renderer_vulkan: Introduce shader HLE system with copy shader implementation. (#1683)

* renderer_vulkan: Introduce shader HLE system with copy shader implementation.

Co-authored-by: TheTurtle <47210458+raphaelthegreat@users.noreply.github.com>

* buffer_cache: Handle obtaining buffer views partially within buffers.

* vk_shader_hle: Make more efficient

---------

Co-authored-by: TheTurtle <47210458+raphaelthegreat@users.noreply.github.com>

CMakeLists.txt

@@ -738,6 +738,8 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
                src/video_core/renderer_vulkan/vk_resource_pool.h
                src/video_core/renderer_vulkan/vk_scheduler.cpp
                src/video_core/renderer_vulkan/vk_scheduler.h
+               src/video_core/renderer_vulkan/vk_shader_hle.cpp
+               src/video_core/renderer_vulkan/vk_shader_hle.h
                src/video_core/renderer_vulkan/vk_shader_util.cpp
                src/video_core/renderer_vulkan/vk_shader_util.h
                src/video_core/renderer_vulkan/vk_swapchain.cpp

src/video_core/buffer_cache/buffer_cache.cpp

@@ -360,7 +360,8 @@ std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, b
     return {&buffer, buffer.Offset(device_addr)};
 }
 
-std::pair<Buffer*, u32> BufferCache::ObtainViewBuffer(VAddr gpu_addr, u32 size) {
+std::pair<Buffer*, u32> BufferCache::ObtainViewBuffer(VAddr gpu_addr, u32 size, bool prefer_gpu) {
+    // Check if any buffer contains the full requested range.
     const u64 page = gpu_addr >> CACHING_PAGEBITS;
     const BufferId buffer_id = page_table[page];
     if (buffer_id) {
@@ -370,6 +371,13 @@ std::pair<Buffer*, u32> BufferCache::ObtainViewBuffer(VAddr gpu_addr, u32 size)
             return {&buffer, buffer.Offset(gpu_addr)};
         }
     }
+    // If no buffer contains the full requested range but some buffer within was GPU-modified,
+    // fall back to ObtainBuffer to create a full buffer and avoid losing GPU modifications.
+    // This is only done if the request prefers to use GPU memory, otherwise we can skip it.
+    if (prefer_gpu && memory_tracker.IsRegionGpuModified(gpu_addr, size)) {
+        return ObtainBuffer(gpu_addr, size, false, false);
+    }
+    // In all other cases, just do a CPU copy to the staging buffer.
     const u32 offset = staging_buffer.Copy(gpu_addr, size, 16);
     return {&staging_buffer, offset};
 }

src/video_core/buffer_cache/buffer_cache.h

@@ -96,7 +96,8 @@ public:
                                                        BufferId buffer_id = {});
 
     /// Attempts to obtain a buffer without modifying the cache contents.
-    [[nodiscard]] std::pair<Buffer*, u32> ObtainViewBuffer(VAddr gpu_addr, u32 size);
+    [[nodiscard]] std::pair<Buffer*, u32> ObtainViewBuffer(VAddr gpu_addr, u32 size,
+                                                           bool prefer_gpu);
 
     /// Return true when a region is registered on the cache
     [[nodiscard]] bool IsRegionRegistered(VAddr addr, size_t size);

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -8,6 +8,7 @@
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_rasterizer.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_shader_hle.h"
 #include "video_core/texture_cache/image_view.h"
 #include "video_core/texture_cache/texture_cache.h"
 #include "vk_rasterizer.h"
@@ -318,6 +319,11 @@ void Rasterizer::DispatchDirect() {
         return;
     }
 
+    const auto& cs = pipeline->GetStage(Shader::Stage::Compute);
+    if (ExecuteShaderHLE(cs, liverpool->regs, *this)) {
+        return;
+    }
+
     if (!BindResources(pipeline)) {
         return;
     }

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -28,6 +28,14 @@ public:
                         AmdGpu::Liverpool* liverpool);
     ~Rasterizer();
 
+    [[nodiscard]] Scheduler& GetScheduler() noexcept {
+        return scheduler;
+    }
+
+    [[nodiscard]] VideoCore::BufferCache& GetBufferCache() noexcept {
+        return buffer_cache;
+    }
+
     [[nodiscard]] VideoCore::TextureCache& GetTextureCache() noexcept {
         return texture_cache;
     }

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -10,6 +10,10 @@
 #include "video_core/renderer_vulkan/vk_master_semaphore.h"
 #include "video_core/renderer_vulkan/vk_resource_pool.h"
 
+namespace tracy {
+class VkCtxScope;
+}
+
 namespace Vulkan {
 
 class Instance;

src/video_core/renderer_vulkan/vk_shader_hle.cpp

@@ -0,0 +1,139 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "shader_recompiler/info.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_shader_hle.h"
+
+#include "vk_rasterizer.h"
+
+namespace Vulkan {
+
+static constexpr u64 COPY_SHADER_HASH = 0xfefebf9f;
+
+bool ExecuteCopyShaderHLE(const Shader::Info& info, const AmdGpu::Liverpool::Regs& regs,
+                          Rasterizer& rasterizer) {
+    auto& scheduler = rasterizer.GetScheduler();
+    auto& buffer_cache = rasterizer.GetBufferCache();
+
+    // Copy shader defines three formatted buffers as inputs: control, source, and destination.
+    const auto ctl_buf_sharp = info.texture_buffers[0].GetSharp(info);
+    const auto src_buf_sharp = info.texture_buffers[1].GetSharp(info);
+    const auto dst_buf_sharp = info.texture_buffers[2].GetSharp(info);
+    const auto buf_stride = src_buf_sharp.GetStride();
+    ASSERT(buf_stride == dst_buf_sharp.GetStride());
+
+    struct CopyShaderControl {
+        u32 dst_idx;
+        u32 src_idx;
+        u32 end;
+    };
+    static_assert(sizeof(CopyShaderControl) == 12);
+    ASSERT(ctl_buf_sharp.GetStride() == sizeof(CopyShaderControl));
+    const auto ctl_buf = reinterpret_cast<const CopyShaderControl*>(ctl_buf_sharp.base_address);
+
+    static std::vector<vk::BufferCopy> copies;
+    copies.clear();
+    copies.reserve(regs.cs_program.dim_x);
+
+    for (u32 i = 0; i < regs.cs_program.dim_x; i++) {
+        const auto& [dst_idx, src_idx, end] = ctl_buf[i];
+        const u32 local_dst_offset = dst_idx * buf_stride;
+        const u32 local_src_offset = src_idx * buf_stride;
+        const u32 local_size = (end + 1) * buf_stride;
+        copies.emplace_back(local_src_offset, local_dst_offset, local_size);
+    }
+
+    scheduler.EndRendering();
+
+    static constexpr vk::MemoryBarrier READ_BARRIER{
+        .srcAccessMask = vk::AccessFlagBits::eMemoryWrite,
+        .dstAccessMask = vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite,
+    };
+    static constexpr vk::MemoryBarrier WRITE_BARRIER{
+        .srcAccessMask = vk::AccessFlagBits::eTransferWrite,
+        .dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite,
+    };
+    scheduler.CommandBuffer().pipelineBarrier(
+        vk::PipelineStageFlagBits::eAllCommands, vk::PipelineStageFlagBits::eTransfer,
+        vk::DependencyFlagBits::eByRegion, READ_BARRIER, {}, {});
+
+    static constexpr vk::DeviceSize MaxDistanceForMerge = 64_MB;
+    u32 batch_start = 0;
+    u32 batch_end = 1;
+
+    while (batch_end < copies.size()) {
+        // Place first copy into the current batch
+        const auto& copy = copies[batch_start];
+        auto src_offset_min = copy.srcOffset;
+        auto src_offset_max = copy.srcOffset + copy.size;
+        auto dst_offset_min = copy.dstOffset;
+        auto dst_offset_max = copy.dstOffset + copy.size;
+
+        for (int i = batch_start + 1; i < copies.size(); i++) {
+            // Compute new src and dst bounds if we were to batch this copy
+            const auto [src_offset, dst_offset, size] = copies[i];
+            auto new_src_offset_min = std::min(src_offset_min, src_offset);
+            auto new_src_offset_max = std::max(src_offset_max, src_offset + size);
+            if (new_src_offset_max - new_src_offset_min > MaxDistanceForMerge) {
+                continue;
+            }
+
+            auto new_dst_offset_min = std::min(dst_offset_min, dst_offset);
+            auto new_dst_offset_max = std::max(dst_offset_max, dst_offset + size);
+            if (new_dst_offset_max - new_dst_offset_min > MaxDistanceForMerge) {
+                continue;
+            }
+
+            // We can batch this copy
+            src_offset_min = new_src_offset_min;
+            src_offset_max = new_src_offset_max;
+            dst_offset_min = new_dst_offset_min;
+            dst_offset_max = new_dst_offset_max;
+            if (i != batch_end) {
+                std::swap(copies[i], copies[batch_end]);
+            }
+            ++batch_end;
+        }
+
+        // Obtain buffers for the total source and destination ranges.
+        const auto [src_buf, src_buf_offset] =
+            buffer_cache.ObtainBuffer(src_buf_sharp.base_address + src_offset_min,
+                                      src_offset_max - src_offset_min, false, false);
+        const auto [dst_buf, dst_buf_offset] =
+            buffer_cache.ObtainBuffer(dst_buf_sharp.base_address + dst_offset_min,
+                                      dst_offset_max - dst_offset_min, true, false);
+
+        // Apply found buffer base.
+        const auto vk_copies = std::span{copies}.subspan(batch_start, batch_end - batch_start);
+        for (auto& copy : vk_copies) {
+            copy.srcOffset = copy.srcOffset - src_offset_min + src_buf_offset;
+            copy.dstOffset = copy.dstOffset - dst_offset_min + dst_buf_offset;
+        }
+
+        // Execute buffer copies.
+        LOG_TRACE(Render_Vulkan, "HLE buffer copy: src_size = {}, dst_size = {}",
+                  src_offset_max - src_offset_min, dst_offset_max - dst_offset_min);
+        scheduler.CommandBuffer().copyBuffer(src_buf->Handle(), dst_buf->Handle(), vk_copies);
+        batch_start = batch_end;
+        ++batch_end;
+    }
+
+    scheduler.CommandBuffer().pipelineBarrier(
+        vk::PipelineStageFlagBits::eTransfer, vk::PipelineStageFlagBits::eAllCommands,
+        vk::DependencyFlagBits::eByRegion, WRITE_BARRIER, {}, {});
+
+    return true;
+}
+
+bool ExecuteShaderHLE(const Shader::Info& info, const AmdGpu::Liverpool::Regs& regs,
+                      Rasterizer& rasterizer) {
+    switch (info.pgm_hash) {
+    case COPY_SHADER_HASH:
+        return ExecuteCopyShaderHLE(info, regs, rasterizer);
+    default:
+        return false;
+    }
+}
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_shader_hle.h

@@ -0,0 +1,20 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "video_core/amdgpu/liverpool.h"
+
+namespace Shader {
+struct Info;
+}
+
+namespace Vulkan {
+
+class Rasterizer;
+
+/// Attempts to execute a shader using HLE if possible.
+bool ExecuteShaderHLE(const Shader::Info& info, const AmdGpu::Liverpool::Regs& regs,
+                      Rasterizer& rasterizer);
+
+} // namespace Vulkan

src/video_core/texture_cache/texture_cache.cpp

@@ -466,6 +466,9 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
     const auto& num_mips = image.info.resources.levels;
     ASSERT(num_mips == image.info.mips_layout.size());
 
+    const bool is_gpu_modified = True(image.flags & ImageFlagBits::GpuModified);
+    const bool is_gpu_dirty = True(image.flags & ImageFlagBits::GpuDirty);
+
     boost::container::small_vector<vk::BufferImageCopy, 14> image_copy{};
     for (u32 m = 0; m < num_mips; m++) {
         const u32 width = std::max(image.info.size.width >> m, 1u);
@@ -475,8 +478,6 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
         const auto& mip = image.info.mips_layout[m];
 
         // Protect GPU modified resources from accidental CPU reuploads.
-        const bool is_gpu_modified = True(image.flags & ImageFlagBits::GpuModified);
-        const bool is_gpu_dirty = True(image.flags & ImageFlagBits::GpuDirty);
         if (is_gpu_modified && !is_gpu_dirty) {
             const u8* addr = std::bit_cast<u8*>(image.info.guest_address);
             const u64 hash = XXH3_64bits(addr + mip.offset, mip.size);
@@ -515,7 +516,8 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
 
     const VAddr image_addr = image.info.guest_address;
     const size_t image_size = image.info.guest_size_bytes;
-    const auto [vk_buffer, buf_offset] = buffer_cache.ObtainViewBuffer(image_addr, image_size);
+    const auto [vk_buffer, buf_offset] =
+        buffer_cache.ObtainViewBuffer(image_addr, image_size, is_gpu_dirty);
     // The obtained buffer may be written by a shader so we need to emit a barrier to prevent RAW
     // hazard
     if (auto barrier = vk_buffer->GetBarrier(vk::AccessFlagBits2::eTransferRead,