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

DMAEngine: Accelerate BufferClear [accelerateDMA Part 2]

1 files changed, 70 insertions, 8 deletions

diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 2871682f6..7373cb62d 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/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;
@@ -463,23 +470,28 @@ 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);
     const std::optional<VAddr> cpu_dest_address = gpu_memory.GpuToCpuAddress(dest_address);
     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);
@@ -525,6 +538,33 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
 }
 
 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) {
     const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
@@ -782,6 +822,27 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
 }
 
 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);
     for (u64 page = addr >> PAGE_BITS; page < page_end;) {
@@ -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) {