Base DMA implementation

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -164,7 +164,7 @@ using BufferAlias = EmitContext::BufferAlias;
 
 Id EmitReadConst(EmitContext& ctx, IR::Inst* inst) {
     const u32 flatbuf_off_dw = inst->Flags<u32>();
-    const auto& srt_flatbuf = ctx.buffers.back();
+    const auto& srt_flatbuf = ctx.buffers[ctx.flatbuf_index];
     ASSERT(srt_flatbuf.binding >= 0 && flatbuf_off_dw > 0 &&
            srt_flatbuf.buffer_type == BufferType::ReadConstUbo);
     const auto [id, pointer_type] = srt_flatbuf[BufferAlias::U32];

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -195,7 +195,8 @@ EmitContext::SpirvAttribute EmitContext::GetAttributeInfo(AmdGpu::NumberFormat f
 }
 
 Id EmitContext::GetBufferSize(const u32 sharp_idx) {
-    const auto& srt_flatbuf = buffers.back();
+    // Can this be done with memory access? Like we do now with ReadConst
+    const auto& srt_flatbuf = buffers[flatbuf_index];
     ASSERT(srt_flatbuf.buffer_type == BufferType::ReadConstUbo);
     const auto [id, pointer_type] = srt_flatbuf[BufferAlias::U32];
 
@@ -693,6 +694,12 @@ EmitContext::BufferSpv EmitContext::DefineBuffer(bool is_storage, bool is_writte
     case Shader::BufferType::ReadConstUbo:
         Name(id, "srt_flatbuf_ubo");
         break;
+    case Shader::BufferType::BdaPagetable:
+        Name(id, "bda_pagetable");
+        break;
+    case Shader::BufferType::FaultReadback:
+        Name(id, "fault_readback");
+        break;
     case Shader::BufferType::SharedMemory:
         Name(id, "ssbo_shmem");
         break;
@@ -718,6 +725,15 @@ void EmitContext::DefineBuffers() {
         const auto buf_sharp = desc.GetSharp(info);
         const bool is_storage = desc.IsStorage(buf_sharp, profile);
 
+        // Set indexes for special buffers.
+        if (desc.buffer_type == BufferType::ReadConstUbo) {
+            flatbuf_index = buffers.size();
+        } else if (desc.buffer_type == BufferType::BdaPagetable) {
+            bda_pagetable_index = buffers.size();
+        } else if (desc.buffer_type == BufferType::FaultReadback) {
+            fault_readback_index = buffers.size();
+        }
+
         // Define aliases depending on the shader usage.
         auto& spv_buffer = buffers.emplace_back(binding.buffer++, desc.buffer_type);
         if (True(desc.used_types & IR::Type::U32)) {

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -269,6 +269,10 @@ public:
     boost::container::small_vector<TextureDefinition, 8> images;
     boost::container::small_vector<Id, 4> samplers;
 
+    size_t flatbuf_index{};
+    size_t bda_pagetable_index{};
+    size_t fault_readback_index{};
+
     Id sampler_type{};
     Id sampler_pointer_type{};
 

src/shader_recompiler/frontend/translate/scalar_memory.cpp

@@ -39,21 +39,22 @@ void Translator::EmitScalarMemory(const GcnInst& inst) {
 
 void Translator::S_LOAD_DWORD(int num_dwords, const GcnInst& inst) {
     const auto& smrd = inst.control.smrd;
-    const u32 dword_offset = [&] -> u32 {
+    const IR::ScalarReg sbase{inst.src[0].code * 2};
+    const IR::U32 dword_offset = [&] -> IR::U32 {
         if (smrd.imm) {
-            return smrd.offset;
+            return ir.Imm32(smrd.offset);
         }
         if (smrd.offset == SQ_SRC_LITERAL) {
-            return inst.src[1].code;
+            return ir.Imm32(inst.src[1].code);
         }
-        UNREACHABLE();
+        return ir.ShiftRightLogical(ir.GetScalarReg(IR::ScalarReg(smrd.offset)), ir.Imm32(2));
     }();
-    const IR::ScalarReg sbase{inst.src[0].code * 2};
     const IR::Value base =
         ir.CompositeConstruct(ir.GetScalarReg(sbase), ir.GetScalarReg(sbase + 1));
     IR::ScalarReg dst_reg{inst.dst[0].code};
     for (u32 i = 0; i < num_dwords; i++) {
-        ir.SetScalarReg(dst_reg++, ir.ReadConst(base, ir.Imm32(dword_offset + i)));
+        IR::U32 index = ir.IAdd(dword_offset, ir.Imm32(i));
+        ir.SetScalarReg(dst_reg++, ir.ReadConst(base, index));
     }
 }
 

src/shader_recompiler/info.h

@@ -42,6 +42,8 @@ constexpr u32 NUM_TEXTURE_TYPES = 7;
 enum class BufferType : u32 {
     Guest,
     ReadConstUbo,
+    BdaPagetable,
+    FaultReadback,
     GdsBuffer,
     SharedMemory,
 };
@@ -209,6 +211,7 @@ struct Info {
     bool stores_tess_level_inner{};
     bool translation_failed{};
     bool has_readconst{};
+    bool uses_dma{};
     u8 mrt_mask{0u};
     bool has_fetch_shader{false};
     u32 fetch_shader_sgpr_base{0u};

src/shader_recompiler/ir/passes/shader_info_collection_pass.cpp

@@ -87,6 +87,21 @@ void Visit(Info& info, const IR::Inst& inst) {
             });
             info.has_readconst = true;
         }
+        if (!info.uses_dma) {
+            // For now, we only need U32, but we may need
+            // to add more types in the future for other porposes.
+            info.buffers.push_back({
+                .used_types = IR::Type::U32,
+                .inline_cbuf = AmdGpu::Buffer::Null(),
+                .buffer_type = BufferType::BdaPagetable,
+            });
+            info.buffers.push_back({
+                .used_types = IR::Type::U32,
+                .inline_cbuf = AmdGpu::Buffer::Null(),
+                .buffer_type = BufferType::FaultReadback,
+            });
+            info.uses_dma = true;
+        }
         break;
     case IR::Opcode::PackUfloat10_11_11:
         info.uses_pack_10_11_11 = true;

src/video_core/buffer_cache/buffer_cache.cpp

@@ -379,10 +379,7 @@ void BufferCache::ImportMemory(u64 start, u64 end) {
         }
         WriteDataBuffer(bda_pagetable_buffer, range_start * sizeof(vk::DeviceAddress), bda_addrs.data(),
                         bda_addrs.size() * sizeof(vk::DeviceAddress));
-        {
-            std::scoped_lock lk{mutex};
-            imported_buffers.emplace_back(std::move(buffer));
-        }
+        imported_buffers.emplace_back(std::move(buffer));
         // Mark the pages as covered
         imported_regions += range;
     }
@@ -850,7 +847,7 @@ void BufferCache::WriteDataBuffer(Buffer& buffer, VAddr address, const void* val
     vk::Buffer src_buffer = staging_buffer.Handle();
     if (num_bytes < StagingBufferSize) {
         const auto [staging, offset] = staging_buffer.Map(num_bytes);
-        std::memcpy(staging + offset, value, num_bytes);
+        std::memcpy(staging, value, num_bytes);
         copy.srcOffset = offset;
         staging_buffer.Commit();
     } else {

src/video_core/buffer_cache/buffer_cache.h

@@ -79,13 +79,13 @@ public:
     }
 
     /// Retrieves the device local DBA page table buffer.
-    [[nodiscard]] Buffer& GetBdaPageTableBuffer() noexcept {
-        return bda_pagetable_buffer;
+    [[nodiscard]] Buffer* GetBdaPageTableBuffer() noexcept {
+        return &bda_pagetable_buffer;
     }
 
     /// Retrieves the fault readback buffer.
-    [[nodiscard]] Buffer& GetFaultReadbackBuffer() noexcept {
-        return fault_readback_buffer;
+    [[nodiscard]] Buffer* GetFaultReadbackBuffer() noexcept {
+        return &fault_readback_buffer;
     }
 
     /// Retrieves the buffer with the specified id.

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -459,9 +459,24 @@ bool Rasterizer::BindResources(const Pipeline* pipeline) {
         stage->PushUd(binding, push_data);
         BindBuffers(*stage, binding, push_data);
         BindTextures(*stage, binding);
+        
+        dma_enabled |= stage->uses_dma;
     }
 
     pipeline->BindResources(set_writes, buffer_barriers, push_data);
+
+    if (dma_enabled) {
+        // First, import any queued host memory, then sync every mapped
+        // region that is cached on GPU memory.
+        buffer_cache.ImportQueuedRegions();
+        {
+            std::shared_lock lock(mapped_ranges_mutex);
+            for (const auto& range : mapped_ranges) {
+                buffer_cache.SynchronizeRange(range.lower(), range.upper() - range.lower());
+            }
+        }
+    }
+
     return true;
 }
 
@@ -526,6 +541,12 @@ void Rasterizer::BindBuffers(const Shader::Info& stage, Shader::Backend::Binding
                 const u64 offset = vk_buffer.Copy(stage.flattened_ud_buf.data(), ubo_size,
                                                   instance.UniformMinAlignment());
                 buffer_infos.emplace_back(vk_buffer.Handle(), offset, ubo_size);
+            } else if (desc.buffer_type == Shader::BufferType::BdaPagetable) {
+                const auto* bda_buffer = buffer_cache.GetBdaPageTableBuffer();
+                buffer_infos.emplace_back(bda_buffer->Handle(), 0, bda_buffer->SizeBytes());
+            } else if (desc.buffer_type == Shader::BufferType::FaultReadback) {
+                const auto* fault_buffer = buffer_cache.GetFaultReadbackBuffer();
+                buffer_infos.emplace_back(fault_buffer->Handle(), 0, fault_buffer->SizeBytes());
             } else if (desc.buffer_type == Shader::BufferType::SharedMemory) {
                 auto& lds_buffer = buffer_cache.GetStreamBuffer();
                 const auto& cs_program = liverpool->GetCsRegs();

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -95,6 +95,13 @@ private:
             texture_cache.GetImage(image_id).binding.Reset();
         }
         bound_images.clear();
+
+        if (dma_enabled) {
+            dma_enabled = false;
+            // If a shader accesses a buffer that is not cached, we need to
+            // cache it.
+            buffer_cache.CreateFaultBuffers();
+        }
     }
 
     bool IsComputeMetaClear(const Pipeline* pipeline);
@@ -126,6 +133,7 @@ private:
     boost::container::static_vector<BufferBindingInfo, Shader::NumBuffers> buffer_bindings;
     using ImageBindingInfo = std::pair<VideoCore::ImageId, VideoCore::TextureCache::TextureDesc>;
     boost::container::static_vector<ImageBindingInfo, Shader::NumImages> image_bindings;
+    bool dma_enabled{false};
 };
 
 } // namespace Vulkan