astc_decoder: Reduce workgroup size

This reduces the amount of over dispatching when there are odd dimensions (i.e. ASTC 8x5), which rarely evenly divide into 32x32.

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp

@@ -166,6 +166,8 @@ NvResult nvhost_nvdec_common::MapBuffer(const std::vector<u8>& input, std::vecto
             LOG_ERROR(Service_NVDRV, "failed to map size={}", object->size);
         } else {
             cmd_buffer.map_address = object->dma_map_addr;
+            AddBufferMap(object->dma_map_addr, object->size, object->addr,
+                         object->status == nvmap::Object::Status::Allocated);
         }
     }
     std::memcpy(output.data(), &params, sizeof(IoctlMapBuffer));
@@ -176,11 +178,30 @@ NvResult nvhost_nvdec_common::MapBuffer(const std::vector<u8>& input, std::vecto
 }
 
 NvResult nvhost_nvdec_common::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& output) {
-    // This is intntionally stubbed.
-    // Skip unmapping buffers here, as to not break the continuity of the VP9 reference frame
-    // addresses, and risk invalidating data before the async GPU thread is done with it
+    IoctlMapBuffer params{};
+    std::memcpy(&params, input.data(), sizeof(IoctlMapBuffer));
+    std::vector<MapBufferEntry> cmd_buffer_handles(params.num_entries);
+    SliceVectors(input, cmd_buffer_handles, params.num_entries, sizeof(IoctlMapBuffer));
+
+    auto& gpu = system.GPU();
+
+    for (auto& cmd_buffer : cmd_buffer_handles) {
+        const auto object{nvmap_dev->GetObject(cmd_buffer.map_handle)};
+        if (!object) {
+            LOG_ERROR(Service_NVDRV, "invalid cmd_buffer nvmap_handle={:X}", cmd_buffer.map_handle);
+            std::memcpy(output.data(), &params, output.size());
+            return NvResult::InvalidState;
+        }
+        if (const auto size{RemoveBufferMap(object->dma_map_addr)}; size) {
+            gpu.MemoryManager().Unmap(object->dma_map_addr, *size);
+        } else {
+            // This occurs quite frequently, however does not seem to impact functionality
+            LOG_DEBUG(Service_NVDRV, "invalid offset=0x{:X} dma=0x{:X}", object->addr,
+                      object->dma_map_addr);
+        }
+        object->dma_map_addr = 0;
+    }
     std::memset(output.data(), 0, output.size());
-    LOG_DEBUG(Service_NVDRV, "(STUBBED) called");
     return NvResult::Success;
 }
 
@@ -191,4 +212,33 @@ NvResult nvhost_nvdec_common::SetSubmitTimeout(const std::vector<u8>& input,
     return NvResult::Success;
 }
 
+std::optional<nvhost_nvdec_common::BufferMap> nvhost_nvdec_common::FindBufferMap(
+    GPUVAddr gpu_addr) const {
+    const auto it = std::find_if(
+        buffer_mappings.begin(), buffer_mappings.upper_bound(gpu_addr), [&](const auto& entry) {
+            return (gpu_addr >= entry.second.StartAddr() && gpu_addr < entry.second.EndAddr());
+        });
+
+    ASSERT(it != buffer_mappings.end());
+    return it->second;
+}
+
+void nvhost_nvdec_common::AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr,
+                                       bool is_allocated) {
+    buffer_mappings.insert_or_assign(gpu_addr, BufferMap{gpu_addr, size, cpu_addr, is_allocated});
+}
+
+std::optional<std::size_t> nvhost_nvdec_common::RemoveBufferMap(GPUVAddr gpu_addr) {
+    const auto iter{buffer_mappings.find(gpu_addr)};
+    if (iter == buffer_mappings.end()) {
+        return std::nullopt;
+    }
+    std::size_t size = 0;
+    if (iter->second.IsAllocated()) {
+        size = iter->second.Size();
+    }
+    buffer_mappings.erase(iter);
+    return size;
+}
+
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h

@@ -23,6 +23,45 @@ public:
     ~nvhost_nvdec_common() override;
 
 protected:
+    class BufferMap final {
+    public:
+        constexpr BufferMap() = default;
+
+        constexpr BufferMap(GPUVAddr start_addr_, std::size_t size_)
+            : start_addr{start_addr_}, end_addr{start_addr_ + size_} {}
+
+        constexpr BufferMap(GPUVAddr start_addr_, std::size_t size_, VAddr cpu_addr_,
+                            bool is_allocated_)
+            : start_addr{start_addr_}, end_addr{start_addr_ + size_}, cpu_addr{cpu_addr_},
+              is_allocated{is_allocated_} {}
+
+        constexpr VAddr StartAddr() const {
+            return start_addr;
+        }
+
+        constexpr VAddr EndAddr() const {
+            return end_addr;
+        }
+
+        constexpr std::size_t Size() const {
+            return end_addr - start_addr;
+        }
+
+        constexpr VAddr CpuAddr() const {
+            return cpu_addr;
+        }
+
+        constexpr bool IsAllocated() const {
+            return is_allocated;
+        }
+
+    private:
+        GPUVAddr start_addr{};
+        GPUVAddr end_addr{};
+        VAddr cpu_addr{};
+        bool is_allocated{};
+    };
+
     struct IoctlSetNvmapFD {
         s32_le nvmap_fd{};
     };
@@ -115,11 +154,17 @@ protected:
     NvResult UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& output);
     NvResult SetSubmitTimeout(const std::vector<u8>& input, std::vector<u8>& output);
 
+    std::optional<BufferMap> FindBufferMap(GPUVAddr gpu_addr) const;
+    void AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr, bool is_allocated);
+    std::optional<std::size_t> RemoveBufferMap(GPUVAddr gpu_addr);
+
     s32_le nvmap_fd{};
     u32_le submit_timeout{};
     std::shared_ptr<nvmap> nvmap_dev;
     SyncpointManager& syncpoint_manager;
     std::array<u32, MaxSyncPoints> device_syncpoints{};
+    // This is expected to be ordered, therefore we must use a map, not unordered_map
+    std::map<GPUVAddr, BufferMap> buffer_mappings;
 };
 }; // namespace Devices
 } // namespace Service::Nvidia

src/video_core/command_classes/codecs/vp9.cpp

@@ -11,9 +11,6 @@
 
 namespace Tegra::Decoder {
 namespace {
-constexpr u32 diff_update_probability = 252;
-constexpr u32 frame_sync_code = 0x498342;
-
 // Default compressed header probabilities once frame context resets
 constexpr Vp9EntropyProbs default_probs{
     .y_mode_prob{
@@ -364,7 +361,8 @@ Vp9PictureInfo VP9::GetVp9PictureInfo(const NvdecCommon::NvdecRegisters& state)
     InsertEntropy(state.vp9_entropy_probs_offset, vp9_info.entropy);
 
     // surface_luma_offset[0:3] contains the address of the reference frame offsets in the following
-    // order: last, golden, altref, current.
+    // order: last, golden, altref, current. It may be worthwhile to track the updates done here
+    // to avoid buffering frame data needed for reference frame updating in the header composition.
     std::copy(state.surface_luma_offset.begin(), state.surface_luma_offset.begin() + 4,
               vp9_info.frame_offsets.begin());
 
@@ -386,18 +384,33 @@ Vp9FrameContainer VP9::GetCurrentFrame(const NvdecCommon::NvdecRegisters& state)
         gpu.MemoryManager().ReadBlock(state.frame_bitstream_offset, current_frame.bit_stream.data(),
                                       current_frame.info.bitstream_size);
     }
-    if (!next_frame.bit_stream.empty()) {
+    // Buffer two frames, saving the last show frame info
+    if (!next_next_frame.bit_stream.empty()) {
         Vp9FrameContainer temp{
             .info = current_frame.info,
             .bit_stream = std::move(current_frame.bit_stream),
         };
-        next_frame.info.show_frame = current_frame.info.last_frame_shown;
-        current_frame.info = next_frame.info;
-        current_frame.bit_stream = std::move(next_frame.bit_stream);
-        next_frame = std::move(temp);
+        next_next_frame.info.show_frame = current_frame.info.last_frame_shown;
+        current_frame.info = next_next_frame.info;
+        current_frame.bit_stream = std::move(next_next_frame.bit_stream);
+        next_next_frame = std::move(temp);
+
+        if (!next_frame.bit_stream.empty()) {
+            Vp9FrameContainer temp2{
+                .info = current_frame.info,
+                .bit_stream = std::move(current_frame.bit_stream),
+            };
+            next_frame.info.show_frame = current_frame.info.last_frame_shown;
+            current_frame.info = next_frame.info;
+            current_frame.bit_stream = std::move(next_frame.bit_stream);
+            next_frame = std::move(temp2);
+        } else {
+            next_frame.info = current_frame.info;
+            next_frame.bit_stream = std::move(current_frame.bit_stream);
+        }
     } else {
-        next_frame.info = current_frame.info;
-        next_frame.bit_stream = std::move(current_frame.bit_stream);
+        next_next_frame.info = current_frame.info;
+        next_next_frame.bit_stream = std::move(current_frame.bit_stream);
     }
     return current_frame;
 }
@@ -600,64 +613,86 @@ VpxBitStreamWriter VP9::ComposeUncompressedHeader() {
 
         // Reset context
         prev_frame_probs = default_probs;
-        swap_ref_indices = false;
+        swap_next_golden = false;
         loop_filter_ref_deltas.fill(0);
         loop_filter_mode_deltas.fill(0);
-        frame_ctxs.fill(default_probs);
+
+        // allow frames offsets to stabilize before checking for golden frames
+        grace_period = 4;
+
+        // On key frames, all frame slots are set to the current frame,
+        // so the value of the selected slot doesn't really matter.
+        frame_ctxs.fill({current_frame_number, false, default_probs});
 
         // intra only, meaning the frame can be recreated with no other references
         current_frame_info.intra_only = true;
+
     } else {
+
         if (!current_frame_info.show_frame) {
             uncomp_writer.WriteBit(current_frame_info.intra_only);
+            if (!current_frame_info.last_frame_was_key) {
+                swap_next_golden = !swap_next_golden;
+            }
         } else {
             current_frame_info.intra_only = false;
         }
         if (!current_frame_info.error_resilient_mode) {
             uncomp_writer.WriteU(0, 2); // Reset frame context.
         }
-        const auto& curr_offsets = current_frame_info.frame_offsets;
-        const auto& next_offsets = next_frame.info.frame_offsets;
-        const bool ref_frames_different = curr_offsets[1] != curr_offsets[2];
-        const bool next_references_swap =
-            (next_offsets[1] == curr_offsets[2]) || (next_offsets[2] == curr_offsets[1]);
-        const bool needs_ref_swap = ref_frames_different && next_references_swap;
-        if (needs_ref_swap) {
-            swap_ref_indices = !swap_ref_indices;
-        }
-        union {
-            u32 raw;
-            BitField<0, 1, u32> refresh_last;
-            BitField<1, 2, u32> refresh_golden;
-            BitField<2, 1, u32> refresh_alt;
-        } refresh_frame_flags;
 
-        refresh_frame_flags.raw = 0;
-        for (u32 index = 0; index < 3; ++index) {
-            // Refresh indices that use the current frame as an index
-            if (curr_offsets[3] == next_offsets[index]) {
-                refresh_frame_flags.raw |= 1u << index;
+        // Last, Golden, Altref frames
+        std::array<s32, 3> ref_frame_index{0, 1, 2};
+
+        // Set when next frame is hidden
+        // altref and golden references are swapped
+        if (swap_next_golden) {
+            ref_frame_index = std::array<s32, 3>{0, 2, 1};
+        }
+
+        // update Last Frame
+        u64 refresh_frame_flags = 1;
+
+        // golden frame may refresh, determined if the next golden frame offset is changed
+        bool golden_refresh = false;
+        if (grace_period <= 0) {
+            for (s32 index = 1; index < 3; ++index) {
+                if (current_frame_info.frame_offsets[index] !=
+                    next_frame.info.frame_offsets[index]) {
+                    current_frame_info.refresh_frame[index] = true;
+                    golden_refresh = true;
+                    grace_period = 3;
+                }
             }
         }
-        if (swap_ref_indices) {
-            const u32 temp = refresh_frame_flags.refresh_golden;
-            refresh_frame_flags.refresh_golden.Assign(refresh_frame_flags.refresh_alt.Value());
-            refresh_frame_flags.refresh_alt.Assign(temp);
+
+        if (current_frame_info.show_frame &&
+            (!next_frame.info.show_frame || next_frame.info.is_key_frame)) {
+            // Update golden frame
+            refresh_frame_flags = swap_next_golden ? 2 : 4;
         }
+
+        if (!current_frame_info.show_frame) {
+            // Update altref
+            refresh_frame_flags = swap_next_golden ? 2 : 4;
+        } else if (golden_refresh) {
+            refresh_frame_flags = 3;
+        }
+
         if (current_frame_info.intra_only) {
             uncomp_writer.WriteU(frame_sync_code, 24);
-            uncomp_writer.WriteU(refresh_frame_flags.raw, 8);
+            uncomp_writer.WriteU(static_cast<s32>(refresh_frame_flags), 8);
             uncomp_writer.WriteU(current_frame_info.frame_size.width - 1, 16);
             uncomp_writer.WriteU(current_frame_info.frame_size.height - 1, 16);
             uncomp_writer.WriteBit(false); // Render and frame size different.
         } else {
-            const bool swap_indices = needs_ref_swap ^ swap_ref_indices;
-            const auto ref_frame_index = swap_indices ? std::array{0, 2, 1} : std::array{0, 1, 2};
-            uncomp_writer.WriteU(refresh_frame_flags.raw, 8);
-            for (size_t index = 1; index < 4; index++) {
+            uncomp_writer.WriteU(static_cast<s32>(refresh_frame_flags), 8);
+
+            for (s32 index = 1; index < 4; index++) {
                 uncomp_writer.WriteU(ref_frame_index[index - 1], 3);
                 uncomp_writer.WriteU(current_frame_info.ref_frame_sign_bias[index], 1);
             }
+
             uncomp_writer.WriteBit(true);  // Frame size with refs.
             uncomp_writer.WriteBit(false); // Render and frame size different.
             uncomp_writer.WriteBit(current_frame_info.allow_high_precision_mv);
@@ -679,9 +714,10 @@ VpxBitStreamWriter VP9::ComposeUncompressedHeader() {
         frame_ctx_idx = 1;
     }
 
-    uncomp_writer.WriteU(frame_ctx_idx, 2);       // Frame context index.
-    prev_frame_probs = frame_ctxs[frame_ctx_idx]; // reference probabilities for compressed header
-    frame_ctxs[frame_ctx_idx] = current_frame_info.entropy;
+    uncomp_writer.WriteU(frame_ctx_idx, 2); // Frame context index.
+    prev_frame_probs =
+        frame_ctxs[frame_ctx_idx].probs; // reference probabilities for compressed header
+    frame_ctxs[frame_ctx_idx] = {current_frame_number, false, current_frame_info.entropy};
 
     uncomp_writer.WriteU(current_frame_info.first_level, 6);
     uncomp_writer.WriteU(current_frame_info.sharpness_level, 3);
@@ -776,6 +812,7 @@ const std::vector<u8>& VP9::ComposeFrameHeader(const NvdecCommon::NvdecRegisters
         current_frame_info = curr_frame.info;
         bitstream = std::move(curr_frame.bit_stream);
     }
+
     // The uncompressed header routine sets PrevProb parameters needed for the compressed header
     auto uncomp_writer = ComposeUncompressedHeader();
     std::vector<u8> compressed_header = ComposeCompressedHeader();
@@ -791,6 +828,13 @@ const std::vector<u8>& VP9::ComposeFrameHeader(const NvdecCommon::NvdecRegisters
               frame.begin() + uncompressed_header.size());
     std::copy(bitstream.begin(), bitstream.end(),
               frame.begin() + uncompressed_header.size() + compressed_header.size());
+
+    // keep track of frame number
+    current_frame_number++;
+    grace_period--;
+
+    // don't display hidden frames
+    hidden = !current_frame_info.show_frame;
     return frame;
 }
 

src/video_core/command_classes/codecs/vp9.h

@@ -14,6 +14,7 @@
 
 namespace Tegra {
 class GPU;
+enum class FrameType { KeyFrame = 0, InterFrame = 1 };
 namespace Decoder {
 
 /// The VpxRangeEncoder, and VpxBitStreamWriter classes are used to compose the
@@ -123,7 +124,7 @@ public:
 
     /// Returns true if the most recent frame was a hidden frame.
     [[nodiscard]] bool WasFrameHidden() const {
-        return !current_frame_info.show_frame;
+        return hidden;
     }
 
 private:
@@ -177,12 +178,19 @@ private:
     std::array<s8, 4> loop_filter_ref_deltas{};
     std::array<s8, 2> loop_filter_mode_deltas{};
 
+    bool hidden = false;
+    s64 current_frame_number = -2; // since we buffer 2 frames
+    s32 grace_period = 6;          // frame offsets need to stabilize
+    std::array<FrameContexts, 4> frame_ctxs{};
     Vp9FrameContainer next_frame{};
-    std::array<Vp9EntropyProbs, 4> frame_ctxs{};
-    bool swap_ref_indices{};
+    Vp9FrameContainer next_next_frame{};
+    bool swap_next_golden{};
 
     Vp9PictureInfo current_frame_info{};
     Vp9EntropyProbs prev_frame_probs{};
+
+    s32 diff_update_probability = 252;
+    s32 frame_sync_code = 0x498342;
 };
 
 } // namespace Decoder

src/video_core/command_classes/codecs/vp9_types.h

@@ -296,6 +296,12 @@ struct RefPoolElement {
     bool refresh{};
 };
 
+struct FrameContexts {
+    s64 from;
+    bool adapted;
+    Vp9EntropyProbs probs;
+};
+
 #define ASSERT_POSITION(field_name, position)                                                      \
     static_assert(offsetof(Vp9EntropyProbs, field_name) == position,                               \
                   "Field " #field_name " has invalid position")

src/video_core/host_shaders/astc_decoder.comp

@@ -10,33 +10,27 @@
 #define END_PUSH_CONSTANTS };
 #define UNIFORM(n)
 #define BINDING_INPUT_BUFFER 0
-#define BINDING_ENC_BUFFER 1
-#define BINDING_SWIZZLE_BUFFER 2
-#define BINDING_OUTPUT_IMAGE 3
+#define BINDING_OUTPUT_IMAGE 1
 
 #else // ^^^ Vulkan ^^^ // vvv OpenGL vvv
 
 #define BEGIN_PUSH_CONSTANTS
 #define END_PUSH_CONSTANTS
 #define UNIFORM(n) layout(location = n) uniform
-#define BINDING_SWIZZLE_BUFFER 0
-#define BINDING_INPUT_BUFFER 1
-#define BINDING_ENC_BUFFER 2
+#define BINDING_INPUT_BUFFER 0
 #define BINDING_OUTPUT_IMAGE 0
 
 #endif
 
-layout(local_size_x = 32, local_size_y = 32, local_size_z = 1) in;
+layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
 
 BEGIN_PUSH_CONSTANTS
 UNIFORM(1) uvec2 block_dims;
-
-UNIFORM(2) uint bytes_per_block_log2;
-UNIFORM(3) uint layer_stride;
-UNIFORM(4) uint block_size;
-UNIFORM(5) uint x_shift;
-UNIFORM(6) uint block_height;
-UNIFORM(7) uint block_height_mask;
+UNIFORM(2) uint layer_stride;
+UNIFORM(3) uint block_size;
+UNIFORM(4) uint x_shift;
+UNIFORM(5) uint block_height;
+UNIFORM(6) uint block_height_mask;
 END_PUSH_CONSTANTS
 
 struct EncodingData {
@@ -55,45 +49,35 @@ struct TexelWeightParams {
     bool void_extent_hdr;
 };
 
-// Swizzle data
-layout(binding = BINDING_SWIZZLE_BUFFER, std430) readonly buffer SwizzleTable {
-    uint swizzle_table[];
-};
-
 layout(binding = BINDING_INPUT_BUFFER, std430) readonly buffer InputBufferU32 {
-    uint astc_data[];
-};
-
-// ASTC Encodings data
-layout(binding = BINDING_ENC_BUFFER, std430) readonly buffer EncodingsValues {
-    EncodingData encoding_values[];
+    uvec4 astc_data[];
 };
 
 layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly image2DArray dest_image;
 
-const uint GOB_SIZE_X = 64;
-const uint GOB_SIZE_Y = 8;
-const uint GOB_SIZE_Z = 1;
-const uint GOB_SIZE = GOB_SIZE_X * GOB_SIZE_Y * GOB_SIZE_Z;
-
 const uint GOB_SIZE_X_SHIFT = 6;
 const uint GOB_SIZE_Y_SHIFT = 3;
-const uint GOB_SIZE_Z_SHIFT = 0;
-const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT + GOB_SIZE_Z_SHIFT;
+const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT;
 
-const uvec2 SWIZZLE_MASK = uvec2(GOB_SIZE_X - 1, GOB_SIZE_Y - 1);
-
-const int BLOCK_SIZE_IN_BYTES = 16;
-
-const int BLOCK_INFO_ERROR = 0;
-const int BLOCK_INFO_VOID_EXTENT_HDR = 1;
-const int BLOCK_INFO_VOID_EXTENT_LDR = 2;
-const int BLOCK_INFO_NORMAL = 3;
+const uint BYTES_PER_BLOCK_LOG2 = 4;
 
 const int JUST_BITS = 0;
 const int QUINT = 1;
 const int TRIT = 2;
 
+// ASTC Encodings data, sorted in ascending order based on their BitLength value
+// (see GetBitLength() function)
+EncodingData encoding_values[22] = EncodingData[](
+    EncodingData(JUST_BITS, 0, 0, 0), EncodingData(JUST_BITS, 1, 0, 0), EncodingData(TRIT, 0, 0, 0),
+    EncodingData(JUST_BITS, 2, 0, 0), EncodingData(QUINT, 0, 0, 0), EncodingData(TRIT, 1, 0, 0),
+    EncodingData(JUST_BITS, 3, 0, 0), EncodingData(QUINT, 1, 0, 0), EncodingData(TRIT, 2, 0, 0),
+    EncodingData(JUST_BITS, 4, 0, 0), EncodingData(QUINT, 2, 0, 0), EncodingData(TRIT, 3, 0, 0),
+    EncodingData(JUST_BITS, 5, 0, 0), EncodingData(QUINT, 3, 0, 0), EncodingData(TRIT, 4, 0, 0),
+    EncodingData(JUST_BITS, 6, 0, 0), EncodingData(QUINT, 4, 0, 0), EncodingData(TRIT, 5, 0, 0),
+    EncodingData(JUST_BITS, 7, 0, 0), EncodingData(QUINT, 5, 0, 0), EncodingData(TRIT, 6, 0, 0),
+    EncodingData(JUST_BITS, 8, 0, 0)
+);
+
 // The following constants are expanded variants of the Replicate()
 // function calls corresponding to the following arguments:
 // value: index into the generated table
@@ -135,44 +119,37 @@ const uint REPLICATE_7_BIT_TO_8_TABLE[128] =
 // Input ASTC texture globals
 uint current_index = 0;
 int bitsread = 0;
-uint total_bitsread = 0;
-uint local_buff[16];
+int total_bitsread = 0;
+uvec4 local_buff;
 
 // Color data globals
-uint color_endpoint_data[16];
+uvec4 color_endpoint_data;
 int color_bitsread = 0;
-uint total_color_bitsread = 0;
-int color_index = 0;
 
 // Four values, two endpoints, four maximum paritions
 uint color_values[32];
 int colvals_index = 0;
 
 // Weight data globals
-uint texel_weight_data[16];
+uvec4 texel_weight_data;
 int texel_bitsread = 0;
-uint total_texel_bitsread = 0;
-int texel_index = 0;
 
 bool texel_flag = false;
 
 // Global "vectors" to be pushed into when decoding
-EncodingData result_vector[100];
+EncodingData result_vector[144];
 int result_index = 0;
 
-EncodingData texel_vector[100];
+EncodingData texel_vector[144];
 int texel_vector_index = 0;
 
 uint unquantized_texel_weights[2][144];
 
 uint SwizzleOffset(uvec2 pos) {
-    pos = pos & SWIZZLE_MASK;
-    return swizzle_table[pos.y * 64 + pos.x];
-}
-
-uint ReadTexel(uint offset) {
-    // extract the 8-bit value from the 32-bit packed data.
-    return bitfieldExtract(astc_data[offset / 4], int((offset * 8) & 24), 8);
+    uint x = pos.x;
+    uint y = pos.y;
+    return ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 +
+                          (y % 2) * 16 + (x % 16);
 }
 
 // Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
@@ -278,14 +255,10 @@ uint Hash52(uint p) {
     return p;
 }
 
-uint SelectPartition(uint seed, uint x, uint y, uint z, uint partition_count, bool small_block) {
-    if (partition_count == 1) {
-        return 0;
-    }
+uint Select2DPartition(uint seed, uint x, uint y, uint partition_count, bool small_block) {
     if (small_block) {
         x <<= 1;
         y <<= 1;
-        z <<= 1;
     }
 
     seed += (partition_count - 1) * 1024;
@@ -299,10 +272,6 @@ uint SelectPartition(uint seed, uint x, uint y, uint z, uint partition_count, bo
     uint seed6 = uint((rnum >> 20) & 0xF);
     uint seed7 = uint((rnum >> 24) & 0xF);
     uint seed8 = uint((rnum >> 28) & 0xF);
-    uint seed9 = uint((rnum >> 18) & 0xF);
-    uint seed10 = uint((rnum >> 22) & 0xF);
-    uint seed11 = uint((rnum >> 26) & 0xF);
-    uint seed12 = uint(((rnum >> 30) | (rnum << 2)) & 0xF);
 
     seed1 = (seed1 * seed1);
     seed2 = (seed2 * seed2);
@@ -312,12 +281,8 @@ uint SelectPartition(uint seed, uint x, uint y, uint z, uint partition_count, bo
     seed6 = (seed6 * seed6);
     seed7 = (seed7 * seed7);
     seed8 = (seed8 * seed8);
-    seed9 = (seed9 * seed9);
-    seed10 = (seed10 * seed10);
-    seed11 = (seed11 * seed11);
-    seed12 = (seed12 * seed12);
 
-    int sh1, sh2, sh3;
+    uint sh1, sh2;
     if ((seed & 1) > 0) {
         sh1 = (seed & 2) > 0 ? 4 : 5;
         sh2 = (partition_count == 3) ? 6 : 5;
@@ -325,25 +290,19 @@ uint SelectPartition(uint seed, uint x, uint y, uint z, uint partition_count, bo
         sh1 = (partition_count == 3) ? 6 : 5;
         sh2 = (seed & 2) > 0 ? 4 : 5;
     }
-    sh3 = (seed & 0x10) > 0 ? sh1 : sh2;
+    seed1 >>= sh1;
+    seed2 >>= sh2;
+    seed3 >>= sh1;
+    seed4 >>= sh2;
+    seed5 >>= sh1;
+    seed6 >>= sh2;
+    seed7 >>= sh1;
+    seed8 >>= sh2;
 
-    seed1 = (seed1 >> sh1);
-    seed2 = (seed2 >> sh2);
-    seed3 = (seed3 >> sh1);
-    seed4 = (seed4 >> sh2);
-    seed5 = (seed5 >> sh1);
-    seed6 = (seed6 >> sh2);
-    seed7 = (seed7 >> sh1);
-    seed8 = (seed8 >> sh2);
-    seed9 = (seed9 >> sh3);
-    seed10 = (seed10 >> sh3);
-    seed11 = (seed11 >> sh3);
-    seed12 = (seed12 >> sh3);
-
-    uint a = seed1 * x + seed2 * y + seed11 * z + (rnum >> 14);
-    uint b = seed3 * x + seed4 * y + seed12 * z + (rnum >> 10);
-    uint c = seed5 * x + seed6 * y + seed9 * z + (rnum >> 6);
-    uint d = seed7 * x + seed8 * y + seed10 * z + (rnum >> 2);
+    uint a = seed1 * x + seed2 * y + (rnum >> 14);
+    uint b = seed3 * x + seed4 * y + (rnum >> 10);
+    uint c = seed5 * x + seed6 * y + (rnum >> 6);
+    uint d = seed7 * x + seed8 * y + (rnum >> 2);
 
     a &= 0x3F;
     b &= 0x3F;
@@ -368,58 +327,37 @@ uint SelectPartition(uint seed, uint x, uint y, uint z, uint partition_count, bo
     }
 }
 
-uint Select2DPartition(uint seed, uint x, uint y, uint partition_count, bool small_block) {
-    return SelectPartition(seed, x, y, 0, partition_count, small_block);
-}
-
-uint ReadBit() {
-    if (current_index >= local_buff.length()) {
+uint ExtractBits(uvec4 payload, int offset, int bits) {
+    if (bits <= 0) {
         return 0;
     }
-    uint bit = bitfieldExtract(local_buff[current_index], bitsread, 1);
-    ++bitsread;
-    ++total_bitsread;
-    if (bitsread == 8) {
-        ++current_index;
-        bitsread = 0;
+    int last_offset = offset + bits - 1;
+    int shifted_offset = offset >> 5;
+    if ((last_offset >> 5) == shifted_offset) {
+        return bitfieldExtract(payload[shifted_offset], offset & 31, bits);
     }
-    return bit;
+    int first_bits = 32 - (offset & 31);
+    int result_first = int(bitfieldExtract(payload[shifted_offset], offset & 31, first_bits));
+    int result_second = int(bitfieldExtract(payload[shifted_offset + 1], 0, bits - first_bits));
+    return result_first | (result_second << first_bits);
 }
 
 uint StreamBits(uint num_bits) {
-    uint ret = 0;
-    for (uint i = 0; i < num_bits; i++) {
-        ret |= ((ReadBit() & 1) << i);
-    }
+    int int_bits = int(num_bits);
+    uint ret = ExtractBits(local_buff, total_bitsread, int_bits);
+    total_bitsread += int_bits;
     return ret;
 }
 
-uint ReadColorBit() {
-    uint bit = 0;
-    if (texel_flag) {
-        bit = bitfieldExtract(texel_weight_data[texel_index], texel_bitsread, 1);
-        ++texel_bitsread;
-        ++total_texel_bitsread;
-        if (texel_bitsread == 8) {
-            ++texel_index;
-            texel_bitsread = 0;
-        }
-    } else {
-        bit = bitfieldExtract(color_endpoint_data[color_index], color_bitsread, 1);
-        ++color_bitsread;
-        ++total_color_bitsread;
-        if (color_bitsread == 8) {
-            ++color_index;
-            color_bitsread = 0;
-        }
-    }
-    return bit;
-}
-
 uint StreamColorBits(uint num_bits) {
     uint ret = 0;
-    for (uint i = 0; i < num_bits; i++) {
-        ret |= ((ReadColorBit() & 1) << i);
+    int int_bits = int(num_bits);
+    if (texel_flag) {
+        ret = ExtractBits(texel_weight_data, texel_bitsread, int_bits);
+        texel_bitsread += int_bits;
+    } else {
+        ret = ExtractBits(color_endpoint_data, color_bitsread, int_bits);
+        color_bitsread += int_bits;
     }
     return ret;
 }
@@ -596,22 +534,16 @@ void DecodeColorValues(uvec4 modes, uint num_partitions, uint color_data_bits) {
     for (uint i = 0; i < num_partitions; i++) {
         num_values += ((modes[i] >> 2) + 1) << 1;
     }
-    int range = 256;
-    while (--range > 0) {
-        EncodingData val = encoding_values[range];
+    // Find the largest encoding that's within color_data_bits
+    // TODO(ameerj): profile with binary search
+    int range = 0;
+    while (++range < encoding_values.length()) {
         uint bit_length = GetBitLength(num_values, range);
-        if (bit_length <= color_data_bits) {
-            while (--range > 0) {
-                EncodingData newval = encoding_values[range];
-                if (newval.encoding != val.encoding && newval.num_bits != val.num_bits) {
-                    break;
-                }
-            }
-            ++range;
+        if (bit_length > color_data_bits) {
             break;
         }
     }
-    DecodeIntegerSequence(range, num_values);
+    DecodeIntegerSequence(range - 1, num_values);
     uint out_index = 0;
     for (int itr = 0; itr < result_index; ++itr) {
         if (out_index >= num_values) {
@@ -1028,7 +960,7 @@ int FindLayout(uint mode) {
     return 5;
 }
 
-TexelWeightParams DecodeBlockInfo(uint block_index) {
+TexelWeightParams DecodeBlockInfo() {
     TexelWeightParams params = TexelWeightParams(uvec2(0), 0, false, false, false, false);
     uint mode = StreamBits(11);
     if ((mode & 0x1ff) == 0x1fc) {
@@ -1110,10 +1042,10 @@ TexelWeightParams DecodeBlockInfo(uint block_index) {
     }
     weight_index -= 2;
     if ((mode_layout != 9) && ((mode & 0x200) != 0)) {
-        const int max_weights[6] = int[6](9, 11, 15, 19, 23, 31);
+        const int max_weights[6] = int[6](7, 8, 9, 10, 11, 12);
         params.max_weight = max_weights[weight_index];
     } else {
-        const int max_weights[6] = int[6](1, 2, 3, 4, 5, 7);
+        const int max_weights[6] = int[6](1, 2, 3, 4, 5, 6);
         params.max_weight = max_weights[weight_index];
     }
     return params;
@@ -1144,8 +1076,8 @@ void FillVoidExtentLDR(ivec3 coord) {
     }
 }
 
-void DecompressBlock(ivec3 coord, uint block_index) {
-    TexelWeightParams params = DecodeBlockInfo(block_index);
+void DecompressBlock(ivec3 coord) {
+    TexelWeightParams params = DecodeBlockInfo();
     if (params.error_state) {
         FillError(coord);
         return;
@@ -1212,7 +1144,7 @@ void DecompressBlock(ivec3 coord, uint block_index) {
     // Read color data...
     uint color_data_bits = remaining_bits;
     while (remaining_bits > 0) {
-        int nb = int(min(remaining_bits, 8U));
+        int nb = int(min(remaining_bits, 32U));
         uint b = StreamBits(nb);
         color_endpoint_data[ced_pointer] = uint(bitfieldExtract(b, 0, nb));
         ++ced_pointer;
@@ -1254,25 +1186,20 @@ void DecompressBlock(ivec3 coord, uint block_index) {
         ComputeEndpoints(endpoints[i][0], endpoints[i][1], color_endpoint_mode[i]);
     }
 
-    for (uint i = 0; i < 16; i++) {
-        texel_weight_data[i] = local_buff[i];
-    }
-    for (uint i = 0; i < 8; i++) {
-#define REVERSE_BYTE(b) ((b * 0x0802U & 0x22110U) | (b * 0x8020U & 0x88440U)) * 0x10101U >> 16
-        uint a = REVERSE_BYTE(texel_weight_data[i]);
-        uint b = REVERSE_BYTE(texel_weight_data[15 - i]);
-#undef REVERSE_BYTE
-        texel_weight_data[i] = uint(bitfieldExtract(b, 0, 8));
-        texel_weight_data[15 - i] = uint(bitfieldExtract(a, 0, 8));
-    }
+    texel_weight_data = local_buff;
+    texel_weight_data = bitfieldReverse(texel_weight_data).wzyx;
     uint clear_byte_start =
         (GetPackedBitSize(params.size, params.dual_plane, params.max_weight) >> 3) + 1;
-    texel_weight_data[clear_byte_start - 1] =
-        texel_weight_data[clear_byte_start - 1] &
+
+    uint byte_insert = ExtractBits(texel_weight_data, int(clear_byte_start - 1) * 8, 8) &
         uint(
             ((1 << (GetPackedBitSize(params.size, params.dual_plane, params.max_weight) % 8)) - 1));
-    for (uint i = 0; i < 16 - clear_byte_start; i++) {
-        texel_weight_data[clear_byte_start + i] = 0U;
+    uint vec_index = (clear_byte_start - 1) >> 2;
+    texel_weight_data[vec_index] =
+        bitfieldInsert(texel_weight_data[vec_index], byte_insert, int((clear_byte_start - 1) % 4) * 8, 8);
+    for (uint i = clear_byte_start; i < 16; ++i) {
+        uint idx = i >> 2;
+        texel_weight_data[idx] = bitfieldInsert(texel_weight_data[idx], 0, int(i % 4) * 8, 8);
     }
     texel_flag = true; // use texel "vector" and bit stream in integer decoding
     DecodeIntegerSequence(params.max_weight, GetNumWeightValues(params.size, params.dual_plane));
@@ -1281,8 +1208,11 @@ void DecompressBlock(ivec3 coord, uint block_index) {
 
     for (uint j = 0; j < block_dims.y; j++) {
         for (uint i = 0; i < block_dims.x; i++) {
-            uint local_partition = Select2DPartition(partition_index, i, j, num_partitions,
+            uint local_partition = 0;
+            if (num_partitions > 1) {
+                local_partition = Select2DPartition(partition_index, i, j, num_partitions,
                                                      (block_dims.y * block_dims.x) < 32);
+            }
             vec4 p;
             uvec4 C0 = ReplicateByteTo16(endpoints[local_partition][0]);
             uvec4 C1 = ReplicateByteTo16(endpoints[local_partition][1]);
@@ -1303,7 +1233,7 @@ void DecompressBlock(ivec3 coord, uint block_index) {
 
 void main() {
     uvec3 pos = gl_GlobalInvocationID;
-    pos.x <<= bytes_per_block_log2;
+    pos.x <<= BYTES_PER_BLOCK_LOG2;
 
     // Read as soon as possible due to its latency
     const uint swizzle = SwizzleOffset(pos.xy);
@@ -1321,13 +1251,8 @@ void main() {
     if (any(greaterThanEqual(coord, imageSize(dest_image)))) {
         return;
     }
-    uint block_index =
-        pos.z * gl_WorkGroupSize.x * gl_WorkGroupSize.y + pos.y * gl_WorkGroupSize.x + pos.x;
-
     current_index = 0;
     bitsread = 0;
-    for (int i = 0; i < 16; i++) {
-        local_buff[i] = ReadTexel(offset + i);
-    }
-    DecompressBlock(coord, block_index);
+    local_buff = astc_data[offset / 16];
+    DecompressBlock(coord);
 }

src/video_core/renderer_opengl/util_shaders.cpp

@@ -60,19 +60,14 @@ UtilShaders::UtilShaders(ProgramManager& program_manager_)
       copy_bc4_program(MakeProgram(OPENGL_COPY_BC4_COMP)) {
     const auto swizzle_table = Tegra::Texture::MakeSwizzleTable();
     swizzle_table_buffer.Create();
-    astc_buffer.Create();
     glNamedBufferStorage(swizzle_table_buffer.handle, sizeof(swizzle_table), &swizzle_table, 0);
-    glNamedBufferStorage(astc_buffer.handle, sizeof(ASTC_ENCODINGS_VALUES), &ASTC_ENCODINGS_VALUES,
-                         0);
 }
 
 UtilShaders::~UtilShaders() = default;
 
 void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
                              std::span<const VideoCommon::SwizzleParameters> swizzles) {
-    static constexpr GLuint BINDING_SWIZZLE_BUFFER = 0;
-    static constexpr GLuint BINDING_INPUT_BUFFER = 1;
-    static constexpr GLuint BINDING_ENC_BUFFER = 2;
+    static constexpr GLuint BINDING_INPUT_BUFFER = 0;
     static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
 
     const Extent2D tile_size{
@@ -80,34 +75,32 @@ void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
         .height = VideoCore::Surface::DefaultBlockHeight(image.info.format),
     };
     program_manager.BindComputeProgram(astc_decoder_program.handle);
-    glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_SWIZZLE_BUFFER, swizzle_table_buffer.handle);
-    glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_ENC_BUFFER, astc_buffer.handle);
-
     glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
     glUniform2ui(1, tile_size.width, tile_size.height);
+
     // Ensure buffer data is valid before dispatching
     glFlush();
     for (const SwizzleParameters& swizzle : swizzles) {
         const size_t input_offset = swizzle.buffer_offset + map.offset;
-        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
-        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
+        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 8U);
+        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 8U);
 
         const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
         ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0}));
         ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0}));
+        ASSERT(params.bytes_per_block_log2 == 4);
 
-        glUniform1ui(2, params.bytes_per_block_log2);
-        glUniform1ui(3, params.layer_stride);
-        glUniform1ui(4, params.block_size);
-        glUniform1ui(5, params.x_shift);
-        glUniform1ui(6, params.block_height);
-        glUniform1ui(7, params.block_height_mask);
+        glUniform1ui(2, params.layer_stride);
+        glUniform1ui(3, params.block_size);
+        glUniform1ui(4, params.x_shift);
+        glUniform1ui(5, params.block_height);
+        glUniform1ui(6, params.block_height_mask);
 
-        glBindImageTexture(BINDING_OUTPUT_IMAGE, image.StorageHandle(), swizzle.level, GL_TRUE, 0,
-                           GL_WRITE_ONLY, GL_RGBA8);
         // ASTC texture data
         glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,
                           image.guest_size_bytes - swizzle.buffer_offset);
+        glBindImageTexture(BINDING_OUTPUT_IMAGE, image.StorageHandle(), swizzle.level, GL_TRUE, 0,
+                           GL_WRITE_ONLY, GL_RGBA8);
 
         glDispatchCompute(num_dispatches_x, num_dispatches_y, image.info.resources.layers);
     }

src/video_core/renderer_opengl/util_shaders.h

@@ -62,7 +62,6 @@ private:
     ProgramManager& program_manager;
 
     OGLBuffer swizzle_table_buffer;
-    OGLBuffer astc_buffer;
 
     OGLProgram astc_decoder_program;
     OGLProgram block_linear_unswizzle_2d_program;

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -30,16 +30,12 @@
 namespace Vulkan {
 
 using Tegra::Texture::SWIZZLE_TABLE;
-using Tegra::Texture::ASTC::ASTC_ENCODINGS_VALUES;
-using namespace Tegra::Texture::ASTC;
 
 namespace {
 
 constexpr u32 ASTC_BINDING_INPUT_BUFFER = 0;
-constexpr u32 ASTC_BINDING_ENC_BUFFER = 1;
-constexpr u32 ASTC_BINDING_SWIZZLE_BUFFER = 2;
-constexpr u32 ASTC_BINDING_OUTPUT_IMAGE = 3;
-constexpr size_t ASTC_NUM_BINDINGS = 4;
+constexpr u32 ASTC_BINDING_OUTPUT_IMAGE = 1;
+constexpr size_t ASTC_NUM_BINDINGS = 2;
 
 template <size_t size>
 inline constexpr VkPushConstantRange COMPUTE_PUSH_CONSTANT_RANGE{
@@ -75,7 +71,7 @@ constexpr DescriptorBankInfo INPUT_OUTPUT_BANK_INFO{
     .score = 2,
 };
 
-constexpr std::array<VkDescriptorSetLayoutBinding, 4> ASTC_DESCRIPTOR_SET_BINDINGS{{
+constexpr std::array<VkDescriptorSetLayoutBinding, ASTC_NUM_BINDINGS> ASTC_DESCRIPTOR_SET_BINDINGS{{
     {
         .binding = ASTC_BINDING_INPUT_BUFFER,
         .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
@@ -83,20 +79,6 @@ constexpr std::array<VkDescriptorSetLayoutBinding, 4> ASTC_DESCRIPTOR_SET_BINDIN
         .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
         .pImmutableSamplers = nullptr,
     },
-    {
-        .binding = ASTC_BINDING_ENC_BUFFER,
-        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-        .descriptorCount = 1,
-        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
-        .pImmutableSamplers = nullptr,
-    },
-    {
-        .binding = ASTC_BINDING_SWIZZLE_BUFFER,
-        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-        .descriptorCount = 1,
-        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
-        .pImmutableSamplers = nullptr,
-    },
     {
         .binding = ASTC_BINDING_OUTPUT_IMAGE,
         .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
@@ -108,12 +90,12 @@ constexpr std::array<VkDescriptorSetLayoutBinding, 4> ASTC_DESCRIPTOR_SET_BINDIN
 
 constexpr DescriptorBankInfo ASTC_BANK_INFO{
     .uniform_buffers = 0,
-    .storage_buffers = 3,
+    .storage_buffers = 1,
     .texture_buffers = 0,
     .image_buffers = 0,
     .textures = 0,
     .images = 1,
-    .score = 4,
+    .score = 2,
 };
 
 constexpr VkDescriptorUpdateTemplateEntryKHR INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE{
@@ -135,22 +117,6 @@ constexpr std::array<VkDescriptorUpdateTemplateEntryKHR, ASTC_NUM_BINDINGS>
             .offset = ASTC_BINDING_INPUT_BUFFER * sizeof(DescriptorUpdateEntry),
             .stride = sizeof(DescriptorUpdateEntry),
         },
-        {
-            .dstBinding = ASTC_BINDING_ENC_BUFFER,
-            .dstArrayElement = 0,
-            .descriptorCount = 1,
-            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-            .offset = ASTC_BINDING_ENC_BUFFER * sizeof(DescriptorUpdateEntry),
-            .stride = sizeof(DescriptorUpdateEntry),
-        },
-        {
-            .dstBinding = ASTC_BINDING_SWIZZLE_BUFFER,
-            .dstArrayElement = 0,
-            .descriptorCount = 1,
-            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-            .offset = ASTC_BINDING_SWIZZLE_BUFFER * sizeof(DescriptorUpdateEntry),
-            .stride = sizeof(DescriptorUpdateEntry),
-        },
         {
             .dstBinding = ASTC_BINDING_OUTPUT_IMAGE,
             .dstArrayElement = 0,
@@ -163,7 +129,6 @@ constexpr std::array<VkDescriptorUpdateTemplateEntryKHR, ASTC_NUM_BINDINGS>
 
 struct AstcPushConstants {
     std::array<u32, 2> blocks_dims;
-    u32 bytes_per_block_log2;
     u32 layer_stride;
     u32 block_size;
     u32 x_shift;
@@ -354,46 +319,6 @@ ASTCDecoderPass::ASTCDecoderPass(const Device& device_, VKScheduler& scheduler_,
 
 ASTCDecoderPass::~ASTCDecoderPass() = default;
 
-void ASTCDecoderPass::MakeDataBuffer() {
-    constexpr size_t TOTAL_BUFFER_SIZE = sizeof(ASTC_ENCODINGS_VALUES) + sizeof(SWIZZLE_TABLE);
-    data_buffer = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
-        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
-        .pNext = nullptr,
-        .flags = 0,
-        .size = TOTAL_BUFFER_SIZE,
-        .usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
-        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
-        .queueFamilyIndexCount = 0,
-        .pQueueFamilyIndices = nullptr,
-    });
-    data_buffer_commit = memory_allocator.Commit(data_buffer, MemoryUsage::Upload);
-
-    const auto staging_ref = staging_buffer_pool.Request(TOTAL_BUFFER_SIZE, MemoryUsage::Upload);
-    std::memcpy(staging_ref.mapped_span.data(), &ASTC_ENCODINGS_VALUES,
-                sizeof(ASTC_ENCODINGS_VALUES));
-    // Tack on the swizzle table at the end of the buffer
-    std::memcpy(staging_ref.mapped_span.data() + sizeof(ASTC_ENCODINGS_VALUES), &SWIZZLE_TABLE,
-                sizeof(SWIZZLE_TABLE));
-
-    scheduler.Record([src = staging_ref.buffer, offset = staging_ref.offset, dst = *data_buffer,
-                      TOTAL_BUFFER_SIZE](vk::CommandBuffer cmdbuf) {
-        static constexpr VkMemoryBarrier write_barrier{
-            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
-            .pNext = nullptr,
-            .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
-        };
-        const VkBufferCopy copy{
-            .srcOffset = offset,
-            .dstOffset = 0,
-            .size = TOTAL_BUFFER_SIZE,
-        };
-        cmdbuf.CopyBuffer(src, dst, copy);
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-                               0, write_barrier);
-    });
-}
-
 void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
                                std::span<const VideoCommon::SwizzleParameters> swizzles) {
     using namespace VideoCommon::Accelerated;
@@ -402,9 +327,6 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         VideoCore::Surface::DefaultBlockHeight(image.info.format),
     };
     scheduler.RequestOutsideRenderPassOperationContext();
-    if (!data_buffer) {
-        MakeDataBuffer();
-    }
     const VkPipeline vk_pipeline = *pipeline;
     const VkImageAspectFlags aspect_mask = image.AspectMask();
     const VkImage vk_image = image.Handle();
@@ -436,16 +358,13 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         });
     for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
         const size_t input_offset = swizzle.buffer_offset + map.offset;
-        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
-        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
+        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 8U);
+        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 8U);
         const u32 num_dispatches_z = image.info.resources.layers;
 
         update_descriptor_queue.Acquire();
         update_descriptor_queue.AddBuffer(map.buffer, input_offset,
                                           image.guest_size_bytes - swizzle.buffer_offset);
-        update_descriptor_queue.AddBuffer(*data_buffer, 0, sizeof(ASTC_ENCODINGS_VALUES));
-        update_descriptor_queue.AddBuffer(*data_buffer, sizeof(ASTC_ENCODINGS_VALUES),
-                                          sizeof(SWIZZLE_TABLE));
         update_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
         const void* const descriptor_data{update_descriptor_queue.UpdateData()};
 
@@ -453,11 +372,11 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
         ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0}));
         ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0}));
+        ASSERT(params.bytes_per_block_log2 == 4);
         scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, block_dims,
                           params, descriptor_data](vk::CommandBuffer cmdbuf) {
             const AstcPushConstants uniforms{
                 .blocks_dims = block_dims,
-                .bytes_per_block_log2 = params.bytes_per_block_log2,
                 .layer_stride = params.layer_stride,
                 .block_size = params.block_size,
                 .x_shift = params.x_shift,

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -96,15 +96,10 @@ public:
                   std::span<const VideoCommon::SwizzleParameters> swizzles);
 
 private:
-    void MakeDataBuffer();
-
     VKScheduler& scheduler;
     StagingBufferPool& staging_buffer_pool;
     VKUpdateDescriptorQueue& update_descriptor_queue;
     MemoryAllocator& memory_allocator;
-
-    vk::Buffer data_buffer;
-    MemoryCommit data_buffer_commit;
 };
 
 } // namespace Vulkan

src/video_core/textures/astc.cpp

@@ -151,6 +151,76 @@ private:
     const IntType& m_Bits;
 };
 
+enum class IntegerEncoding { JustBits, Quint, Trit };
+
+struct IntegerEncodedValue {
+    constexpr IntegerEncodedValue() = default;
+
+    constexpr IntegerEncodedValue(IntegerEncoding encoding_, u32 num_bits_)
+        : encoding{encoding_}, num_bits{num_bits_} {}
+
+    constexpr bool MatchesEncoding(const IntegerEncodedValue& other) const {
+        return encoding == other.encoding && num_bits == other.num_bits;
+    }
+
+    // Returns the number of bits required to encode num_vals values.
+    u32 GetBitLength(u32 num_vals) const {
+        u32 total_bits = num_bits * num_vals;
+        if (encoding == IntegerEncoding::Trit) {
+            total_bits += (num_vals * 8 + 4) / 5;
+        } else if (encoding == IntegerEncoding::Quint) {
+            total_bits += (num_vals * 7 + 2) / 3;
+        }
+        return total_bits;
+    }
+
+    IntegerEncoding encoding{};
+    u32 num_bits = 0;
+    u32 bit_value = 0;
+    union {
+        u32 quint_value = 0;
+        u32 trit_value;
+    };
+};
+
+// Returns a new instance of this struct that corresponds to the
+// can take no more than mav_value values
+static constexpr IntegerEncodedValue CreateEncoding(u32 mav_value) {
+    while (mav_value > 0) {
+        u32 check = mav_value + 1;
+
+        // Is mav_value a power of two?
+        if (!(check & (check - 1))) {
+            return IntegerEncodedValue(IntegerEncoding::JustBits, std::popcount(mav_value));
+        }
+
+        // Is mav_value of the type 3*2^n - 1?
+        if ((check % 3 == 0) && !((check / 3) & ((check / 3) - 1))) {
+            return IntegerEncodedValue(IntegerEncoding::Trit, std::popcount(check / 3 - 1));
+        }
+
+        // Is mav_value of the type 5*2^n - 1?
+        if ((check % 5 == 0) && !((check / 5) & ((check / 5) - 1))) {
+            return IntegerEncodedValue(IntegerEncoding::Quint, std::popcount(check / 5 - 1));
+        }
+
+        // Apparently it can't be represented with a bounded integer sequence...
+        // just iterate.
+        mav_value--;
+    }
+    return IntegerEncodedValue(IntegerEncoding::JustBits, 0);
+}
+
+static constexpr std::array<IntegerEncodedValue, 256> MakeEncodedValues() {
+    std::array<IntegerEncodedValue, 256> encodings{};
+    for (std::size_t i = 0; i < encodings.size(); ++i) {
+        encodings[i] = CreateEncoding(static_cast<u32>(i));
+    }
+    return encodings;
+}
+
+static constexpr std::array<IntegerEncodedValue, 256> ASTC_ENCODINGS_VALUES = MakeEncodedValues();
+
 namespace Tegra::Texture::ASTC {
 using IntegerEncodedVector = boost::container::static_vector<
     IntegerEncodedValue, 256,
@@ -521,35 +591,41 @@ static TexelWeightParams DecodeBlockInfo(InputBitStream& strm) {
     return params;
 }
 
-static void FillVoidExtentLDR(InputBitStream& strm, std::span<u32> outBuf, u32 blockWidth,
-                              u32 blockHeight) {
-    // Don't actually care about the void extent, just read the bits...
-    for (s32 i = 0; i < 4; ++i) {
-        strm.ReadBits<13>();
+// Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
+// is the same as [(num_bits - 1):0] and repeats all the way down.
+template <typename IntType>
+static constexpr IntType Replicate(IntType val, u32 num_bits, u32 to_bit) {
+    if (num_bits == 0 || to_bit == 0) {
+        return 0;
     }
-
-    // Decode the RGBA components and renormalize them to the range [0, 255]
-    u16 r = static_cast<u16>(strm.ReadBits<16>());
-    u16 g = static_cast<u16>(strm.ReadBits<16>());
-    u16 b = static_cast<u16>(strm.ReadBits<16>());
-    u16 a = static_cast<u16>(strm.ReadBits<16>());
-
-    u32 rgba = (r >> 8) | (g & 0xFF00) | (static_cast<u32>(b) & 0xFF00) << 8 |
-               (static_cast<u32>(a) & 0xFF00) << 16;
-
-    for (u32 j = 0; j < blockHeight; j++) {
-        for (u32 i = 0; i < blockWidth; i++) {
-            outBuf[j * blockWidth + i] = rgba;
+    const IntType v = val & static_cast<IntType>((1 << num_bits) - 1);
+    IntType res = v;
+    u32 reslen = num_bits;
+    while (reslen < to_bit) {
+        u32 comp = 0;
+        if (num_bits > to_bit - reslen) {
+            u32 newshift = to_bit - reslen;
+            comp = num_bits - newshift;
+            num_bits = newshift;
         }
+        res = static_cast<IntType>(res << num_bits);
+        res = static_cast<IntType>(res | (v >> comp));
+        reslen += num_bits;
     }
+    return res;
 }
 
-static void FillError(std::span<u32> outBuf, u32 blockWidth, u32 blockHeight) {
-    for (u32 j = 0; j < blockHeight; j++) {
-        for (u32 i = 0; i < blockWidth; i++) {
-            outBuf[j * blockWidth + i] = 0xFFFF00FF;
-        }
+static constexpr std::size_t NumReplicateEntries(u32 num_bits) {
+    return std::size_t(1) << num_bits;
+}
+
+template <typename IntType, u32 num_bits, u32 to_bit>
+static constexpr auto MakeReplicateTable() {
+    std::array<IntType, NumReplicateEntries(num_bits)> table{};
+    for (IntType value = 0; value < static_cast<IntType>(std::size(table)); ++value) {
+        table[value] = Replicate(value, num_bits, to_bit);
     }
+    return table;
 }
 
 static constexpr auto REPLICATE_BYTE_TO_16_TABLE = MakeReplicateTable<u32, 8, 16>();
@@ -572,6 +648,9 @@ static constexpr auto REPLICATE_2_BIT_TO_8_TABLE = MakeReplicateTable<u32, 2, 8>
 static constexpr auto REPLICATE_3_BIT_TO_8_TABLE = MakeReplicateTable<u32, 3, 8>();
 static constexpr auto REPLICATE_4_BIT_TO_8_TABLE = MakeReplicateTable<u32, 4, 8>();
 static constexpr auto REPLICATE_5_BIT_TO_8_TABLE = MakeReplicateTable<u32, 5, 8>();
+static constexpr auto REPLICATE_6_BIT_TO_8_TABLE = MakeReplicateTable<u32, 6, 8>();
+static constexpr auto REPLICATE_7_BIT_TO_8_TABLE = MakeReplicateTable<u32, 7, 8>();
+static constexpr auto REPLICATE_8_BIT_TO_8_TABLE = MakeReplicateTable<u32, 8, 8>();
 /// Use a precompiled table with the most common usages, if it's not in the expected range, fallback
 /// to the runtime implementation
 static constexpr u32 FastReplicateTo8(u32 value, u32 num_bits) {
@@ -1316,6 +1395,37 @@ static void ComputeEndpoints(Pixel& ep1, Pixel& ep2, const u32*& colorValues,
 #undef READ_INT_VALUES
 }
 
+static void FillVoidExtentLDR(InputBitStream& strm, std::span<u32> outBuf, u32 blockWidth,
+                              u32 blockHeight) {
+    // Don't actually care about the void extent, just read the bits...
+    for (s32 i = 0; i < 4; ++i) {
+        strm.ReadBits<13>();
+    }
+
+    // Decode the RGBA components and renormalize them to the range [0, 255]
+    u16 r = static_cast<u16>(strm.ReadBits<16>());
+    u16 g = static_cast<u16>(strm.ReadBits<16>());
+    u16 b = static_cast<u16>(strm.ReadBits<16>());
+    u16 a = static_cast<u16>(strm.ReadBits<16>());
+
+    u32 rgba = (r >> 8) | (g & 0xFF00) | (static_cast<u32>(b) & 0xFF00) << 8 |
+               (static_cast<u32>(a) & 0xFF00) << 16;
+
+    for (u32 j = 0; j < blockHeight; j++) {
+        for (u32 i = 0; i < blockWidth; i++) {
+            outBuf[j * blockWidth + i] = rgba;
+        }
+    }
+}
+
+static void FillError(std::span<u32> outBuf, u32 blockWidth, u32 blockHeight) {
+    for (u32 j = 0; j < blockHeight; j++) {
+        for (u32 i = 0; i < blockWidth; i++) {
+            outBuf[j * blockWidth + i] = 0xFFFF00FF;
+        }
+    }
+}
+
 static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
                             const u32 blockHeight, std::span<u32, 12 * 12> outBuf) {
     InputBitStream strm(inBuf);

src/video_core/textures/astc.h

@@ -9,117 +9,6 @@
 
 namespace Tegra::Texture::ASTC {
 
-enum class IntegerEncoding { JustBits, Quint, Trit };
-
-struct IntegerEncodedValue {
-    constexpr IntegerEncodedValue() = default;
-
-    constexpr IntegerEncodedValue(IntegerEncoding encoding_, u32 num_bits_)
-        : encoding{encoding_}, num_bits{num_bits_} {}
-
-    constexpr bool MatchesEncoding(const IntegerEncodedValue& other) const {
-        return encoding == other.encoding && num_bits == other.num_bits;
-    }
-
-    // Returns the number of bits required to encode num_vals values.
-    u32 GetBitLength(u32 num_vals) const {
-        u32 total_bits = num_bits * num_vals;
-        if (encoding == IntegerEncoding::Trit) {
-            total_bits += (num_vals * 8 + 4) / 5;
-        } else if (encoding == IntegerEncoding::Quint) {
-            total_bits += (num_vals * 7 + 2) / 3;
-        }
-        return total_bits;
-    }
-
-    IntegerEncoding encoding{};
-    u32 num_bits = 0;
-    u32 bit_value = 0;
-    union {
-        u32 quint_value = 0;
-        u32 trit_value;
-    };
-};
-
-// Returns a new instance of this struct that corresponds to the
-// can take no more than mav_value values
-constexpr IntegerEncodedValue CreateEncoding(u32 mav_value) {
-    while (mav_value > 0) {
-        u32 check = mav_value + 1;
-
-        // Is mav_value a power of two?
-        if (!(check & (check - 1))) {
-            return IntegerEncodedValue(IntegerEncoding::JustBits, std::popcount(mav_value));
-        }
-
-        // Is mav_value of the type 3*2^n - 1?
-        if ((check % 3 == 0) && !((check / 3) & ((check / 3) - 1))) {
-            return IntegerEncodedValue(IntegerEncoding::Trit, std::popcount(check / 3 - 1));
-        }
-
-        // Is mav_value of the type 5*2^n - 1?
-        if ((check % 5 == 0) && !((check / 5) & ((check / 5) - 1))) {
-            return IntegerEncodedValue(IntegerEncoding::Quint, std::popcount(check / 5 - 1));
-        }
-
-        // Apparently it can't be represented with a bounded integer sequence...
-        // just iterate.
-        mav_value--;
-    }
-    return IntegerEncodedValue(IntegerEncoding::JustBits, 0);
-}
-
-constexpr std::array<IntegerEncodedValue, 256> MakeEncodedValues() {
-    std::array<IntegerEncodedValue, 256> encodings{};
-    for (std::size_t i = 0; i < encodings.size(); ++i) {
-        encodings[i] = CreateEncoding(static_cast<u32>(i));
-    }
-    return encodings;
-}
-
-constexpr std::array<IntegerEncodedValue, 256> ASTC_ENCODINGS_VALUES = MakeEncodedValues();
-
-// Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
-// is the same as [(num_bits - 1):0] and repeats all the way down.
-template <typename IntType>
-constexpr IntType Replicate(IntType val, u32 num_bits, u32 to_bit) {
-    if (num_bits == 0 || to_bit == 0) {
-        return 0;
-    }
-    const IntType v = val & static_cast<IntType>((1 << num_bits) - 1);
-    IntType res = v;
-    u32 reslen = num_bits;
-    while (reslen < to_bit) {
-        u32 comp = 0;
-        if (num_bits > to_bit - reslen) {
-            u32 newshift = to_bit - reslen;
-            comp = num_bits - newshift;
-            num_bits = newshift;
-        }
-        res = static_cast<IntType>(res << num_bits);
-        res = static_cast<IntType>(res | (v >> comp));
-        reslen += num_bits;
-    }
-    return res;
-}
-
-constexpr std::size_t NumReplicateEntries(u32 num_bits) {
-    return std::size_t(1) << num_bits;
-}
-
-template <typename IntType, u32 num_bits, u32 to_bit>
-constexpr auto MakeReplicateTable() {
-    std::array<IntType, NumReplicateEntries(num_bits)> table{};
-    for (IntType value = 0; value < static_cast<IntType>(std::size(table)); ++value) {
-        table[value] = Replicate(value, num_bits, to_bit);
-    }
-    return table;
-}
-
-constexpr auto REPLICATE_6_BIT_TO_8_TABLE = MakeReplicateTable<u32, 6, 8>();
-constexpr auto REPLICATE_7_BIT_TO_8_TABLE = MakeReplicateTable<u32, 7, 8>();
-constexpr auto REPLICATE_8_BIT_TO_8_TABLE = MakeReplicateTable<u32, 8, 8>();
-
 void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
                 uint32_t block_width, uint32_t block_height, std::span<uint8_t> output);