8949a9ab..8842f62d7 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h @@ -62,6 +62,7 @@ class Device; class PipelineStatistics; class RenderPassCache; class RescalingPushConstant; +class RenderAreaPushConstant; class Scheduler; class UpdateDescriptorQueue; @@ -113,7 +114,8 @@ private: template <typename Spec> void ConfigureImpl(bool is_indexed); - void ConfigureDraw(const RescalingPushConstant& rescaling); + void ConfigureDraw(const RescalingPushConstant& rescaling, + const RenderAreaPushConstant& render_are); void MakePipeline(VkRenderPass render_pass); diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 7e40c2df1..bf9f825f4 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -682,6 +682,22 @@ void RasterizerVulkan::UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& reg if (!state_tracker.TouchViewports()) { return; } + if (!regs.viewport_transform_enabled) { + const auto x = static_cast<float>(regs.render_area.x); + const auto y = static_cast<float>(regs.render_area.y); + const auto width = static_cast<float>(regs.render_area.width); + const auto height = static_cast<float>(regs.render_area.height); + VkViewport viewport{ + .x = x, + .y = y, + .width = width != 0.0f ? width : 1.0f, + .height = height != 0.0f ? height : 1.0f, + .minDepth = 0.0f, + .maxDepth = 1.0f, + }; + scheduler.Record([viewport](vk::CommandBuffer cmdbuf) { cmdbuf.SetViewport(0, viewport); }); + return; + } const bool is_rescaling{texture_cache.IsRescaling()}; const float scale = is_rescaling ? Settings::values.resolution_info.up_factor : 1.0f; const std::array viewports{ diff --git a/src/video_core/shader_environment.cpp b/src/video_core/shader_environment.cpp index 808d88eec..3ead06dd6 100644 --- a/src/video_core/shader_environment.cpp +++ b/src/video_core/shader_environment.cpp @@ -191,6 +191,8 @@ void GenericEnvironment::Serialize(std::ofstream& file) const { .write(reinterpret_cast<const char*>(&start_address), sizeof(start_address)) .write(reinterpret_cast<const char*>(&cached_lowest), sizeof(cached_lowest)) .write(reinterpret_cast<const char*>(&cached_highest), sizeof(cached_highest)) + .write(reinterpret_cast<const char*>(&viewport_transform_state), + sizeof(viewport_transform_state)) .write(reinterpret_cast<const char*>(&stage), sizeof(stage)) .write(reinterpret_cast<const char*>(code.data()), code_size); for (const auto& [key, type] : texture_types) { @@ -311,6 +313,12 @@ Shader::TextureType GraphicsEnvironment::ReadTextureType(u32 handle) { return ReadTextureTypeImpl(regs.tic.Address(), regs.tic.limit, via_header_index, handle); } +u32 GraphicsEnvironment::ReadViewportTransformState() { + const auto& regs{maxwell3d->regs}; + viewport_transform_state = regs.viewport_transform_enabled; + return viewport_transform_state; +} + ComputeEnvironment::ComputeEnvironment(Tegra::Engines::KeplerCompute& kepler_compute_, Tegra::MemoryManager& gpu_memory_, GPUVAddr program_base_, u32 start_address_) @@ -342,6 +350,10 @@ Shader::TextureType ComputeEnvironment::ReadTextureType(u32 handle) { return ReadTextureTypeImpl(regs.tic.Address(), regs.tic.limit, qmd.linked_tsc != 0, handle); } +u32 ComputeEnvironment::ReadViewportTransformState() { + return viewport_transform_state; +} + void FileEnvironment::Deserialize(std::ifstream& file) { u64 code_size{}; u64 num_texture_types{}; @@ -354,6 +366,7 @@ void FileEnvironment::Deserialize(std::ifstream& file) { .read(reinterpret_cast<char*>(&start_address), sizeof(start_address)) .read(reinterpret_cast<char*>(&read_lowest), sizeof(read_lowest)) .read(reinterpret_cast<char*>(&read_highest), sizeof(read_highest)) + .read(reinterpret_cast<char*>(&viewport_transform_state), sizeof(viewport_transform_state)) .read(reinterpret_cast<char*>(&stage), sizeof(stage)); code = std::make_unique<u64[]>(Common::DivCeil(code_size, sizeof(u64))); file.read(reinterpret_cast<char*>(code.get()), code_size); @@ -411,6 +424,10 @@ Shader::TextureType FileEnvironment::ReadTextureType(u32 handle) { return it->second; } +u32 FileEnvironment::ReadViewportTransformState() { + return viewport_transform_state; +} + u32 FileEnvironment::LocalMemorySize() const { return local_memory_size; } diff --git a/src/video_core/shader_environment.h b/src/video_core/shader_environment.h index 5a145f33a..a0659fd7c 100644 --- a/src/video_core/shader_environment.h +++ b/src/video_core/shader_environment.h @@ -85,6 +85,8 @@ protected: u32 cached_highest = 0; u32 initial_offset = 0; + u32 viewport_transform_state = 1; + bool has_unbound_instructions = false; }; @@ -102,6 +104,8 @@ public: Shader::TextureType ReadTextureType(u32 handle) override; + u32 ReadViewportTransformState() override; + private: Tegra::Engines::Maxwell3D* maxwell3d{}; size_t stage_index{}; @@ -120,6 +124,8 @@ public: Shader::TextureType ReadTextureType(u32 handle) override; + u32 ReadViewportTransformState() override; + private: Tegra::Engines::KeplerCompute* kepler_compute{}; }; @@ -143,6 +149,8 @@ public: [[nodiscard]] Shader::TextureType ReadTextureType(u32 handle) override; + [[nodiscard]] u32 ReadViewportTransformState() override; + [[nodiscard]] u32 LocalMemorySize() const override; [[nodiscard]] u32 SharedMemorySize() const override; @@ -164,6 +172,7 @@ private: u32 read_lowest{}; u32 read_highest{}; u32 initial_offset{}; + u32 viewport_transform_state = 1; }; void SerializePipeline(std::span<const char> key, std::span<const GenericEnvironment* const> envs, |