Merge pull request #153 from shadps4-emu/small_additions

video_core: multiple render targets binding

src/shader_recompiler/frontend/translate/translate.cpp

@@ -321,9 +321,13 @@ void Translate(IR::Block* block, std::span<const GcnInst> inst_list, Info& info)
         case Opcode::V_MAX_F32:
             translator.V_MAX_F32(inst);
             break;
+        case Opcode::V_RSQ_F32:
+            translator.V_RSQ_F32(inst);
+            break;
         case Opcode::S_ANDN2_B64:
             translator.S_ANDN2_B64(inst);
             break;
+        case Opcode::S_NOP:
         case Opcode::S_CBRANCH_EXECZ:
         case Opcode::S_CBRANCH_SCC0:
         case Opcode::S_MOV_B64:

src/shader_recompiler/frontend/translate/translate.h

@@ -67,6 +67,7 @@ public:
     void V_FMA_F32(const GcnInst& inst);
     void V_CMP_F32(ConditionOp op, const GcnInst& inst);
     void V_MAX_F32(const GcnInst& inst);
+    void V_RSQ_F32(const GcnInst& inst);
 
     // Vector Memory
     void BUFFER_LOAD_FORMAT(u32 num_dwords, bool is_typed, const GcnInst& inst);

src/shader_recompiler/frontend/translate/vector_alu.cpp

@@ -193,4 +193,9 @@ void Translator::V_MAX_F32(const GcnInst& inst) {
     SetDst(inst.dst[0], ir.FPMax(src0, src1));
 }
 
+void Translator::V_RSQ_F32(const GcnInst& inst) {
+    const IR::F32 src0{GetSrc(inst.src[0], true)};
+    SetDst(inst.dst[0], ir.FPRecipSqrt(src0));
+}
+
 } // namespace Shader::Gcn

src/video_core/amdgpu/liverpool.h

@@ -416,6 +416,10 @@ struct Liverpool {
         BitField<20, 4, u32> output5_mask;
         BitField<24, 4, u32> output6_mask;
         BitField<28, 4, u32> output7_mask;
+
+        [[nodiscard]] u8 GetMask(int buf_id) const {
+            return (raw >> (buf_id * 4)) & 0xffu;
+        }
     };
 
     struct IndexBufferBase {
@@ -631,6 +635,10 @@ struct Liverpool {
         u32 clear_word1;
         INSERT_PADDING_WORDS(2);
 
+        operator bool() const {
+            return info.format != DataFormat::FormatInvalid;
+        }
+
         u32 Pitch() const {
             return (pitch.tile_max + 1) << 3;
         }

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -113,6 +113,11 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
         vk::DynamicState::eBlendConstants,
     };
 
+    if (instance.IsColorWriteEnableSupported()) {
+        dynamic_states.push_back(vk::DynamicState::eColorWriteEnableEXT);
+        dynamic_states.push_back(vk::DynamicState::eColorWriteMaskEXT);
+    }
+
     const vk::PipelineDynamicStateCreateInfo dynamic_info = {
         .dynamicStateCount = static_cast<u32>(dynamic_states.size()),
         .pDynamicStates = dynamic_states.data(),
@@ -144,18 +149,21 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
         },
     };
 
-    u32 shader_count = 2;
+    u32 shader_count = 1;
     std::array<vk::PipelineShaderStageCreateInfo, MaxShaderStages> shader_stages;
     shader_stages[0] = vk::PipelineShaderStageCreateInfo{
         .stage = vk::ShaderStageFlagBits::eVertex,
         .module = modules[0],
         .pName = "main",
     };
-    shader_stages[1] = vk::PipelineShaderStageCreateInfo{
+    if (modules[4]) {
-        .stage = vk::ShaderStageFlagBits::eFragment,
+        shader_stages[1] = vk::PipelineShaderStageCreateInfo{
-        .module = modules[4],
+            .stage = vk::ShaderStageFlagBits::eFragment,
-        .pName = "main",
+            .module = modules[4],
-    };
+            .pName = "main",
+        };
+        ++shader_count;
+    }
 
     const auto it = std::ranges::find(key.color_formats, vk::Format::eUndefined);
     const u32 num_color_formats = std::distance(key.color_formats.begin(), it);
@@ -177,8 +185,11 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
             .srcAlphaBlendFactor = LiverpoolToVK::BlendFactor(control.alpha_src_factor),
             .dstAlphaBlendFactor = LiverpoolToVK::BlendFactor(control.color_dst_factor),
             .alphaBlendOp = LiverpoolToVK::BlendOp(control.alpha_func),
-            .colorWriteMask = vk::ColorComponentFlagBits::eR | vk::ColorComponentFlagBits::eG |
+            .colorWriteMask =
-                              vk::ColorComponentFlagBits::eB | vk::ColorComponentFlagBits::eA,
+                instance.IsColorWriteEnableSupported()
+                    ? vk::ColorComponentFlagBits::eR | vk::ColorComponentFlagBits::eG |
+                          vk::ColorComponentFlagBits::eB | vk::ColorComponentFlagBits::eA
+                    : key.write_masks[i],
         };
     }
 

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -39,6 +39,7 @@ struct GraphicsPipelineKey {
     Liverpool::PolygonMode polygon_mode;
     Liverpool::CullMode cull_mode;
     std::array<Liverpool::BlendControl, Liverpool::NumColorBuffers> blend_controls;
+    std::array<vk::ColorComponentFlags, Liverpool::NumColorBuffers> write_masks;
 
     bool operator==(const GraphicsPipelineKey& key) const noexcept {
         return std::memcmp(this, &key, sizeof(GraphicsPipelineKey)) == 0;
@@ -66,6 +67,10 @@ public:
         return key.stage_hashes[0] == EmbeddedVsHash;
     }
 
+    [[nodiscard]] auto GetWriteMasks() const {
+        return key.write_masks;
+    }
+
 private:
     void BuildDescSetLayout();
 

src/video_core/renderer_vulkan/vk_instance.cpp

@@ -115,6 +115,7 @@ bool Instance::CreateDevice() {
         vk::PhysicalDeviceCustomBorderColorFeaturesEXT, vk::PhysicalDeviceIndexTypeUint8FeaturesEXT,
         vk::PhysicalDeviceFragmentShaderInterlockFeaturesEXT,
         vk::PhysicalDevicePipelineCreationCacheControlFeaturesEXT,
+        vk::PhysicalDeviceColorWriteEnableFeaturesEXT,
         vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR>();
     const vk::StructureChain properties_chain =
         physical_device.getProperties2<vk::PhysicalDeviceProperties2,
@@ -152,6 +153,9 @@ bool Instance::CreateDevice() {
     index_type_uint8 = add_extension(VK_KHR_INDEX_TYPE_UINT8_EXTENSION_NAME);
     add_extension(VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);
     add_extension(VK_KHR_MAINTENANCE_4_EXTENSION_NAME);
+    // The next two extensions are required to be available together in order to support write masks
+    color_write_en = add_extension(VK_EXT_COLOR_WRITE_ENABLE_EXTENSION_NAME);
+    color_write_en &= add_extension(VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
 
     const auto family_properties = physical_device.getQueueFamilyProperties();
     if (family_properties.empty()) {
@@ -191,6 +195,7 @@ bool Instance::CreateDevice() {
         vk::PhysicalDeviceFeatures2{
             .features{
                 .robustBufferAccess = features.robustBufferAccess,
+                .independentBlend = true,
                 .geometryShader = features.geometryShader,
                 .logicOp = features.logicOp,
                 .samplerAnisotropy = features.samplerAnisotropy,
@@ -216,12 +221,23 @@ bool Instance::CreateDevice() {
         vk::PhysicalDeviceIndexTypeUint8FeaturesEXT{
             .indexTypeUint8 = true,
         },
+        vk::PhysicalDeviceColorWriteEnableFeaturesEXT{
+            .colorWriteEnable = true,
+        },
+        vk::PhysicalDeviceExtendedDynamicState3FeaturesEXT{
+            .extendedDynamicState3ColorWriteMask = true,
+        },
     };
 
     if (!index_type_uint8) {
         device_chain.unlink<vk::PhysicalDeviceIndexTypeUint8FeaturesEXT>();
     }
 
+    if (!color_write_en) {
+        device_chain.unlink<vk::PhysicalDeviceColorWriteEnableFeaturesEXT>();
+        device_chain.unlink<vk::PhysicalDeviceExtendedDynamicState3FeaturesEXT>();
+    }
+
     try {
         device = physical_device.createDeviceUnique(device_chain.get());
     } catch (vk::ExtensionNotPresentError& err) {

src/video_core/renderer_vulkan/vk_instance.h

@@ -111,6 +111,11 @@ public:
         return external_memory_host;
     }
 
+    /// Returns true when VK_EXT_color_write_enable is supported
+    bool IsColorWriteEnableSupported() const {
+        return color_write_en;
+    }
+
     /// Returns the vendor ID of the physical device
     u32 GetVendorID() const {
         return properties.vendorID;
@@ -218,6 +223,7 @@ private:
     bool fragment_shader_barycentric{};
     bool shader_stencil_export{};
     bool external_memory_host{};
+    bool color_write_en{};
     u64 min_imported_host_pointer_alignment{};
     bool tooling_info{};
     bool debug_utils_supported{};

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -93,17 +93,26 @@ void PipelineCache::RefreshGraphicsKey() {
     key.stencil_ref_back = regs.stencil_ref_back;
     key.prim_type = regs.primitive_type;
     key.polygon_mode = regs.polygon_control.PolyMode();
-    key.blend_controls = regs.blend_control;
 
     const auto& db = regs.depth_buffer;
     key.depth_format = key.depth.depth_enable
                            ? LiverpoolToVK::DepthFormat(db.z_info.format, db.stencil_info.format)
                            : vk::Format::eUndefined;
-    for (u32 i = 0; i < Liverpool::NumColorBuffers; i++) {
+    // `RenderingInfo` is assumed to be initialized with a contiguous array of valid color
-        const auto& cb = regs.color_buffers[i];
+    // attachments. This might be not a case as HW color buffers can be bound in an arbitrary order.
-        key.color_formats[i] = cb.base_address
+    // We need to do some arrays compaction at this stage
-                                   ? LiverpoolToVK::SurfaceFormat(cb.info.format, cb.NumFormat())
+    int remapped_cb{};
-                                   : vk::Format::eUndefined;
+    for (auto cb = 0u; cb < Liverpool::NumColorBuffers; ++cb) {
+        auto const& col_buf = regs.color_buffers[cb];
+        if (!col_buf) {
+            continue;
+        }
+        key.color_formats[remapped_cb] =
+            LiverpoolToVK::SurfaceFormat(col_buf.info.format, col_buf.NumFormat());
+        key.blend_controls[remapped_cb] = regs.blend_control[cb];
+        key.write_masks[remapped_cb] = vk::ColorComponentFlags{regs.color_target_mask.GetMask(cb)};
+
+        ++remapped_cb;
     }
 
     for (u32 i = 0; i < MaxShaderStages; i++) {

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -39,14 +39,21 @@ void Rasterizer::Draw(bool is_indexed) {
     const GraphicsPipeline* pipeline = pipeline_cache.GetGraphicsPipeline();
     pipeline->BindResources(memory, vertex_index_buffer, texture_cache);
 
-    const auto& image_view = texture_cache.RenderTarget(regs.color_buffers[0]);
+    boost::container::static_vector<vk::RenderingAttachmentInfo, Liverpool::NumColorBuffers>
+        color_attachments{};
+    for (const auto& col_buf : regs.color_buffers) {
+        if (!col_buf) {
+            continue;
+        }
+        const auto& image_view = texture_cache.RenderTarget(col_buf);
 
-    const vk::RenderingAttachmentInfo color_info = {
+        color_attachments.push_back({
-        .imageView = *image_view.image_view,
+            .imageView = *image_view.image_view,
-        .imageLayout = vk::ImageLayout::eGeneral,
+            .imageLayout = vk::ImageLayout::eGeneral,
-        .loadOp = vk::AttachmentLoadOp::eLoad,
+            .loadOp = vk::AttachmentLoadOp::eLoad,
-        .storeOp = vk::AttachmentStoreOp::eStore,
+            .storeOp = vk::AttachmentStoreOp::eStore,
-    };
+        });
+    }
 
     // TODO: Don't restart renderpass every draw
     const auto& scissor = regs.screen_scissor;
@@ -57,11 +64,11 @@ void Rasterizer::Draw(bool is_indexed) {
                 .extent = {scissor.GetWidth(), scissor.GetHeight()},
             },
         .layerCount = 1,
-        .colorAttachmentCount = 1,
+        .colorAttachmentCount = static_cast<u32>(color_attachments.size()),
-        .pColorAttachments = &color_info,
+        .pColorAttachments = color_attachments.data(),
     };
 
-    UpdateDynamicState();
+    UpdateDynamicState(*pipeline);
 
     cmdbuf.beginRendering(rendering_info);
     cmdbuf.bindPipeline(vk::PipelineBindPoint::eGraphics, pipeline->Handle());
@@ -125,12 +132,22 @@ u32 Rasterizer::SetupIndexBuffer(bool& is_indexed) {
     return regs.num_indices;
 }
 
-void Rasterizer::UpdateDynamicState() {
+void Rasterizer::UpdateDynamicState(const GraphicsPipeline& pipeline) {
     UpdateViewportScissorState();
 
     auto& regs = liverpool->regs;
     const auto cmdbuf = scheduler.CommandBuffer();
     cmdbuf.setBlendConstants(&regs.blend_constants.red);
+
+    if (instance.IsColorWriteEnableSupported()) {
+        const auto& write_masks = pipeline.GetWriteMasks();
+        std::array<vk::Bool32, Liverpool::NumColorBuffers> write_ens{};
+        std::transform(write_masks.cbegin(), write_masks.cend(), write_ens.begin(),
+                       [](auto in) { return in ? vk::True : vk::False; });
+
+        cmdbuf.setColorWriteEnableEXT(write_ens);
+        cmdbuf.setColorWriteMaskEXT(0, write_masks);
+    }
 }
 
 void Rasterizer::UpdateViewportScissorState() {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -37,7 +37,7 @@ private:
     u32 SetupIndexBuffer(bool& is_indexed);
     void MapMemory(VAddr addr, size_t size);
 
-    void UpdateDynamicState();
+    void UpdateDynamicState(const GraphicsPipeline& pipeline);
     void UpdateViewportScissorState();
     void UpdateDepthStencilState();
 

src/video_core/texture_cache/texture_cache.cpp

@@ -15,6 +15,7 @@
 
 #define PAGE_NOACCESS PROT_NONE
 #define PAGE_READWRITE (PROT_READ | PROT_WRITE)
+#define PAGE_READONLY PROT_READ
 #else
 #include <windows.h>