Merge pull request #6629 from FernandoS27/accel-dma-2

DMAEngine: Accelerate BufferClear [accelerateDMA Part 2]

src/video_core/buffer_cache/buffer_cache.h

@@ -164,11 +164,16 @@ public:
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
 
-    [[nodiscard]] bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount);
+    bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount);
+
+    bool DMAClear(GPUVAddr src_address, u64 amount, u32 value);
 
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
+    /// Return true when a region is registered on the cache
+    [[nodiscard]] bool IsRegionRegistered(VAddr addr, size_t size);
+
     /// Return true when a CPU region is modified from the CPU
     [[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size);
 
@@ -324,6 +329,8 @@ private:
 
     [[nodiscard]] bool HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept;
 
+    void ClearDownload(IntervalType subtract_interval);
+
     VideoCore::RasterizerInterface& rasterizer;
     Tegra::Engines::Maxwell3D& maxwell3d;
     Tegra::Engines::KeplerCompute& kepler_compute;
@@ -462,6 +469,14 @@ void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) {
     });
 }
 
+template <class P>
+void BufferCache<P>::ClearDownload(IntervalType subtract_interval) {
+    uncommitted_ranges.subtract(subtract_interval);
+    for (auto& interval_set : committed_ranges) {
+        interval_set.subtract(subtract_interval);
+    }
+}
+
 template <class P>
 bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) {
     const std::optional<VAddr> cpu_src_address = gpu_memory.GpuToCpuAddress(src_address);
@@ -469,17 +484,14 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     if (!cpu_src_address || !cpu_dest_address) {
         return false;
     }
-    const bool source_dirty = IsRegionGpuModified(*cpu_src_address, amount);
-    const bool dest_dirty = IsRegionGpuModified(*cpu_dest_address, amount);
+    const bool source_dirty = IsRegionRegistered(*cpu_src_address, amount);
+    const bool dest_dirty = IsRegionRegistered(*cpu_dest_address, amount);
     if (!source_dirty && !dest_dirty) {
         return false;
     }
 
     const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount};
-    uncommitted_ranges.subtract(subtract_interval);
-    for (auto& interval_set : committed_ranges) {
-        interval_set.subtract(subtract_interval);
-    }
+    ClearDownload(subtract_interval);
 
     BufferId buffer_a;
     BufferId buffer_b;
@@ -510,12 +522,13 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     ForEachWrittenRange(*cpu_src_address, amount, mirror);
     // This subtraction in this order is important for overlapping copies.
     common_ranges.subtract(subtract_interval);
+    bool atleast_1_download = tmp_intervals.size() != 0;
     for (const IntervalType add_interval : tmp_intervals) {
         common_ranges.add(add_interval);
     }
 
     runtime.CopyBuffer(dest_buffer, src_buffer, copies);
-    if (source_dirty) {
+    if (atleast_1_download) {
         dest_buffer.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
     std::vector<u8> tmp_buffer(amount);
@@ -524,6 +537,33 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     return true;
 }
 
+template <class P>
+bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
+    const std::optional<VAddr> cpu_dst_address = gpu_memory.GpuToCpuAddress(dst_address);
+    if (!cpu_dst_address) {
+        return false;
+    }
+    const bool dest_dirty = IsRegionRegistered(*cpu_dst_address, amount);
+    if (!dest_dirty) {
+        return false;
+    }
+
+    const size_t size = amount * sizeof(u32);
+    const IntervalType subtract_interval{*cpu_dst_address, *cpu_dst_address + size};
+    ClearDownload(subtract_interval);
+    common_ranges.subtract(subtract_interval);
+
+    BufferId buffer;
+    do {
+        has_deleted_buffers = false;
+        buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size));
+    } while (has_deleted_buffers);
+    auto& dest_buffer = slot_buffers[buffer];
+    const u32 offset = static_cast<u32>(*cpu_dst_address - dest_buffer.CpuAddr());
+    runtime.ClearBuffer(dest_buffer, offset, size, value);
+    return true;
+}
+
 template <class P>
 void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
                                                u32 size) {
@@ -781,6 +821,27 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
     return false;
 }
 
+template <class P>
+bool BufferCache<P>::IsRegionRegistered(VAddr addr, size_t size) {
+    const VAddr end_addr = addr + size;
+    const u64 page_end = Common::DivCeil(end_addr, PAGE_SIZE);
+    for (u64 page = addr >> PAGE_BITS; page < page_end;) {
+        const BufferId buffer_id = page_table[page];
+        if (!buffer_id) {
+            ++page;
+            continue;
+        }
+        Buffer& buffer = slot_buffers[buffer_id];
+        const VAddr buf_start_addr = buffer.CpuAddr();
+        const VAddr buf_end_addr = buf_start_addr + buffer.SizeBytes();
+        if (buf_start_addr < end_addr && addr < buf_end_addr) {
+            return true;
+        }
+        page = Common::DivCeil(end_addr, PAGE_SIZE);
+    }
+    return false;
+}
+
 template <class P>
 bool BufferCache<P>::IsRegionCpuModified(VAddr addr, size_t size) {
     const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE);
@@ -1425,6 +1486,7 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
         const VAddr end_address = start_address + range_size;
         ForEachWrittenRange(start_address, range_size, add_download);
         const IntervalType subtract_interval{start_address, end_address};
+        ClearDownload(subtract_interval);
         common_ranges.subtract(subtract_interval);
     });
     if (total_size_bytes == 0) {

src/video_core/engines/maxwell_dma.cpp

@@ -4,6 +4,7 @@
 
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -12,6 +13,9 @@
 #include "video_core/renderer_base.h"
 #include "video_core/textures/decoders.h"
 
+MICROPROFILE_DECLARE(GPU_DMAEngine);
+MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128));
+
 namespace Tegra::Engines {
 
 using namespace Texture;
@@ -43,6 +47,7 @@ void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
 }
 
 void MaxwellDMA::Launch() {
+    MICROPROFILE_SCOPE(GPU_DMAEngine);
     LOG_TRACE(Render_OpenGL, "DMA copy 0x{:x} -> 0x{:x}", static_cast<GPUVAddr>(regs.offset_in),
               static_cast<GPUVAddr>(regs.offset_out));
 
@@ -87,9 +92,11 @@ void MaxwellDMA::CopyPitchToPitch() {
         // TODO: allow multisized components.
         if (is_buffer_clear) {
             ASSERT(regs.remap_const.component_size_minus_one == 3);
+            accelerate.BufferClear(regs.offset_out, regs.line_length_in, regs.remap_consta_value);
             std::vector<u32> tmp_buffer(regs.line_length_in, regs.remap_consta_value);
-            memory_manager.WriteBlock(regs.offset_out, reinterpret_cast<u8*>(tmp_buffer.data()),
-                                      regs.line_length_in * sizeof(u32));
+            memory_manager.WriteBlockUnsafe(regs.offset_out,
+                                            reinterpret_cast<u8*>(tmp_buffer.data()),
+                                            regs.line_length_in * sizeof(u32));
             return;
         }
         UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
@@ -179,8 +186,13 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
         write_buffer.resize(dst_size);
     }
 
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
-    memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    if (Settings::IsGPULevelExtreme()) {
+        memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    } else {
+        memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
+    }
 
     // If the input is linear and the output is tiled, swizzle the input and copy it over.
     if (regs.dst_params.block_size.depth > 0) {

src/video_core/engines/maxwell_dma.h

@@ -31,6 +31,8 @@ class AccelerateDMAInterface {
 public:
     /// Write the value to the register identified by method.
     virtual bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) = 0;
+
+    virtual bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) = 0;
 };
 
 /**

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -98,6 +98,12 @@ void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
     }
 }
 
+void BufferCacheRuntime::ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value) {
+    glClearNamedBufferSubData(dest_buffer.Handle(), GL_R32UI, static_cast<GLintptr>(offset),
+                              static_cast<GLsizeiptr>(size / sizeof(u32)), GL_RGBA, GL_UNSIGNED_INT,
+                              &value);
+}
+
 void BufferCacheRuntime::BindIndexBuffer(Buffer& buffer, u32 offset, u32 size) {
     if (has_unified_vertex_buffers) {
         buffer.MakeResident(GL_READ_ONLY);

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -57,6 +57,8 @@ public:
     void CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
                     std::span<const VideoCommon::BufferCopy> copies);
 
+    void ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value);
+
     void BindIndexBuffer(Buffer& buffer, u32 offset, u32 size);
 
     void BindVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size, u32 stride);

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -1407,4 +1407,9 @@ bool AccelerateDMA::BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64
     return buffer_cache.DMACopy(src_address, dest_address, amount);
 }
 
+bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) {
+    std::scoped_lock lock{buffer_cache.mutex};
+    return buffer_cache.DMAClear(src_address, amount, value);
+}
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -65,6 +65,8 @@ public:
 
     bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) override;
 
+    bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
+
 private:
     BufferCache& buffer_cache;
 };

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -136,6 +136,30 @@ void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
     });
 }
 
+void BufferCacheRuntime::ClearBuffer(VkBuffer dest_buffer, u32 offset, size_t size, u32 value) {
+    static constexpr VkMemoryBarrier READ_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
+    };
+    static constexpr VkMemoryBarrier WRITE_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+    };
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([dest_buffer, offset, size, value](vk::CommandBuffer cmdbuf) {
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, READ_BARRIER);
+        cmdbuf.FillBuffer(dest_buffer, offset, size, value);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               0, WRITE_BARRIER);
+    });
+}
+
 void BufferCacheRuntime::BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format,
                                          u32 base_vertex, u32 num_indices, VkBuffer buffer,
                                          u32 offset, [[maybe_unused]] u32 size) {

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -60,6 +60,8 @@ public:
     void CopyBuffer(VkBuffer src_buffer, VkBuffer dst_buffer,
                     std::span<const VideoCommon::BufferCopy> copies);
 
+    void ClearBuffer(VkBuffer dest_buffer, u32 offset, size_t size, u32 value);
+
     void BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format, u32 num_indices,
                          u32 base_vertex, VkBuffer buffer, u32 offset, u32 size);
 

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -706,6 +706,11 @@ void RasterizerVulkan::FlushWork() {
 
 AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_) : buffer_cache{buffer_cache_} {}
 
+bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) {
+    std::scoped_lock lock{buffer_cache.mutex};
+    return buffer_cache.DMAClear(src_address, amount, value);
+}
+
 bool AccelerateDMA::BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) {
     std::scoped_lock lock{buffer_cache.mutex};
     return buffer_cache.DMACopy(src_address, dest_address, amount);

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -56,6 +56,8 @@ public:
 
     bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
 
+    bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
+
 private:
     BufferCache& buffer_cache;
 };