liverpool: More dma data impl

src/core/address_space.h

@@ -45,6 +45,10 @@ public:
     explicit AddressSpace();
     ~AddressSpace();
 
+    [[nodiscard]] u8* BackingBase() const noexcept {
+        return backing_base;
+    }
+
     [[nodiscard]] VAddr SystemManagedVirtualBase() noexcept {
         return reinterpret_cast<VAddr>(system_managed_base);
     }

src/core/memory.cpp

@@ -54,6 +54,17 @@ void MemoryManager::SetupMemoryRegions(u64 flexible_size) {
              total_flexible_size, total_direct_size);
 }
 
+bool MemoryManager::TryWriteBacking(void* address, const void* data, u32 num_bytes) {
+    const VAddr virtual_addr = std::bit_cast<VAddr>(address);
+    const auto& vma = FindVMA(virtual_addr)->second;
+    if (vma.type != VMAType::Direct) {
+        return false;
+    }
+    u8* backing = impl.BackingBase() + vma.phys_base + (virtual_addr - vma.base);
+    memcpy(backing, data, num_bytes);
+    return true;
+}
+
 PAddr MemoryManager::PoolExpand(PAddr search_start, PAddr search_end, size_t size, u64 alignment) {
     std::scoped_lock lk{mutex};
 

src/core/memory.h

@@ -149,6 +149,8 @@ public:
         return impl.SystemReservedVirtualBase();
     }
 
+    bool TryWriteBacking(void* address, const void* data, u32 num_bytes);
+
     void SetupMemoryRegions(u64 flexible_size);
 
     PAddr PoolExpand(PAddr search_start, PAddr search_end, size_t size, u64 alignment);

src/video_core/amdgpu/liverpool.cpp

@@ -8,6 +8,7 @@
 #include "common/thread.h"
 #include "core/debug_state.h"
 #include "core/libraries/videoout/driver.h"
+#include "core/memory.h"
 #include "video_core/amdgpu/liverpool.h"
 #include "video_core/amdgpu/pm4_cmds.h"
 #include "video_core/renderdoc.h"
@@ -504,7 +505,12 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
             }
             case PM4ItOpcode::EventWriteEos: {
                 const auto* event_eos = reinterpret_cast<const PM4CmdEventWriteEos*>(header);
-                event_eos->SignalFence();
+                event_eos->SignalFence([](void* address, u64 data, u32 num_bytes) {
+                    auto* memory = Core::Memory::Instance();
+                    if (!memory->TryWriteBacking(address, &data, num_bytes)) {
+                        memcpy(address, &data, num_bytes);
+                    }
+                });
                 if (event_eos->command == PM4CmdEventWriteEos::Command::GdsStore) {
                     ASSERT(event_eos->size == 1);
                     if (rasterizer) {
@@ -517,13 +523,42 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
             }
             case PM4ItOpcode::EventWriteEop: {
                 const auto* event_eop = reinterpret_cast<const PM4CmdEventWriteEop*>(header);
-                event_eop->SignalFence();
+                event_eop->SignalFence([](void* address, u64 data, u32 num_bytes) {
+                    auto* memory = Core::Memory::Instance();
+                    if (!memory->TryWriteBacking(address, &data, num_bytes)) {
+                        memcpy(address, &data, num_bytes);
+                    }
+                });
                 break;
             }
             case PM4ItOpcode::DmaData: {
                 const auto* dma_data = reinterpret_cast<const PM4DmaData*>(header);
+                if (dma_data->dst_addr_lo == 0x3022C) {
+                    break;
+                }
                 if (dma_data->src_sel == DmaDataSrc::Data && dma_data->dst_sel == DmaDataDst::Gds) {
-                    rasterizer->InlineDataToGds(dma_data->dst_addr_lo, dma_data->data);
+                    rasterizer->InlineData(dma_data->dst_addr_lo, &dma_data->data, sizeof(u32),
+                                           true);
+                } else if (dma_data->src_sel == DmaDataSrc::Memory &&
+                           dma_data->dst_sel == DmaDataDst::Gds) {
+                    rasterizer->InlineData(dma_data->dst_addr_lo,
+                                           dma_data->SrcAddress<const void*>(),
+                                           dma_data->NumBytes(), true);
+                } else if (dma_data->src_sel == DmaDataSrc::Data &&
+                           dma_data->dst_sel == DmaDataDst::Memory) {
+                    rasterizer->InlineData(dma_data->DstAddress<VAddr>(), &dma_data->data,
+                                           sizeof(u32), false);
+                } else if (dma_data->src_sel == DmaDataSrc::Gds &&
+                           dma_data->dst_sel == DmaDataDst::Memory) {
+                    LOG_WARNING(Render_Vulkan, "GDS memory read");
+                } else if (dma_data->src_sel == DmaDataSrc::Memory &&
+                           dma_data->dst_sel == DmaDataDst::Memory) {
+                    rasterizer->InlineData(dma_data->DstAddress<VAddr>(),
+                                           dma_data->SrcAddress<const void*>(),
+                                           dma_data->NumBytes(), false);
+                } else {
+                    UNREACHABLE_MSG("WriteData src_sel = {}, dst_sel = {}",
+                                    u32(dma_data->src_sel.Value()), u32(dma_data->dst_sel.Value()));
                 }
                 break;
             }

src/video_core/amdgpu/pm4_cmds.h

@@ -313,25 +313,26 @@ struct PM4CmdEventWriteEop {
         return data_lo | u64(data_hi) << 32;
     }
 
-    void SignalFence() const {
+    void SignalFence(auto&& write_mem) const {
+        u32* address = Address<u32>();
         switch (data_sel.Value()) {
         case DataSelect::None: {
             break;
         }
         case DataSelect::Data32Low: {
-            *Address<u32>() = DataDWord();
+            write_mem(address, DataDWord(), sizeof(u32));
             break;
         }
         case DataSelect::Data64: {
-            *Address<u64>() = DataQWord();
+            write_mem(address, DataQWord(), sizeof(u64));
             break;
         }
         case DataSelect::GpuClock64: {
-            *Address<u64>() = GetGpuClock64();
+            write_mem(address, GetGpuClock64(), sizeof(u64));
             break;
         }
         case DataSelect::PerfCounter: {
-            *Address<u64>() = Common::FencedRDTSC();
+            write_mem(address, Common::FencedRDTSC(), sizeof(u64));
             break;
         }
         default: {
@@ -401,6 +402,20 @@ struct PM4DmaData {
     u32 dst_addr_lo;
     u32 dst_addr_hi;
     u32 command;
+
+    template <typename T>
+    T SrcAddress() const {
+        return reinterpret_cast<T>(src_addr_lo | u64(src_addr_hi) << 32);
+    }
+
+    template <typename T>
+    T DstAddress() const {
+        return reinterpret_cast<T>(dst_addr_lo | u64(dst_addr_hi) << 32);
+    }
+
+    u32 NumBytes() const noexcept {
+        return command & 0x1fffff;
+    }
 };
 
 struct PM4CmdWaitRegMem {
@@ -534,11 +549,11 @@ struct PM4CmdEventWriteEos {
         return this->data;
     }
 
-    void SignalFence() const {
+    void SignalFence(auto&& write_mem) const {
         const auto cmd = command.Value();
         switch (cmd) {
         case Command::SignalFence: {
-            *Address() = DataDWord();
+            write_mem(Address(), DataDWord(), sizeof(u32));
             break;
         }
         case Command::GdsStore: {

src/video_core/buffer_cache/buffer_cache.cpp

@@ -20,7 +20,7 @@ static constexpr size_t StagingBufferSize = 1_GB;
 static constexpr size_t UboStreamBufferSize = 64_MB;
 
 BufferCache::BufferCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
-                         const AmdGpu::Liverpool* liverpool_, TextureCache& texture_cache_,
+                         AmdGpu::Liverpool* liverpool_, TextureCache& texture_cache_,
                          PageManager& tracker_)
     : instance{instance_}, scheduler{scheduler_}, liverpool{liverpool_},
       texture_cache{texture_cache_}, tracker{tracker_},
@@ -70,11 +70,10 @@ void BufferCache::InvalidateMemory(VAddr device_addr, u64 size) {
 void BufferCache::DownloadBufferMemory(Buffer& buffer, VAddr device_addr, u64 size) {
     boost::container::small_vector<vk::BufferCopy, 1> copies;
     u64 total_size_bytes = 0;
-    u64 largest_copy = 0;
     memory_tracker.ForEachDownloadRange<true>(
         device_addr, size, [&](u64 device_addr_out, u64 range_size) {
             const VAddr buffer_addr = buffer.CpuAddr();
-            const auto add_download = [&](VAddr start, VAddr end, u64) {
+            const auto add_download = [&](VAddr start, VAddr end) {
                 const u64 new_offset = start - buffer_addr;
                 const u64 new_size = end - start;
                 copies.push_back(vk::BufferCopy{
@@ -82,12 +81,10 @@ void BufferCache::DownloadBufferMemory(Buffer& buffer, VAddr device_addr, u64 si
                     .dstOffset = total_size_bytes,
                     .size = new_size,
                 });
-                // Align up to avoid cache conflicts
+                total_size_bytes += new_size;
-                constexpr u64 align = 64ULL;
-                constexpr u64 mask = ~(align - 1ULL);
-                total_size_bytes += (new_size + align - 1) & mask;
-                largest_copy = std::max(largest_copy, new_size);
             };
+            gpu_modified_ranges.ForEachInRange(device_addr_out, range_size, add_download);
+            gpu_modified_ranges.Subtract(device_addr_out, range_size);
         });
     if (total_size_bytes == 0) {
         return;
@@ -181,6 +178,9 @@ bool BufferCache::BindVertexBuffers(const Shader::Info& vs_info) {
             .divisor = 1,
         });
     }
+    if (ranges.empty()) {
+        return false;
+    }
 
     std::ranges::sort(ranges, [](const BufferRange& lhv, const BufferRange& rhv) {
         return lhv.base_address < rhv.base_address;
@@ -269,34 +269,45 @@ u32 BufferCache::BindIndexBuffer(bool& is_indexed, u32 index_offset) {
     return regs.num_indices;
 }
 
-void BufferCache::InlineDataToGds(u32 gds_offset, u32 value) {
+void BufferCache::InlineData(VAddr address, const void* value, u32 num_bytes, bool is_gds) {
-    ASSERT_MSG(gds_offset % 4 == 0, "GDS offset must be dword aligned");
+    ASSERT_MSG(address % 4 == 0, "GDS offset must be dword aligned");
+    if (!is_gds && !IsRegionRegistered(address, num_bytes)) {
+        memcpy(std::bit_cast<void*>(address), value, num_bytes);
+        return;
+    }
     scheduler.EndRendering();
     const auto cmdbuf = scheduler.CommandBuffer();
+    const Buffer* buffer = [&] {
+        if (is_gds) {
+            return &gds_buffer;
+        }
+        const BufferId buffer_id = FindBuffer(address, num_bytes);
+        return &slot_buffers[buffer_id];
+    }();
     const vk::BufferMemoryBarrier2 buf_barrier = {
         .srcStageMask = vk::PipelineStageFlagBits2::eTransfer,
         .srcAccessMask = vk::AccessFlagBits2::eTransferWrite,
         .dstStageMask = vk::PipelineStageFlagBits2::eAllCommands,
         .dstAccessMask = vk::AccessFlagBits2::eMemoryRead,
-        .buffer = gds_buffer.Handle(),
+        .buffer = buffer->Handle(),
-        .offset = gds_offset,
+        .offset = buffer->Offset(address),
-        .size = sizeof(u32),
+        .size = num_bytes,
     };
     cmdbuf.pipelineBarrier2(vk::DependencyInfo{
         .dependencyFlags = vk::DependencyFlagBits::eByRegion,
         .bufferMemoryBarrierCount = 1,
         .pBufferMemoryBarriers = &buf_barrier,
     });
-    cmdbuf.updateBuffer(gds_buffer.Handle(), gds_offset, sizeof(u32), &value);
+    cmdbuf.updateBuffer(buffer->Handle(), buf_barrier.offset, num_bytes, value);
 }
 
 std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, bool is_written,
                                                   bool is_texel_buffer, BufferId buffer_id) {
+    // For small uniform buffers that have not been modified by gpu
+    // use device local stream buffer to reduce renderpass breaks.
     static constexpr u64 StreamThreshold = CACHING_PAGESIZE;
     const bool is_gpu_dirty = memory_tracker.IsRegionGpuModified(device_addr, size);
     if (!is_written && size <= StreamThreshold && !is_gpu_dirty) {
-        // For small uniform buffers that have not been modified by gpu
-        // use device local stream buffer to reduce renderpass breaks.
         const u64 offset = stream_buffer.Copy(device_addr, size, instance.UniformMinAlignment());
         return {&stream_buffer, offset};
     }
@@ -308,7 +319,7 @@ std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, b
     SynchronizeBuffer(buffer, device_addr, size, is_texel_buffer);
     if (is_written) {
         memory_tracker.MarkRegionAsGpuModified(device_addr, size);
-        gpu_regions.Add(device_addr, size);
+        gpu_modified_ranges.Add(device_addr, size);
     }
     return {&buffer, buffer.Offset(device_addr)};
 }
@@ -477,7 +488,7 @@ void BufferCache::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
     cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
                            vk::PipelineStageFlagBits::eAllCommands,
                            vk::DependencyFlagBits::eByRegion, WRITE_BARRIER, {}, {});
-    DeleteBuffer(overlap_id, true);
+    DeleteBuffer(overlap_id);
 }
 
 BufferId BufferCache::CreateBuffer(VAddr device_addr, u32 wanted_size) {
@@ -532,7 +543,7 @@ void BufferCache::SynchronizeBuffer(Buffer& buffer, VAddr device_addr, u32 size,
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
     VAddr buffer_start = buffer.CpuAddr();
-    const auto add_copy = [&](VAddr device_addr_out, u64 range_size) {
+    memory_tracker.ForEachUploadRange(device_addr, size, [&](u64 device_addr_out, u64 range_size) {
         copies.push_back(vk::BufferCopy{
             .srcOffset = total_size_bytes,
             .dstOffset = device_addr_out - buffer_start,
@@ -540,13 +551,6 @@ void BufferCache::SynchronizeBuffer(Buffer& buffer, VAddr device_addr, u32 size,
         });
         total_size_bytes += range_size;
         largest_copy = std::max(largest_copy, range_size);
-    };
-    memory_tracker.ForEachUploadRange(device_addr, size, [&](u64 device_addr_out, u64 range_size) {
-        bool has_gpu = false;
-        gpu_regions.ForEachInRange(device_addr_out, range_size, [&](VAddr, VAddr) { has_gpu = true; });
-        add_copy(device_addr_out, range_size);
-        // Prevent uploading to gpu modified regions.
-        // gpu_modified_ranges.ForEachNotInRange(device_addr_out, range_size, add_copy);
     });
     SCOPE_EXIT {
         if (is_texel_buffer) {
@@ -659,12 +663,8 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, VAddr device_addr,
     return true;
 }
 
-void BufferCache::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
+void BufferCache::DeleteBuffer(BufferId buffer_id) {
-    // Mark the whole buffer as CPU written to stop tracking CPU writes
     Buffer& buffer = slot_buffers[buffer_id];
-    if (!do_not_mark) {
-        memory_tracker.MarkRegionAsCpuModified(buffer.CpuAddr(), buffer.SizeBytes());
-    }
     Unregister(buffer_id);
     scheduler.DeferOperation([this, buffer_id] { slot_buffers.erase(buffer_id); });
     buffer.is_deleted = true;

src/video_core/buffer_cache/buffer_cache.h

@@ -54,7 +54,7 @@ public:
 
 public:
     explicit BufferCache(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
-                         const AmdGpu::Liverpool* liverpool, TextureCache& texture_cache,
+                         AmdGpu::Liverpool* liverpool, TextureCache& texture_cache,
                          PageManager& tracker);
     ~BufferCache();
 
@@ -81,12 +81,13 @@ public:
     /// Bind host index buffer for the current draw.
     u32 BindIndexBuffer(bool& is_indexed, u32 index_offset);
 
-    /// Writes a value to GDS buffer.
+    /// Writes a value to GPU buffer.
-    void InlineDataToGds(u32 gds_offset, u32 value);
+    void InlineData(VAddr address, const void* value, u32 num_bytes, bool is_gds);
 
     /// Obtains a buffer for the specified region.
     [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(VAddr gpu_addr, u32 size, bool is_written,
-                                                       bool is_texel_buffer = false, BufferId buffer_id = {});
+                                                       bool is_texel_buffer = false,
+                                                       BufferId buffer_id = {});
 
     /// Attempts to obtain a buffer without modifying the cache contents.
     [[nodiscard]] std::pair<Buffer*, u32> ObtainViewBuffer(VAddr gpu_addr, u32 size);
@@ -139,11 +140,11 @@ private:
 
     bool SynchronizeBufferFromImage(Buffer& buffer, VAddr device_addr, u32 size);
 
-    void DeleteBuffer(BufferId buffer_id, bool do_not_mark = false);
+    void DeleteBuffer(BufferId buffer_id);
 
     const Vulkan::Instance& instance;
     Vulkan::Scheduler& scheduler;
-    const AmdGpu::Liverpool* liverpool;
+    AmdGpu::Liverpool* liverpool;
     TextureCache& texture_cache;
     PageManager& tracker;
     StreamBuffer staging_buffer;
@@ -151,7 +152,7 @@ private:
     Buffer gds_buffer;
     std::mutex mutex;
     Common::SlotVector<Buffer> slot_buffers;
-    RangeSet gpu_regions;
+    RangeSet gpu_modified_ranges;
     vk::BufferView null_buffer_view;
     MemoryTracker memory_tracker;
     PageTable page_table;

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -119,14 +119,15 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
 
     info->PushUd(binding, push_data);
 
+    buffer_infos.clear();
+    buffer_views.clear();
+    image_infos.clear();
+
     // Most of the time when a metadata is updated with a shader it gets cleared. It means
     // we can skip the whole dispatch and update the tracked state instead. Also, it is not
     // intended to be consumed and in such rare cases (e.g. HTile introspection, CRAA) we
     // will need its full emulation anyways. For cases of metadata read a warning will be logged.
-    for (const auto& desc : info->buffers) {
+    for (const auto& desc : info->texture_buffers) {
-        if (desc.is_gds_buffer) {
-            continue;
-        }
         const VAddr address = desc.GetSharp(*info).base_address;
         if (desc.is_written) {
             if (texture_cache.TouchMeta(address, true)) {
@@ -140,14 +141,15 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
         }
     }
 
-    BindBuffers(buffer_cache, texture_cache, *info, binding, push_data,
+    BindBuffers(buffer_cache, texture_cache, *info, binding, push_data, set_writes,
-                       set_writes, buffer_barriers);
+                buffer_barriers);
 
     BindTextures(texture_cache, *info, binding, set_writes);
 
     if (set_writes.empty()) {
         return false;
     }
+
     const auto cmdbuf = scheduler.CommandBuffer();
     if (!buffer_barriers.empty()) {
         const auto dependencies = vk::DependencyInfo{
@@ -158,8 +160,10 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
         scheduler.EndRendering();
         cmdbuf.pipelineBarrier2(dependencies);
     }
+
     cmdbuf.pushConstants(*pipeline_layout, vk::ShaderStageFlagBits::eCompute, 0u, sizeof(push_data),
                          &push_data);
+
     // Bind descriptor set.
     if (uses_push_descriptors) {
         cmdbuf.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, *pipeline_layout, 0,
@@ -171,8 +175,7 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
         set_write.dstSet = desc_set;
     }
     instance.GetDevice().updateDescriptorSets(set_writes, {});
-    cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eCompute, *pipeline_layout, 0, desc_set,
+    cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eCompute, *pipeline_layout, 0, desc_set, {});
-                              {});
 
     return true;
 }

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -5,8 +5,8 @@
 #include <boost/container/small_vector.hpp>
 #include <boost/container/static_vector.hpp>
 
-#include "common/scope_exit.h"
 #include "common/assert.h"
+#include "common/scope_exit.h"
 #include "video_core/amdgpu/resource.h"
 #include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_graphics_pipeline.h"
@@ -389,6 +389,10 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
     Shader::PushData push_data{};
     Shader::Backend::Bindings binding{};
 
+    buffer_infos.clear();
+    buffer_views.clear();
+    image_infos.clear();
+
     for (const auto* stage : stages) {
         if (!stage) {
             continue;
@@ -399,8 +403,8 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
         }
         stage->PushUd(binding, push_data);
 
-        BindBuffers(buffer_cache, texture_cache, *stage, binding, push_data,
+        BindBuffers(buffer_cache, texture_cache, *stage, binding, push_data, set_writes,
-                           set_writes, buffer_barriers);
+                    buffer_barriers);
 
         BindTextures(texture_cache, *stage, binding, set_writes);
     }
@@ -414,6 +418,7 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
     if (set_writes.empty()) {
         return;
     }
+
     if (!buffer_barriers.empty()) {
         const auto dependencies = vk::DependencyInfo{
             .dependencyFlags = vk::DependencyFlagBits::eByRegion,
@@ -423,6 +428,7 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
         scheduler.EndRendering();
         cmdbuf.pipelineBarrier2(dependencies);
     }
+
     // Bind descriptor set.
     if (uses_push_descriptors) {
         cmdbuf.pushDescriptorSetKHR(vk::PipelineBindPoint::eGraphics, *pipeline_layout, 0,
@@ -434,8 +440,7 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
         set_write.dstSet = desc_set;
     }
     instance.GetDevice().updateDescriptorSets(set_writes, {});
-    cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eGraphics, *pipeline_layout, 0,
+    cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eGraphics, *pipeline_layout, 0, desc_set, {});
-                              desc_set, {});
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_pipeline_common.cpp

@@ -4,11 +4,11 @@
 #include <boost/container/static_vector.hpp>
 
 #include "shader_recompiler/info.h"
+#include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_pipeline_common.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/texture_cache/texture_cache.h"
-#include "video_core/buffer_cache/buffer_cache.h"
 
 namespace Vulkan {
 
@@ -22,15 +22,14 @@ Pipeline::Pipeline(const Instance& instance_, Scheduler& scheduler_, DescriptorH
 
 Pipeline::~Pipeline() = default;
 
-void Pipeline::BindBuffers(VideoCore::BufferCache& buffer_cache, VideoCore::TextureCache& texture_cache,
+void Pipeline::BindBuffers(VideoCore::BufferCache& buffer_cache,
-                                  const Shader::Info& stage,
+                           VideoCore::TextureCache& texture_cache, const Shader::Info& stage,
-                                  Shader::Backend::Bindings& binding, Shader::PushData& push_data,
+                           Shader::Backend::Bindings& binding, Shader::PushData& push_data,
-                                  DescriptorWrites& set_writes, BufferBarriers& buffer_barriers) const {
+                           DescriptorWrites& set_writes, BufferBarriers& buffer_barriers) const {
     using BufferBindingInfo = std::pair<VideoCore::BufferId, AmdGpu::Buffer>;
     static boost::container::static_vector<BufferBindingInfo, 32> buffer_bindings;
 
     buffer_bindings.clear();
-    buffer_infos.clear();
 
     for (const auto& desc : stage.buffers) {
         const auto vsharp = desc.GetSharp(stage);
@@ -46,7 +45,6 @@ void Pipeline::BindBuffers(VideoCore::BufferCache& buffer_cache, VideoCore::Text
     static boost::container::static_vector<TexBufferBindingInfo, 32> texbuffer_bindings;
 
     texbuffer_bindings.clear();
-    buffer_views.clear();
 
     for (const auto& desc : stage.texture_buffers) {
         const auto vsharp = desc.GetSharp(stage);
@@ -75,16 +73,16 @@ void Pipeline::BindBuffers(VideoCore::BufferCache& buffer_cache, VideoCore::Text
                 buffer_infos.emplace_back(null_buffer.Handle(), 0, VK_WHOLE_SIZE);
             }
         } else {
-            const auto [vk_buffer, offset] =
+            const auto [vk_buffer, offset] = buffer_cache.ObtainBuffer(
-                buffer_cache.ObtainBuffer(vsharp.base_address, vsharp.GetSize(), desc.is_written,
+                vsharp.base_address, vsharp.GetSize(), desc.is_written, false, buffer_id);
-                                          false, buffer_id);
             const u32 alignment =
                 is_storage ? instance.StorageMinAlignment() : instance.UniformMinAlignment();
             const u32 offset_aligned = Common::AlignDown(offset, alignment);
             const u32 adjust = offset - offset_aligned;
             ASSERT(adjust % 4 == 0);
             push_data.AddOffset(binding.buffer, adjust);
-            buffer_infos.emplace_back(vk_buffer->Handle(), offset_aligned, vsharp.GetSize() + adjust);
+            buffer_infos.emplace_back(vk_buffer->Handle(), offset_aligned,
+                                      vsharp.GetSize() + adjust);
         }
 
         set_writes.push_back({
@@ -107,9 +105,8 @@ void Pipeline::BindBuffers(VideoCore::BufferCache& buffer_cache, VideoCore::Text
         vk::BufferView& buffer_view = buffer_views.emplace_back(null_buffer_view);
         if (buffer_id) {
             const u32 alignment = instance.TexelBufferMinAlignment();
-            const auto [vk_buffer, offset] =
+            const auto [vk_buffer, offset] = buffer_cache.ObtainBuffer(
-                buffer_cache.ObtainBuffer(vsharp.base_address, vsharp.GetSize(), desc.is_written, true,
+                vsharp.base_address, vsharp.GetSize(), desc.is_written, true, buffer_id);
-                                          buffer_id);
             const u32 fmt_stride = AmdGpu::NumBits(vsharp.GetDataFmt()) >> 3;
             ASSERT_MSG(fmt_stride == vsharp.GetStride(),
                        "Texel buffer stride must match format stride");
@@ -117,12 +114,13 @@ void Pipeline::BindBuffers(VideoCore::BufferCache& buffer_cache, VideoCore::Text
             const u32 adjust = offset - offset_aligned;
             ASSERT(adjust % fmt_stride == 0);
             push_data.AddOffset(binding.buffer, adjust / fmt_stride);
-            buffer_view = vk_buffer->View(offset_aligned, vsharp.GetSize() + adjust, desc.is_written,
+            buffer_view =
-                                          vsharp.GetDataFmt(), vsharp.GetNumberFmt());
+                vk_buffer->View(offset_aligned, vsharp.GetSize() + adjust, desc.is_written,
+                                vsharp.GetDataFmt(), vsharp.GetNumberFmt());
             if (auto barrier =
-                vk_buffer->GetBarrier(desc.is_written ? vk::AccessFlagBits2::eShaderWrite
+                    vk_buffer->GetBarrier(desc.is_written ? vk::AccessFlagBits2::eShaderWrite
-                                                      : vk::AccessFlagBits2::eShaderRead,
+                                                          : vk::AccessFlagBits2::eShaderRead,
-                                      vk::PipelineStageFlagBits2::eComputeShader)) {
+                                          vk::PipelineStageFlagBits2::eComputeShader)) {
                 buffer_barriers.emplace_back(*barrier);
             }
             if (desc.is_written) {
@@ -151,7 +149,6 @@ void Pipeline::BindTextures(VideoCore::TextureCache& texture_cache, const Shader
     static boost::container::static_vector<ImageBindingInfo, 32> image_bindings;
 
     image_bindings.clear();
-    image_infos.clear();
 
     for (const auto& image_desc : stage.images) {
         const auto tsharp = image_desc.GetSharp(stage);

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -258,8 +258,8 @@ void Rasterizer::BeginRendering(const GraphicsPipeline& pipeline) {
     scheduler.BeginRendering(state);
 }
 
-void Rasterizer::InlineDataToGds(u32 gds_offset, u32 value) {
+void Rasterizer::InlineData(VAddr address, const void* value, u32 num_bytes, bool is_gds) {
-    buffer_cache.InlineDataToGds(gds_offset, value);
+    buffer_cache.InlineData(address, value, num_bytes, is_gds);
 }
 
 u32 Rasterizer::ReadDataFromGds(u32 gds_offset) {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -42,7 +42,7 @@ public:
     void ScopedMarkerInsert(const std::string_view& str);
     void ScopedMarkerInsertColor(const std::string_view& str, const u32 color);
 
-    void InlineDataToGds(u32 gds_offset, u32 value);
+    void InlineData(VAddr address, const void* value, u32 num_bytes, bool is_gds);
     u32 ReadDataFromGds(u32 gsd_offset);
     void InvalidateMemory(VAddr addr, u64 size);
     void MapMemory(VAddr addr, u64 size);