Fixes and QoL (#159)

* to ensure that we're not unlocking submits too early

* a final touch

* video_core: texture_cache: fix for page table corruption

* core: linker: a name for the game main thread

* libraries: gnmdriver: an option to dump application command lists

* libraries: kernel: named guest threads

* video_core: added a heuristic for determination of CB/DB surface extents

* fix for rebase leftover

src/video_core/amdgpu/liverpool.cpp

@@ -61,7 +61,11 @@ void Liverpool::Process(std::stop_token stoken) {
                 --num_submits;
             }
         }
-        num_submits.notify_all();
+
+        if (submit_done) {
+            num_submits.notify_all();
+            submit_done = false;
+        }
     }
 }
 
@@ -163,8 +167,61 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
         }
         case PM4ItOpcode::SetContextReg: {
             const auto* set_data = reinterpret_cast<const PM4CmdSetData*>(header);
-            std::memcpy(&regs.reg_array[ContextRegWordOffset + set_data->reg_offset], header + 2,
-                        (count - 1) * sizeof(u32));
+            const auto reg_addr = ContextRegWordOffset + set_data->reg_offset;
+            const auto* payload = reinterpret_cast<const u32*>(header + 2);
+
+            std::memcpy(&regs.reg_array[reg_addr], payload, (count - 1) * sizeof(u32));
+
+            // In the case of HW, render target memory has alignment as color block operates on
+            // tiles. There is no information of actual resource extents stored in CB context
+            // regs, so any deduction of it from slices/pitch will lead to a larger surface created.
+            // The same applies to the depth targets. Fortunatelly, the guest always sends
+            // a trailing NOP packet right after the context regs setup, so we can use the heuristic
+            // below and extract the hint to determine actual resource dims.
+
+            switch (reg_addr) {
+            case ContextRegs::CbColor0Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor1Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor2Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor3Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor4Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor5Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor6Base:
+                [[fallthrough]];
+            case ContextRegs::CbColor7Base: {
+                const auto col_buf_id = (reg_addr - ContextRegs::CbColor0Base) /
+                                        (ContextRegs::CbColor1Base - ContextRegs::CbColor0Base);
+                ASSERT(col_buf_id < NumColorBuffers);
+
+                const auto nop_offset = header->type3.count;
+                if (nop_offset == 0x0e) {
+                    ASSERT_MSG(payload[nop_offset] == 0xc0001000,
+                               "NOP hint is missing in CB setup sequence");
+                    last_cb_extent[col_buf_id].raw = payload[nop_offset + 1];
+                } else {
+                    last_cb_extent[col_buf_id].raw = 0;
+                }
+                break;
+            }
+            case ContextRegs::DbZInfo: {
+                if (header->type3.count == 8) {
+                    ASSERT_MSG(payload[20] == 0xc0001000,
+                               "NOP hint is missing in DB setup sequence");
+                    last_db_extent.raw = payload[21];
+                } else {
+                    last_db_extent.raw = 0;
+                }
+                break;
+            }
+            default:
+                break;
+            }
             break;
         }
         case PM4ItOpcode::SetShReg: {

src/video_core/amdgpu/liverpool.h

@@ -682,6 +682,18 @@ struct Liverpool {
         Polygon = 21,
     };
 
+    enum ContextRegs : u32 {
+        DbZInfo = 0xA010,
+        CbColor0Base = 0xA318,
+        CbColor1Base = 0xA327,
+        CbColor2Base = 0xA336,
+        CbColor3Base = 0xA345,
+        CbColor4Base = 0xA354,
+        CbColor5Base = 0xA363,
+        CbColor6Base = 0xA372,
+        CbColor7Base = 0xA381,
+    };
+
     union Regs {
         struct {
             INSERT_PADDING_WORDS(0x2C08);
@@ -765,6 +777,21 @@ struct Liverpool {
 
     Regs regs{};
 
+    // See for a comment in context reg parsing code
+    union CbDbExtent {
+        struct {
+            u16 width;
+            u16 height;
+        };
+        u32 raw{0u};
+
+        [[nodiscard]] bool Valid() const {
+            return raw != 0;
+        }
+    };
+    std::array<CbDbExtent, NumColorBuffers> last_cb_extent{};
+    CbDbExtent last_db_extent{};
+
 public:
     Liverpool();
     ~Liverpool();
@@ -777,6 +804,11 @@ public:
         return num_submits == 0;
     }
 
+    void NotifySubmitDone() {
+        submit_done = true;
+        num_submits.notify_all();
+    }
+
     void BindRasterizer(Vulkan::Rasterizer* rasterizer_) {
         rasterizer = rasterizer_;
     }
@@ -841,6 +873,7 @@ private:
     Vulkan::Rasterizer* rasterizer{};
     std::jthread process_thread{};
     std::atomic<u32> num_submits{};
+    std::atomic<bool> submit_done{};
 };
 
 static_assert(GFX6_3D_REG_INDEX(ps_program) == 0x2C08);

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -41,11 +41,14 @@ void Rasterizer::Draw(bool is_indexed) {
 
     boost::container::static_vector<vk::RenderingAttachmentInfo, Liverpool::NumColorBuffers>
         color_attachments{};
-    for (const auto& col_buf : regs.color_buffers) {
+    for (auto col_buf_id = 0u; col_buf_id < Liverpool::NumColorBuffers; ++col_buf_id) {
+        const auto& col_buf = regs.color_buffers[col_buf_id];
         if (!col_buf) {
             continue;
         }
-        const auto& image_view = texture_cache.RenderTarget(col_buf);
+
+        const auto& hint = liverpool->last_cb_extent[col_buf_id];
+        const auto& image_view = texture_cache.RenderTarget(col_buf, hint);
 
         color_attachments.push_back({
             .imageView = *image_view.image_view,

src/video_core/texture_cache/image.cpp

@@ -82,12 +82,13 @@ ImageInfo::ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group) noe
     }
 }
 
-ImageInfo::ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer) noexcept {
+ImageInfo::ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
+                     const AmdGpu::Liverpool::CbDbExtent& hint /*= {}*/) noexcept {
     is_tiled = true;
     pixel_format = LiverpoolToVK::SurfaceFormat(buffer.info.format, buffer.NumFormat());
     type = vk::ImageType::e2D;
-    size.width = buffer.Pitch();
-    size.height = buffer.Height();
+    size.width = hint.Valid() ? hint.width : buffer.Pitch();
+    size.height = hint.Valid() ? hint.height : buffer.Height();
     size.depth = 1;
     pitch = size.width;
     guest_size_bytes = buffer.slice.tile_max * (buffer.view.slice_max + 1);

src/video_core/texture_cache/image.h

@@ -34,7 +34,8 @@ DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)
 struct ImageInfo {
     ImageInfo() = default;
     explicit ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group) noexcept;
-    explicit ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer) noexcept;
+    explicit ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
+                       const AmdGpu::Liverpool::CbDbExtent& hint = {}) noexcept;
     explicit ImageInfo(const AmdGpu::Image& image) noexcept;
 
     bool is_tiled = false;

src/video_core/texture_cache/texture_cache.cpp

@@ -101,6 +101,7 @@ TextureCache::~TextureCache() {
 }
 
 void TextureCache::OnCpuWrite(VAddr address) {
+    std::unique_lock lock{m_page_table};
     ForEachImageInRegion(address, 1 << PageShift, [&](ImageId image_id, Image& image) {
         // Ensure image is reuploaded when accessed again.
         image.flags |= ImageFlagBits::CpuModified;
@@ -110,6 +111,7 @@ void TextureCache::OnCpuWrite(VAddr address) {
 }
 
 Image& TextureCache::FindImage(const ImageInfo& info, VAddr cpu_address) {
+    std::unique_lock lock{m_page_table};
     boost::container::small_vector<ImageId, 2> image_ids;
     ForEachImageInRegion(cpu_address, info.guest_size_bytes, [&](ImageId image_id, Image& image) {
         if (image.cpu_addr == cpu_address) {
@@ -151,8 +153,9 @@ ImageView& TextureCache::FindImageView(const AmdGpu::Image& desc) {
     return slot_image_views[view_id];
 }
 
-ImageView& TextureCache::RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer) {
-    const ImageInfo info{buffer};
+ImageView& TextureCache::RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
+                                      const AmdGpu::Liverpool::CbDbExtent& hint) {
+    const ImageInfo info{buffer, hint};
     auto& image = FindImage(info, buffer.Address());
 
     ImageViewInfo view_info;

src/video_core/texture_cache/texture_cache.h

@@ -42,7 +42,8 @@ public:
     ImageView& FindImageView(const AmdGpu::Image& image);
 
     /// Retrieves the render target with specified properties
-    ImageView& RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer);
+    ImageView& RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
+                            const AmdGpu::Liverpool::CbDbExtent& hint);
 
     /// Reuploads image contents.
     void RefreshImage(Image& image);
@@ -136,6 +137,7 @@ private:
 #ifdef _WIN64
     void* veh_handle{};
 #endif
+    std::mutex m_page_table;
 };
 
 } // namespace VideoCore