Provide custom border color to samplers (#3014)

src/video_core/amdgpu/liverpool.h

@@ -608,6 +608,16 @@ struct Liverpool {
         }
     };
 
+    struct BorderColorBufferBase {
+        u32 base_addr_lo;
+        BitField<0, 8, u32> base_addr_hi;
+
+        template <typename T = VAddr>
+        T Address() const {
+            return std::bit_cast<T>(u64(base_addr_hi) << 40 | u64(base_addr_lo) << 8);
+        }
+    };
+
     struct IndexBufferBase {
         BitField<0, 8, u32> base_addr_hi;
         u32 base_addr_lo;
@@ -1299,7 +1309,9 @@ struct Liverpool {
             Scissor screen_scissor;
             INSERT_PADDING_WORDS(0xA010 - 0xA00C - 2);
             DepthBuffer depth_buffer;
-            INSERT_PADDING_WORDS(0xA080 - 0xA018);
+            INSERT_PADDING_WORDS(8);
+            BorderColorBufferBase ta_bc_base;
+            INSERT_PADDING_WORDS(0xA080 - 0xA020 - 2);
             WindowOffset window_offset;
             ViewportScissor window_scissor;
             INSERT_PADDING_WORDS(0xA08E - 0xA081 - 2);
@@ -1626,6 +1638,7 @@ static_assert(GFX6_3D_REG_INDEX(depth_htile_data_base) == 0xA005);
 static_assert(GFX6_3D_REG_INDEX(screen_scissor) == 0xA00C);
 static_assert(GFX6_3D_REG_INDEX(depth_buffer.z_info) == 0xA010);
 static_assert(GFX6_3D_REG_INDEX(depth_buffer.depth_slice) == 0xA017);
+static_assert(GFX6_3D_REG_INDEX(ta_bc_base) == 0xA020);
 static_assert(GFX6_3D_REG_INDEX(window_offset) == 0xA080);
 static_assert(GFX6_3D_REG_INDEX(window_scissor) == 0xA081);
 static_assert(GFX6_3D_REG_INDEX(color_target_mask) == 0xA08E);

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -716,7 +716,7 @@ void Rasterizer::BindTextures(const Shader::Info& stage, Shader::Backend::Bindin
                 ssharp.max_aniso.Assign(AmdGpu::AnisoRatio::One);
             }
         }
-        const auto vk_sampler = texture_cache.GetSampler(ssharp);
+        const auto vk_sampler = texture_cache.GetSampler(ssharp, liverpool->regs.ta_bc_base);
         image_infos.emplace_back(vk_sampler, VK_NULL_HANDLE, vk::ImageLayout::eGeneral);
         set_writes.push_back({
             .dstSet = VK_NULL_HANDLE,

src/video_core/texture_cache/sampler.cpp

@@ -9,7 +9,8 @@
 
 namespace VideoCore {
 
-Sampler::Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sampler) {
+Sampler::Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sampler,
+                 const AmdGpu::Liverpool::BorderColorBufferBase& border_color_base) {
     if (sampler.force_degamma) {
         LOG_WARNING(Render_Vulkan, "Texture requires gamma correction");
     }
@@ -20,7 +21,33 @@ Sampler::Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sample
     const float maxAnisotropy =
         anisotropyEnable ? std::clamp(sampler.MaxAniso(), 1.0f, instance.MaxSamplerAnisotropy())
                          : 1.0f;
+    auto borderColor = LiverpoolToVK::BorderColor(sampler.border_color_type);
+    if (!instance.IsCustomBorderColorSupported()) {
+        LOG_WARNING(Render_Vulkan, "Custom border color is not supported, falling back to black");
+        borderColor = vk::BorderColor::eFloatOpaqueBlack;
+    }
+
+    const auto customColor = [&]() -> std::optional<vk::SamplerCustomBorderColorCreateInfoEXT> {
+        if (borderColor == vk::BorderColor::eFloatCustomEXT) {
+            const auto borderColorIndex = sampler.border_color_ptr.Value();
+            const auto borderColorBuffer = border_color_base.Address<std::array<float, 4>*>();
+            const auto customBorderColorArray = borderColorBuffer[borderColorIndex];
+
+            const vk::SamplerCustomBorderColorCreateInfoEXT ret{
+                .customBorderColor =
+                    vk::ClearColorValue{
+                        .float32 = customBorderColorArray,
+                    },
+                .format = vk::Format::eR32G32B32A32Sfloat,
+            };
+            return ret;
+        } else {
+            return std::nullopt;
+        }
+    }();
+
     const vk::SamplerCreateInfo sampler_ci = {
+        .pNext = customColor ? &*customColor : nullptr,
         .magFilter = LiverpoolToVK::Filter(sampler.xy_mag_filter),
         .minFilter = LiverpoolToVK::Filter(sampler.xy_min_filter),
         .mipmapMode = LiverpoolToVK::MipFilter(sampler.mip_filter),
@@ -34,7 +61,7 @@ Sampler::Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sample
         .compareOp = LiverpoolToVK::DepthCompare(sampler.depth_compare_func),
         .minLod = sampler.MinLod(),
         .maxLod = sampler.MaxLod(),
-        .borderColor = LiverpoolToVK::BorderColor(sampler.border_color_type),
+        .borderColor = borderColor,
         .unnormalizedCoordinates = false, // Handled in shader due to Vulkan limitations.
     };
     auto [sampler_result, smplr] = instance.GetDevice().createSamplerUnique(sampler_ci);

src/video_core/texture_cache/sampler.h

@@ -14,7 +14,8 @@ namespace VideoCore {
 
 class Sampler {
 public:
-    explicit Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sampler);
+    explicit Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sampler,
+                     const AmdGpu::Liverpool::BorderColorBufferBase& border_color_base);
     ~Sampler();
 
     Sampler(const Sampler&) = delete;

src/video_core/texture_cache/texture_cache.cpp

@@ -636,9 +636,11 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
     image.flags &= ~ImageFlagBits::Dirty;
 }
 
-vk::Sampler TextureCache::GetSampler(const AmdGpu::Sampler& sampler) {
+vk::Sampler TextureCache::GetSampler(
+    const AmdGpu::Sampler& sampler,
+    const AmdGpu::Liverpool::BorderColorBufferBase& border_color_base) {
     const u64 hash = XXH3_64bits(&sampler, sizeof(sampler));
-    const auto [it, new_sampler] = samplers.try_emplace(hash, instance, sampler);
+    const auto [it, new_sampler] = samplers.try_emplace(hash, instance, sampler, border_color_base);
     return it->second.Handle();
 }
 

src/video_core/texture_cache/texture_cache.h

@@ -139,7 +139,9 @@ public:
     void RefreshImage(Image& image, Vulkan::Scheduler* custom_scheduler = nullptr);
 
     /// Retrieves the sampler that matches the provided S# descriptor.
-    [[nodiscard]] vk::Sampler GetSampler(const AmdGpu::Sampler& sampler);
+    [[nodiscard]] vk::Sampler GetSampler(
+        const AmdGpu::Sampler& sampler,
+        const AmdGpu::Liverpool::BorderColorBufferBase& border_color_base);
 
     /// Retrieves the image with the specified id.
     [[nodiscard]] Image& GetImage(ImageId id) {