graphics: represent rebased contiguous mip-tail texture views

Co-authored-by: fxpw <fxpw@outlook.com>

src/graphics/host_gpu/renderer/image/imageInfo.cpp

@@ -21,6 +21,54 @@ bool IsDepthPlane(const ImageInfo& info) {
 	        info.tile_mode == Prospero::TileMode::kLinear);
 }
 
+GuestRange ContiguousMipRange(const ImageInfo& info, uint32_t level) {
+	const auto& block = info.tiling.block;
+	if (info.tile_mode != Prospero::TileMode::kStandard4KB ||
+	    block.family != TileBlockFamily::Standard4KB || info.IsVolume() || info.samples != 1 ||
+	    info.HasMetadata() || level >= info.resources.levels || level >= info.mip_layout.size() ||
+	    info.tiling.texel_width == 0 || info.tiling.texel_height == 0) {
+		return {};
+	}
+	const auto extent = info.MipExtent(level);
+	const auto width  = (extent.width + info.tiling.texel_width - 1u) / info.tiling.texel_width;
+	const auto height = (extent.height + info.tiling.texel_height - 1u) / info.tiling.texel_height;
+	const auto& mip   = info.mip_layout[level];
+	uint32_t last = 0, origin = 0;
+	if (!std::has_single_bit(width) || !std::has_single_bit(height) ||
+	    width > block.block_width || height > block.block_height ||
+	    mip.tail_x > block.block_width - width || mip.tail_y > block.block_height - height ||
+	    mip.tail_x % width != 0 || mip.tail_y % height != 0 ||
+	    !TileGetBlockOffset(block, width - 1u, height - 1u, 0, last) ||
+	    !TileGetBlockOffset(block, mip.tail_x, mip.tail_y, 0, origin)) {
+		return {};
+	}
+	// Standard4KB interleaves monotonically ordered X/Y bits. A dense rectangle
+	// fills exactly this power-of-two interval, including when rebased into a tail.
+	const uint64_t size = uint64_t {last} + block.bytes_per_element;
+	if (!std::has_single_bit(size) || size != uint64_t {width} * height * block.bytes_per_element ||
+	    (origin & (size - 1u)) != 0 || size > block.block_size - origin ||
+	    mip.offset > UINT64_MAX - origin) {
+		return {};
+	}
+	return {mip.offset + origin, size};
+}
+
+uint32_t CompactMipSize(const ImageInfo& info) {
+	if (info.resources.levels != 1 || info.resources.layers != 1) {
+		return 0;
+	}
+	const auto range = ContiguousMipRange(info, 0);
+	if (range.size == 0 || range.address != 0) {
+		return 0;
+	}
+	const auto alignment = std::max<uint64_t>(256, range.size);
+	const auto in_block  = info.data.address & (info.tiling.block.block_size - 1u);
+	return (info.data.address & (alignment - 1u)) == 0 &&
+	               range.size <= info.tiling.block.block_size - in_block
+	           ? static_cast<uint32_t>(range.size)
+	           : 0;
+}
+
 } // namespace
 
 vk::Extent3D ImageInfo::MipExtent(uint32_t level) const {
@@ -34,6 +82,7 @@ void ImageInfo::UpdateSize() {
 	        extent.height == 0 || extent.depth == 0 || resources.levels == 0 ||
 	        resources.levels > mip_layout.size() || resources.layers == 0 || samples == 0 ||
 	        !std::has_single_bit(samples));
+	const bool auto_size = data.size == 0;
 	TileTextureElementLayout element {};
 	EXIT_NOT_IMPLEMENTED(!TileGetTextureElementLayout(guest_format, element));
 	if (bytes_per_block == 0) {
@@ -164,9 +213,86 @@ void ImageInfo::UpdateSize() {
 			linear_offset     = Add(mip.linear_offset, Multiply(mip.linear_size, active.depth));
 			mip_pitch         = std::max(mip_pitch >> 1u, 1u);
 		}
+		const auto compact_size = CompactMipSize(*this);
+		if (compact_size != 0 && (auto_size || data.size == compact_size)) {
+			data.size = mip_layout[0].size = compact_size;
+		}
 	}
 }
 
+uint32_t ImageInfo::DataAlignment() const {
+	if (!IsTiled()) {
+		return 256;
+	}
+	const auto compact_size = CompactMipSize(*this);
+	if (compact_size != 0 && data.size == compact_size && mip_layout[0].size == compact_size) {
+		return std::max(256u, compact_size);
+	}
+	return tiling.block.block_size;
+}
+
+int32_t ImageInfo::SliceOf(const ImageInfo& container, int32_t mip) const noexcept {
+	if (!IsCompatible(container) || tile_mode != container.tile_mode || type != container.type ||
+	    IsVolume() || resources.levels != 1 || resources.layers == 0 ||
+	    container.resources.layers == 0 || mip < 0 ||
+	    static_cast<uint32_t>(mip) >= container.resources.levels ||
+	    container.resources.levels > container.mip_layout.size() ||
+	    !data.Valid() || !container.data.Valid() || data.address < container.data.address ||
+	    data.End() > container.data.End()) {
+		return -1;
+	}
+	if (HasStencil() != container.HasStencil() ||
+	    (HasStencil() && (stencil.address < container.stencil.address ||
+	                      stencil.End() > container.stencil.End()))) {
+		return -1;
+	}
+
+	const auto  level  = static_cast<uint32_t>(mip);
+	const auto& layout = container.mip_layout[level];
+	if (extent.width != std::max(container.extent.width >> level, 1u) ||
+	    extent.height != std::max(container.extent.height >> level, 1u) ||
+	    extent.depth != container.extent.depth || mip_layout[0].pitch != layout.pitch ||
+	    mip_layout[0].height != layout.height || layout.size == 0 ||
+	    layout.size % container.resources.layers != 0 ||
+	    container.data.size % container.resources.layers != 0 ||
+	    data.size % resources.layers != 0) {
+		return -1;
+	}
+
+	// PS5 block slices contain the entire reversed mip chain. A rebased tail
+	// descriptor covers only its active bytes within the shared tail block.
+	const auto slice_stride = container.data.size / container.resources.layers;
+	const auto child_stride = data.size / resources.layers;
+	auto mip_offset = layout.offset;
+	auto mip_size   = layout.size / container.resources.layers;
+	const auto compact_size = CompactMipSize(*this);
+	if (compact_size != 0 && child_stride == compact_size && mip_layout[0].size == compact_size) {
+		const auto range = ContiguousMipRange(container, level);
+		if (range.size != compact_size) {
+			return -1;
+		}
+		mip_offset = range.address;
+		mip_size   = range.size;
+	} else if (mip_layout[0].tail_x != layout.tail_x || mip_layout[0].tail_y != layout.tail_y) {
+		return -1;
+	}
+	if (child_stride != mip_size || mip_offset >= slice_stride ||
+	    child_stride > slice_stride - mip_offset ||
+	    (resources.layers > 1 && child_stride != slice_stride)) {
+		return -1;
+	}
+	const auto address_delta = data.address - container.data.address;
+	if (address_delta < mip_offset || (address_delta - mip_offset) % slice_stride != 0) {
+		return -1;
+	}
+	const auto layer = (address_delta - mip_offset) / slice_stride;
+	if (layer > INT32_MAX || layer >= container.resources.layers ||
+	    resources.layers > container.resources.layers - layer) {
+		return -1;
+	}
+	return static_cast<int32_t>(layer);
+}
+
 uint64_t ImageInfo::LinearSize(uint32_t levels) const {
 	levels = levels == 0 ? resources.levels : levels;
 	EXIT_IF(levels == 0 || levels > resources.levels || levels > mip_layout.size());

src/graphics/host_gpu/renderer/image/imageInfo.h

@@ -88,6 +88,7 @@ struct ImageInfo {
 	uint32_t first_tail_level    = 16;
 
 	void                                           UpdateSize();
+	[[nodiscard]] uint32_t                         DataAlignment() const;
 	[[nodiscard]] uint64_t                         LinearSize(uint32_t levels = 0) const;
 	[[nodiscard]] vk::Extent3D                     MipExtent(uint32_t level) const;
 	[[nodiscard]] std::vector<vk::BufferImageCopy> BufferCopies(uint64_t buffer_offset = 0,
@@ -136,53 +137,7 @@ struct ImageInfo {
 		}
 		return -1;
 	}
-	[[nodiscard]] int32_t SliceOf(const ImageInfo& container, int32_t mip) const noexcept {
+	[[nodiscard]] int32_t SliceOf(const ImageInfo& container, int32_t mip) const noexcept;
-		if (!IsCompatible(container) || tile_mode != container.tile_mode || type != container.type ||
-		    IsVolume() || resources.levels != 1 || resources.layers == 0 ||
-		    container.resources.layers == 0 || mip < 0 ||
-		    static_cast<uint32_t>(mip) >= container.resources.levels ||
-		    container.resources.levels > container.mip_layout.size() ||
-		    !data.Valid() || !container.data.Valid() || data.address < container.data.address ||
-		    data.End() > container.data.End()) {
-			return -1;
-		}
-		if (HasStencil() != container.HasStencil() ||
-		    (HasStencil() && (stencil.address < container.stencil.address ||
-		                      stencil.End() > container.stencil.End()))) {
-			return -1;
-		}
-
-		const auto  level  = static_cast<uint32_t>(mip);
-		const auto& layout = container.mip_layout[level];
-		if (extent.width != std::max(container.extent.width >> level, 1u) ||
-		    extent.height != std::max(container.extent.height >> level, 1u) ||
-		    extent.depth != container.extent.depth || mip_layout[0].pitch != layout.pitch ||
-		    mip_layout[0].height != layout.height || layout.size == 0 ||
-		    layout.size % container.resources.layers != 0 ||
-		    container.data.size % container.resources.layers != 0 ||
-		    data.size % resources.layers != 0) {
-			return -1;
-		}
-
-		// PS5 block slices contain the entire reversed mip chain.
-		const auto slice_stride = container.data.size / container.resources.layers;
-		const auto child_stride = data.size / resources.layers;
-		if (child_stride != layout.size / container.resources.layers ||
-		    layout.offset >= slice_stride || child_stride > slice_stride - layout.offset ||
-		    (resources.layers > 1 && child_stride != slice_stride)) {
-			return -1;
-		}
-		const auto address_delta = data.address - container.data.address;
-		if (address_delta < layout.offset || (address_delta - layout.offset) % slice_stride != 0) {
-			return -1;
-		}
-		const auto layer = (address_delta - layout.offset) / slice_stride;
-		if (layer > INT32_MAX || layer >= container.resources.layers ||
-		    resources.layers > container.resources.layers - layer) {
-			return -1;
-		}
-		return static_cast<int32_t>(layer);
-	}
 };
 
 struct ImageViewInfo {

src/graphics/host_gpu/renderer/pipeline/descriptors.cpp

@@ -608,6 +608,9 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	    block_bytes != 0 ? block_bytes : Prospero::NumBytesPerElement(format);
 	desc.info.samples   = samples;
 	desc.info.tile_mode = tile;
+	if (!resource.r128 && descriptor.MetaCompress() && tile != Prospero::TileMode::kDepth) {
+		desc.info.metadata.kind = ImageMetadataKind::Dcc;
+	}
 	if (levels > physical_levels) {
 		if (!ResolveTextureMipView(desc.info, physical_levels,
 		                           !resource.r128 && descriptor.MetaCompress(), view_levels,
@@ -620,8 +623,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	} else {
 		desc.info.UpdateSize();
 	}
-	const auto alignment =
+	const auto alignment = desc.info.DataAlignment();
-	    tile == Prospero::TileMode::kLinear ? 256u : desc.info.tiling.block.block_size;
 	EXIT_NOT_IMPLEMENTED(desc.info.data.size == 0 || alignment == 0 ||
 	                     (address & (static_cast<uint64_t>(alignment) - 1u)) != 0);
 	if (storage) {
@@ -632,7 +634,6 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 		(void)TileGetDccSize(width, height, volume ? depth : image_layers,
 		                     desc.info.bytes_per_block, physical_levels, tile, metadata_size,
 		                     std::countr_zero(samples));
-		desc.info.metadata.kind          = ImageMetadataKind::Dcc;
 		desc.info.metadata.range         = {descriptor.MetaAddr() << 8u, metadata_size.size};
 		desc.info.metadata.dcc_alpha_msb = descriptor.DccAlphaPos();
 	}

tests/ShaderRecompilerComputeTests.cpp

@@ -18673,6 +18673,36 @@ public:
       check_round_trip("odd multi-mip", total.size, layout);
     }
 
+    {
+      constexpr auto format = Prospero::BufferFormat::k16_16_16_16Float;
+      constexpr auto tile = Prospero::TileMode::kStandard4KB;
+      auto view = MakeTilingImage(format, 16, 16, 1, 1, tile, 0, false);
+      view.data.address = 0x5032b23800;
+      view.UpdateSize();
+      Require(name, "rebased mip footprint",
+              view.data.size == 0x800 && view.DataAlignment() == 0x800 &&
+                  view.mip_layout[0].pitch == 32 &&
+                  view.mip_layout[0].height == 16 &&
+                  view.tiled_slice_stride == 0x1000,
+              "a compact view changed its allocation geometry");
+      check_round_trip("rebased mip exact range", view.data.size, view);
+      check_round_trip("rebased mip adjacent canary", 0x1000, view);
+
+      std::vector<uint8_t> linear(view.LinearSize());
+      fill(&linear, ++case_index);
+      std::vector<uint8_t> expected(0x1000, 0xab), observed;
+      convert_reference(true, &expected, linear, view);
+      gpu_tile(linear, &observed, 0x1000, linear.size(), view, nullptr, 0,
+               0x100, 0x800);
+      compare("rebased mip prefix and neighboring mip", expected, observed);
+      std::vector<uint8_t> round_trip;
+      gpu_detile(observed, &round_trip, observed.size(), linear.size(), view,
+                 nullptr, 0x800);
+      std::vector<uint8_t> expected_linear(linear.size(), 0);
+      convert_reference(false, &expected_linear, expected, view);
+      compare("rebased mip nonzero source", expected_linear, round_trip);
+    }
+
     struct TailMode {
       Prospero::TileMode tile;
       TileBlockFamily family;

tests/shaderCfgTests.cpp

@@ -1296,6 +1296,100 @@ void TestNormalizedImageContracts() {
         "Vulkan image-view compatibility classes diverged from production");
 }
 
+void TestRebasedTextureMipViews() {
+  constexpr auto format = Prospero::BufferFormat::k16_16_16_16Float;
+  constexpr auto tile = Prospero::TileMode::kStandard4KB;
+  TileSizeAlign allocation{};
+  TileGetTextureSize(format, 16, 16, 1, tile, &allocation, nullptr, nullptr);
+  Check(allocation.size == 0x1000 && allocation.align == 0x1000,
+        "rebased texture views changed guest allocation size/alignment");
+
+  ImageInfo chain{};
+  chain.data.address = 0x5032b23000;
+  chain.pixel_format = vk::Format::eR16G16B16A16Sfloat;
+  chain.guest_format = format;
+  chain.extent = {256, 256, 1};
+  chain.resources = {9, 2};
+  chain.tile_mode = tile;
+  chain.UpdateSize();
+  Check(chain.tiled_slice_stride == 0xab000 && chain.data.size == 0x156000 &&
+            chain.first_tail_level == 4 && chain.DataAlignment() == 0x1000,
+        "full texture chain lost its guest allocation layout");
+  const std::array<uint64_t, 9> offsets{
+      0x2b000, 0xb000, 0x3000, 0x1000, 0x800, 0x600, 0x500, 0x400, 0x300};
+  const auto make_view = [&](uint32_t level, uint32_t layer) {
+    auto view = chain;
+    view.data = {chain.data.address + layer * chain.tiled_slice_stride +
+                     offsets[level], 0};
+    view.extent = {256u >> level, 256u >> level, 1};
+    view.resources = {1, 1};
+    view.pitch = 0;
+    view.UpdateSize();
+    return view;
+  };
+  for (uint32_t layer = 0; layer < 2; ++layer) {
+    for (uint32_t level = 0; level < offsets.size(); ++level) {
+      auto view = make_view(level, layer);
+      const uint64_t bytes = uint64_t{view.extent.width} * view.extent.height * 8;
+      Check(view.data.size == bytes && view.mip_layout[0].size == bytes &&
+                view.tiled_slice_stride == std::max<uint64_t>(0x1000, bytes) &&
+                view.DataAlignment() == std::min<uint64_t>(0x1000,
+                                               std::max<uint64_t>(0x100, bytes)) &&
+                view.MipOf(chain) == static_cast<int32_t>(level) &&
+                view.SliceOf(chain, static_cast<int32_t>(level)) ==
+                    static_cast<int32_t>(layer),
+            "rebased mip footprint/origin or parent layer stride was incorrect");
+      view.UpdateSize();
+      Check(view.data.size == bytes && view.mip_layout[0].size == bytes &&
+                view.MipOf(chain) == static_cast<int32_t>(level),
+            "resizing a normalized view lost its compact footprint");
+    }
+  }
+  auto view = make_view(4, 0);
+  Check(view.data.address == 0x5032b23800 && view.mip_layout[0].pitch == 32 &&
+            view.mip_layout[0].height == 16 &&
+            view.mip_layout[0].linear_pitch == 32 && view.LinearSize() == 0x1000,
+        "captured 16x16 view changed its padded tiling or linear transfer size");
+  view.data.address += 0x800;
+  Check(view.MipOf(chain) == -1, "tail view accepted the following mip's address");
+  view = make_view(4, 0);
+  view.mip_layout[0].pitch /= 2;
+  Check(view.MipOf(chain) == -1, "tail view accepted a different padded pitch");
+  view = make_view(4, 0);
+  view.resources.layers = 2;
+  view.data.size *= 2;
+  view.mip_layout[0].size *= 2;
+  Check(view.MipOf(chain) == -1, "tail array view ignored the parent chain stride");
+  view = make_view(4, 0);
+  view.data = {chain.data.address, 0x1000};
+  view.UpdateSize();
+  Check(view.MipOf(chain) == -1,
+        "full-block child ignored the parent mip's nonzero tail origin");
+
+  for (uint32_t variant = 0; variant < 7; ++variant) {
+    view = make_view(4, 0);
+    view.data = {chain.data.address, 0};
+    view.pitch = 0;
+    switch (variant) {
+    case 0: view.extent = {8, 16, 1}; break; // Holes in the Standard4KB span.
+    case 1: view.resources.levels = 2; break;
+    case 2: view.resources.layers = 2; break;
+    case 3: view.type = Prospero::ImageType::kColor3D; break;
+    case 4: view.tile_mode = Prospero::TileMode::kStandard64KB; break;
+    case 5: view.metadata.kind = ImageMetadataKind::Dcc; break;
+    case 6: view.data = {0x5032b23800, 0x1000}; break;
+    }
+    view.UpdateSize();
+    Check(view.data.size >= 0x1000 && view.DataAlignment() >= 0x1000,
+          "unsupported or explicitly sized texture became a compact tail view");
+  }
+  view = make_view(4, 0);
+  view.data = {0x5032b23900, 0};
+  view.UpdateSize();
+  Check(view.data.size == 0x1000 && view.DataAlignment() == 0x1000,
+        "misaligned descriptor received compact tail alignment");
+}
+
 void TestShaderFeatureCollection() {
   using namespace ShaderRecompiler::IR;
 
@@ -15725,6 +15819,7 @@ int main() {
   TestShaderBufferResourceSize();
   TestNativeShaderResourceDependencies();
   TestNormalizedImageContracts();
+  TestRebasedTextureMipViews();
   TestShaderFeatureCollection();
   TestTypedSpirvSerialization();
   TestDeferredSpirvPhiPatching();