summaryrefslogtreecommitdiffstats
path: root/src/video_core
diff options
context:
space:
mode:
Diffstat (limited to 'src/video_core')
-rw-r--r--src/video_core/textures/astc.cpp41
-rw-r--r--src/video_core/vulkan_common/vulkan_device.cpp3
-rw-r--r--src/video_core/vulkan_common/vulkan_device.h35
3 files changed, 43 insertions, 36 deletions
diff --git a/src/video_core/textures/astc.cpp b/src/video_core/textures/astc.cpp
index acd5bdd78..3625b666c 100644
--- a/src/video_core/textures/astc.cpp
+++ b/src/video_core/textures/astc.cpp
@@ -42,21 +42,24 @@ constexpr u32 Popcnt(u32 n) {
class InputBitStream {
public:
- constexpr explicit InputBitStream(const u8* ptr, std::size_t start_offset = 0)
- : cur_byte{ptr}, next_bit{start_offset % 8} {}
+ constexpr explicit InputBitStream(std::span<const u8> data, size_t start_offset = 0)
+ : cur_byte{data.data()}, total_bits{data.size()}, next_bit{start_offset % 8} {}
- constexpr std::size_t GetBitsRead() const {
+ constexpr size_t GetBitsRead() const {
return bits_read;
}
constexpr bool ReadBit() {
- const bool bit = (*cur_byte >> next_bit++) & 1;
+ if (bits_read >= total_bits * 8) {
+ return 0;
+ }
+ const bool bit = ((*cur_byte >> next_bit) & 1) != 0;
+ ++next_bit;
while (next_bit >= 8) {
next_bit -= 8;
- cur_byte++;
+ ++cur_byte;
}
-
- bits_read++;
+ ++bits_read;
return bit;
}
@@ -79,8 +82,9 @@ public:
private:
const u8* cur_byte;
- std::size_t next_bit = 0;
- std::size_t bits_read = 0;
+ size_t total_bits = 0;
+ size_t next_bit = 0;
+ size_t bits_read = 0;
};
class OutputBitStream {
@@ -193,15 +197,15 @@ struct IntegerEncodedValue {
};
};
using IntegerEncodedVector = boost::container::static_vector<
- IntegerEncodedValue, 64,
+ IntegerEncodedValue, 256,
boost::container::static_vector_options<
boost::container::inplace_alignment<alignof(IntegerEncodedValue)>,
boost::container::throw_on_overflow<false>>::type>;
static void DecodeTritBlock(InputBitStream& bits, IntegerEncodedVector& result, u32 nBitsPerValue) {
// Implement the algorithm in section C.2.12
- u32 m[5];
- u32 t[5];
+ std::array<u32, 5> m;
+ std::array<u32, 5> t;
u32 T;
// Read the trit encoded block according to
@@ -866,7 +870,7 @@ public:
}
};
-static void DecodeColorValues(u32* out, u8* data, const u32* modes, const u32 nPartitions,
+static void DecodeColorValues(u32* out, std::span<u8> data, const u32* modes, const u32 nPartitions,
const u32 nBitsForColorData) {
// First figure out how many color values we have
u32 nValues = 0;
@@ -898,7 +902,7 @@ static void DecodeColorValues(u32* out, u8* data, const u32* modes, const u32 nP
// We now have enough to decode our integer sequence.
IntegerEncodedVector decodedColorValues;
- InputBitStream colorStream(data);
+ InputBitStream colorStream(data, 0);
DecodeIntegerSequence(decodedColorValues, colorStream, range, nValues);
// Once we have the decoded values, we need to dequantize them to the 0-255 range
@@ -1441,7 +1445,7 @@ static void ComputeEndpos32s(Pixel& ep1, Pixel& ep2, const u32*& colorValues,
static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
const u32 blockHeight, std::span<u32, 12 * 12> outBuf) {
- InputBitStream strm(inBuf.data());
+ InputBitStream strm(inBuf);
TexelWeightParams weightParams = DecodeBlockInfo(strm);
// Was there an error?
@@ -1619,15 +1623,16 @@ static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
// Make sure that higher non-texel bits are set to zero
const u32 clearByteStart = (weightParams.GetPackedBitSize() >> 3) + 1;
- if (clearByteStart > 0) {
+ if (clearByteStart > 0 && clearByteStart <= texelWeightData.size()) {
texelWeightData[clearByteStart - 1] &=
static_cast<u8>((1 << (weightParams.GetPackedBitSize() % 8)) - 1);
+ std::memset(texelWeightData.data() + clearByteStart, 0,
+ std::min(16U - clearByteStart, 16U));
}
- std::memset(texelWeightData.data() + clearByteStart, 0, std::min(16U - clearByteStart, 16U));
IntegerEncodedVector texelWeightValues;
- InputBitStream weightStream(texelWeightData.data());
+ InputBitStream weightStream(texelWeightData);
DecodeIntegerSequence(texelWeightValues, weightStream, weightParams.m_MaxWeight,
weightParams.GetNumWeightValues());
diff --git a/src/video_core/vulkan_common/vulkan_device.cpp b/src/video_core/vulkan_common/vulkan_device.cpp
index 70495aba0..37d7b45a3 100644
--- a/src/video_core/vulkan_common/vulkan_device.cpp
+++ b/src/video_core/vulkan_common/vulkan_device.cpp
@@ -256,7 +256,7 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
.shaderTessellationAndGeometryPointSize = false,
.shaderImageGatherExtended = true,
.shaderStorageImageExtendedFormats = false,
- .shaderStorageImageMultisample = true,
+ .shaderStorageImageMultisample = is_shader_storage_image_multisample,
.shaderStorageImageReadWithoutFormat = is_formatless_image_load_supported,
.shaderStorageImageWriteWithoutFormat = true,
.shaderUniformBufferArrayDynamicIndexing = false,
@@ -804,6 +804,7 @@ void Device::SetupFamilies(VkSurfaceKHR surface) {
void Device::SetupFeatures() {
const auto supported_features{physical.GetFeatures()};
is_formatless_image_load_supported = supported_features.shaderStorageImageReadWithoutFormat;
+ is_shader_storage_image_multisample = supported_features.shaderStorageImageMultisample;
is_blit_depth_stencil_supported = TestDepthStencilBlits();
is_optimal_astc_supported = IsOptimalAstcSupported(supported_features);
}
diff --git a/src/video_core/vulkan_common/vulkan_device.h b/src/video_core/vulkan_common/vulkan_device.h
index a973c3ce4..4b66dba7a 100644
--- a/src/video_core/vulkan_common/vulkan_device.h
+++ b/src/video_core/vulkan_common/vulkan_device.h
@@ -272,23 +272,24 @@ private:
bool is_optimal_astc_supported{}; ///< Support for native ASTC.
bool is_float16_supported{}; ///< Support for float16 arithmetics.
bool is_warp_potentially_bigger{}; ///< Host warp size can be bigger than guest.
- bool is_formatless_image_load_supported{}; ///< Support for shader image read without format.
- bool is_blit_depth_stencil_supported{}; ///< Support for blitting from and to depth stencil.
- bool nv_viewport_swizzle{}; ///< Support for VK_NV_viewport_swizzle.
- bool khr_uniform_buffer_standard_layout{}; ///< Support for std430 on UBOs.
- bool ext_index_type_uint8{}; ///< Support for VK_EXT_index_type_uint8.
- bool ext_sampler_filter_minmax{}; ///< Support for VK_EXT_sampler_filter_minmax.
- bool ext_depth_range_unrestricted{}; ///< Support for VK_EXT_depth_range_unrestricted.
- bool ext_shader_viewport_index_layer{}; ///< Support for VK_EXT_shader_viewport_index_layer.
- bool ext_tooling_info{}; ///< Support for VK_EXT_tooling_info.
- bool ext_transform_feedback{}; ///< Support for VK_EXT_transform_feedback.
- bool ext_custom_border_color{}; ///< Support for VK_EXT_custom_border_color.
- bool ext_extended_dynamic_state{}; ///< Support for VK_EXT_extended_dynamic_state.
- bool ext_robustness2{}; ///< Support for VK_EXT_robustness2.
- bool ext_shader_stencil_export{}; ///< Support for VK_EXT_shader_stencil_export.
- bool nv_device_diagnostics_config{}; ///< Support for VK_NV_device_diagnostics_config.
- bool has_renderdoc{}; ///< Has RenderDoc attached
- bool has_nsight_graphics{}; ///< Has Nsight Graphics attached
+ bool is_formatless_image_load_supported{}; ///< Support for shader image read without format.
+ bool is_shader_storage_image_multisample{}; ///< Support for image operations on MSAA images.
+ bool is_blit_depth_stencil_supported{}; ///< Support for blitting from and to depth stencil.
+ bool nv_viewport_swizzle{}; ///< Support for VK_NV_viewport_swizzle.
+ bool khr_uniform_buffer_standard_layout{}; ///< Support for std430 on UBOs.
+ bool ext_index_type_uint8{}; ///< Support for VK_EXT_index_type_uint8.
+ bool ext_sampler_filter_minmax{}; ///< Support for VK_EXT_sampler_filter_minmax.
+ bool ext_depth_range_unrestricted{}; ///< Support for VK_EXT_depth_range_unrestricted.
+ bool ext_shader_viewport_index_layer{}; ///< Support for VK_EXT_shader_viewport_index_layer.
+ bool ext_tooling_info{}; ///< Support for VK_EXT_tooling_info.
+ bool ext_transform_feedback{}; ///< Support for VK_EXT_transform_feedback.
+ bool ext_custom_border_color{}; ///< Support for VK_EXT_custom_border_color.
+ bool ext_extended_dynamic_state{}; ///< Support for VK_EXT_extended_dynamic_state.
+ bool ext_robustness2{}; ///< Support for VK_EXT_robustness2.
+ bool ext_shader_stencil_export{}; ///< Support for VK_EXT_shader_stencil_export.
+ bool nv_device_diagnostics_config{}; ///< Support for VK_NV_device_diagnostics_config.
+ bool has_renderdoc{}; ///< Has RenderDoc attached
+ bool has_nsight_graphics{}; ///< Has Nsight Graphics attached
// Asynchronous Graphics Pipeline setting
bool use_asynchronous_shaders{}; ///< Setting to use asynchronous shaders/graphics pipeline
generated by cgit v1.2.3 (git 2.25.1) at 2024-07-15 08:24:11 +0200