Files
punktfunk/crates/pyrowave-sys/vendor/pyrowave/Granite/vulkan/shader.cpp
T
enricobuehler 4c3b11445c feat(host): vendor PyroWave + minimal Granite subset as crates/pyrowave-sys
Phase 0 of design/pyrowave-codec-plan.md — the opt-in wired-LAN ultra-low-
latency codec. Vendored at upstream 509e4f88 (API 0.4.0, Granite 44362775,
volk + vulkan-headers pins in PUNKTFUNK-VENDOR.txt), pruned to the 6.6 MB
the standalone no-renderer build needs; scripts/vendor-pyrowave.sh
reproduces the tree (a pin bump is protocol-affecting, plan §4.2).

build.rs drives the wrapper CMakeLists (static archives incl. a static
C-API lib upstream only ships shared) + bindgen over pyrowave.h; Linux and
Windows only, empty stub elsewhere (Apple gets a native Metal port, §4.7).
Offline-safe by construction: no network, no system lib, vendored Vulkan
headers — same model as the opus dep (flatpak builder has no network).

Phase-0 validation on .21 (RTX 5070 Ti, driver 610.43.03):
- upstream pyrowave-c-test + interop test (incl. dmabuf/DRM-modifier
  Vulkan<->Vulkan) pass, from the pristine AND the pruned tree
- GPU kernel times at ~1.6 bpp noise: encode/decode 0.090/0.042 ms @800p,
  0.146/0.067 @1080p, 0.226/0.103 @1440p, 0.477/0.201 @4K — order of
  magnitude under NVENC's 1-2 ms retrieve, CBR lands within ~100 B of
  target
- cargo test -p pyrowave-sys green (static link + API-version pin check)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 00:35:10 +02:00

1166 lines
40 KiB
C++

/* Copyright (c) 2017-2026 Hans-Kristian Arntzen
*
* Permission is hereby granted, free of charge, to any person obtaining
* a copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
* distribute, sublicense, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
* The above copyright notice and this permission notice shall be
* included in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
#define NOMINMAX
#include "shader.hpp"
#include "device.hpp"
#ifdef GRANITE_VULKAN_SPIRV_CROSS
#include "spirv_cross.hpp"
using namespace spirv_cross;
#endif
#ifdef HAVE_GRANITE_VULKAN_POST_MORTEM
#include "post_mortem.hpp"
#endif
using namespace Util;
namespace Vulkan
{
void ImmutableSamplerBank::hash(Util::Hasher &h, const ImmutableSamplerBank *sampler_bank)
{
h.u32(0);
if (sampler_bank)
{
unsigned index = 0;
for (auto &set : sampler_bank->samplers)
{
for (auto *binding : set)
{
if (binding)
{
h.u32(index);
h.u64(binding->get_hash());
}
index++;
}
}
}
}
// 32 bytes is a decent amount in most cases. For cases with lots of resources, just fallback to heap slice.
static constexpr uint32_t MaxInlineSizePerSet = (256 - VULKAN_PUSH_CONSTANT_SIZE) / VULKAN_NUM_DESCRIPTOR_SETS;
// Worst case we can always fall back to heap slice or indirect table for images. This requires at most one BDA.
static constexpr uint32_t MaxBufferInlineSizePerSet = MaxInlineSizePerSet - sizeof(VkDeviceAddress);
static uint32_t align(uint32_t size, uint32_t alignment)
{
return (size + alignment - 1) & ~(alignment - 1);
}
void PipelineLayout::init_heap_buffers(uint32_t set_index)
{
auto buffer_desc_size =
align(device->get_device_features().descriptor_heap_properties.bufferDescriptorSize,
device->get_device_features().descriptor_heap_properties.bufferDescriptorAlignment);
auto &push_data_offset = heap.push_data_size;
auto &desc_set = layout.sets[set_index];
auto raw_buffer_mask = desc_set.uniform_buffer_mask | desc_set.storage_buffer_mask | desc_set.rtas_mask;
uint32_t num_buffer_descriptors = 0;
bool requires_array_length_or_array = device->get_device_features().enabled_features.robustBufferAccess == VK_TRUE;
Util::for_each_bit(raw_buffer_mask, [&](unsigned bit)
{
num_buffer_descriptors += desc_set.meta[bit].array_size;
if (desc_set.meta[bit].array_size > 1)
requires_array_length_or_array = true;
});
auto required_inline_size = num_buffer_descriptors * sizeof(VkDeviceAddress);
// If we enable robustness, we cannot use PUSH_ADDRESS.
Util::for_each_bit(raw_buffer_mask, [&](unsigned bit)
{
if (desc_set.meta[bit].requires_descriptor_size)
requires_array_length_or_array = true;
});
// Raw PUSH_ADDRESS is always preferred.
if (required_inline_size <= MaxBufferInlineSizePerSet && !requires_array_length_or_array)
{
heap.buffer_strategies[set_index] = DescriptorStrategy::Inline;
if (required_inline_size)
push_data_offset = align(push_data_offset, sizeof(VkDeviceAddress));
heap.push_inline_offsets[set_index] = push_data_offset;
heap.push_inline_size[set_index] += required_inline_size;
push_data_offset += required_inline_size;
}
else if (requires_array_length_or_array)
{
heap.buffer_strategies[set_index] = DescriptorStrategy::HeapSlice;
heap.push_buffer_offsets[set_index] = push_data_offset;
// A single u32 will do.
push_data_offset += sizeof(uint32_t);
// Allocate N descriptors from the heap and write them directly.
heap.heap_slice_size[set_index] = align(heap.heap_slice_size[set_index], buffer_desc_size);
heap.heap_slice_size[set_index] += num_buffer_descriptors * buffer_desc_size;
}
else
{
// Small buffer of BDAs. Don't want to allocate from the precious heap if possible.
heap.buffer_strategies[set_index] = DescriptorStrategy::IndirectTable;
push_data_offset = align(push_data_offset, sizeof(VkDeviceAddress));
heap.push_buffer_offsets[set_index] = push_data_offset;
push_data_offset += sizeof(VkDeviceAddress);
heap.heap_table_size[set_index] += required_inline_size;
}
}
void PipelineLayout::init_heap_image(uint32_t set_index)
{
auto image_desc_size =
align(device->get_device_features().descriptor_heap_properties.imageDescriptorSize,
device->get_device_features().descriptor_heap_properties.imageDescriptorAlignment);
auto &push_data_offset = heap.push_data_size;
auto &desc_set = layout.sets[set_index];
auto image_sampler_mask =
desc_set.sampled_image_mask |
desc_set.separate_image_mask | desc_set.storage_image_mask |
desc_set.sampled_texel_buffer_mask | desc_set.storage_texel_buffer_mask |
desc_set.input_attachment_mask |
desc_set.sampler_mask;
auto sampler_mask = desc_set.sampled_image_mask | desc_set.sampler_mask;
uint32_t num_image_descriptors = 0;
Util::for_each_bit(image_sampler_mask, [&](unsigned bit)
{
num_image_descriptors += desc_set.meta[bit].array_size;
});
bool requires_array_of_image = false;
Util::for_each_bit(image_sampler_mask, [&](unsigned bit)
{
if (desc_set.meta[bit].array_size > 1)
requires_array_of_image = true;
});
uint32_t available_inline_indices = (MaxInlineSizePerSet - heap.push_inline_size[set_index]) / sizeof(uint32_t);
// Array of resources would need either heap slice or indirection table.
if (num_image_descriptors <= available_inline_indices && !requires_array_of_image)
{
heap.image_strategies[set_index] = DescriptorStrategy::Inline;
if (heap.buffer_strategies[set_index] != DescriptorStrategy::Inline)
heap.push_inline_offsets[set_index] = push_data_offset;
heap.push_inline_size[set_index] += num_image_descriptors * sizeof(uint32_t);
push_data_offset += num_image_descriptors * sizeof(uint32_t);
}
else if (sampler_mask != 0 && (layout.bindless_descriptor_set_mask & (1u << set_index)) == 0)
{
// We cannot lower sampler to heap slice since sampler heap is so tiny.
// Force indirection table.
// TODO: It's in theory possible to split this up
// so that samplers are push index inlined while everything else is heap sliced.
// This isn't ideal, but what can you do.
heap.image_strategies[set_index] = DescriptorStrategy::IndirectTable;
// Buffers and images can share the same indirection table.
if (heap.buffer_strategies[set_index] == DescriptorStrategy::IndirectTable)
{
heap.push_image_offsets[set_index] = heap.push_buffer_offsets[set_index];
}
else
{
push_data_offset = align(push_data_offset, sizeof(VkDeviceAddress));
heap.push_image_offsets[set_index] = push_data_offset;
push_data_offset += sizeof(VkDeviceAddress);
}
// Buffers go first, for alignment purposes.
heap.heap_table_size[set_index] += num_image_descriptors * sizeof(uint32_t);
}
else
{
heap.image_strategies[set_index] = DescriptorStrategy::HeapSlice;
if (heap.buffer_strategies[set_index] == DescriptorStrategy::HeapSlice)
{
heap.push_image_offsets[set_index] = heap.push_buffer_offsets[set_index];
}
else
{
heap.push_image_offsets[set_index] = push_data_offset;
push_data_offset += sizeof(uint32_t);
}
if ((layout.bindless_descriptor_set_mask & (1u << set_index)) == 0)
{
// Allocate N descriptors from the heap and write them directly.
heap.heap_slice_size[set_index] = align(heap.heap_slice_size[set_index], image_desc_size);
heap.heap_slice_size[set_index] += num_image_descriptors * image_desc_size;
}
}
}
void PipelineLayout::init_heap_offsets(uint32_t set_index)
{
auto buffer_desc_size =
align(device->get_device_features().descriptor_heap_properties.bufferDescriptorSize,
device->get_device_features().descriptor_heap_properties.bufferDescriptorAlignment);
auto image_desc_size =
align(device->get_device_features().descriptor_heap_properties.imageDescriptorSize,
device->get_device_features().descriptor_heap_properties.imageDescriptorAlignment);
auto sampler_desc_size =
align(device->get_device_features().descriptor_heap_properties.samplerDescriptorSize,
device->get_device_features().descriptor_heap_properties.samplerDescriptorAlignment);
auto &desc_set = layout.sets[set_index];
auto image_sampler_mask =
desc_set.sampled_image_mask |
desc_set.separate_image_mask | desc_set.storage_image_mask |
desc_set.sampled_texel_buffer_mask | desc_set.storage_texel_buffer_mask |
desc_set.input_attachment_mask |
desc_set.sampler_mask;
auto buffer_mask = desc_set.uniform_buffer_mask | desc_set.storage_buffer_mask | desc_set.rtas_mask;
uint32_t push_offset = 0;
uint32_t table_offset = 0;
uint32_t slice_offset = 0;
VkDescriptorSetAndBindingMappingEXT buffer_template = { VK_STRUCTURE_TYPE_DESCRIPTOR_SET_AND_BINDING_MAPPING_EXT };
buffer_template.resourceMask = VK_SPIRV_RESOURCE_TYPE_UNIFORM_BUFFER_BIT_EXT |
VK_SPIRV_RESOURCE_TYPE_READ_WRITE_STORAGE_BUFFER_BIT_EXT |
VK_SPIRV_RESOURCE_TYPE_READ_ONLY_STORAGE_BUFFER_BIT_EXT;
if (device->get_device_features().rtas_features.accelerationStructure)
buffer_template.resourceMask |= VK_SPIRV_RESOURCE_TYPE_ACCELERATION_STRUCTURE_BIT_EXT;
VkDescriptorSetAndBindingMappingEXT image_template = { VK_STRUCTURE_TYPE_DESCRIPTOR_SET_AND_BINDING_MAPPING_EXT };
image_template.resourceMask = VK_SPIRV_RESOURCE_TYPE_SAMPLED_IMAGE_BIT_EXT |
VK_SPIRV_RESOURCE_TYPE_READ_WRITE_STORAGE_BUFFER_BIT_EXT |
VK_SPIRV_RESOURCE_TYPE_READ_ONLY_STORAGE_BUFFER_BIT_EXT |
VK_SPIRV_RESOURCE_TYPE_READ_ONLY_IMAGE_BIT_EXT |
VK_SPIRV_RESOURCE_TYPE_READ_WRITE_IMAGE_BIT_EXT;
switch (heap.buffer_strategies[set_index])
{
case DescriptorStrategy::Inline:
Util::for_each_bit(buffer_mask, [&](unsigned bit)
{
heap.desc_offsets[set_index][bit] = push_offset;
auto mapping = buffer_template;
mapping.descriptorSet = set_index;
mapping.firstBinding = bit;
VK_ASSERT(desc_set.meta[bit].array_size == 1);
mapping.bindingCount = 1;
mapping.source = VK_DESCRIPTOR_MAPPING_SOURCE_PUSH_ADDRESS_EXT;
mapping.sourceData.pushAddressOffset = push_offset + heap.push_inline_offsets[set_index];
push_offset += sizeof(VkDeviceAddress);
heap.mappings.push_back(mapping);
});
break;
case DescriptorStrategy::HeapSlice:
Util::for_each_bit(buffer_mask, [&](unsigned bit)
{
slice_offset = align(slice_offset, buffer_desc_size);
auto mapping = buffer_template;
mapping.descriptorSet = set_index;
mapping.firstBinding = bit;
mapping.bindingCount = desc_set.meta[bit].array_size;
mapping.source = VK_DESCRIPTOR_MAPPING_SOURCE_HEAP_WITH_PUSH_INDEX_EXT;
mapping.sourceData.pushIndex.pushOffset = heap.push_buffer_offsets[set_index];
mapping.sourceData.pushIndex.heapArrayStride = buffer_desc_size;
mapping.sourceData.pushIndex.heapIndexStride = device->get_device_features().resource_heap_resource_desc_size;
mapping.sourceData.pushIndex.heapOffset = slice_offset;
for (unsigned i = 0; i < mapping.bindingCount; i++)
{
heap.desc_offsets[set_index][bit + i] = slice_offset;
slice_offset += buffer_desc_size;
}
heap.mappings.push_back(mapping);
});
break;
case DescriptorStrategy::IndirectTable:
Util::for_each_bit(buffer_mask, [&](unsigned bit)
{
table_offset = align(table_offset, sizeof(VkDeviceAddress));
heap.desc_offsets[set_index][bit] = table_offset;
auto mapping = buffer_template;
mapping.descriptorSet = set_index;
mapping.firstBinding = bit;
VK_ASSERT(desc_set.meta[bit].array_size == 1);
mapping.bindingCount = 1;
mapping.source = VK_DESCRIPTOR_MAPPING_SOURCE_INDIRECT_ADDRESS_EXT;
mapping.sourceData.indirectAddress.pushOffset = heap.push_buffer_offsets[set_index];
mapping.sourceData.indirectAddress.addressOffset = table_offset;
table_offset += sizeof(VkDeviceAddress);
heap.mappings.push_back(mapping);
});
break;
default:
break;
}
bool bindless = (layout.bindless_descriptor_set_mask & (1u << set_index)) != 0;
VK_ASSERT(!bindless || slice_offset == 0);
switch (heap.image_strategies[set_index])
{
case DescriptorStrategy::Inline:
Util::for_each_bit(image_sampler_mask, [&](unsigned bit)
{
heap.desc_offsets[set_index][bit] = push_offset;
auto mapping = image_template;
mapping.descriptorSet = set_index;
mapping.firstBinding = bit;
VK_ASSERT(desc_set.meta[bit].array_size == 1);
mapping.bindingCount = 1;
mapping.resourceMask |= VK_SPIRV_RESOURCE_TYPE_COMBINED_SAMPLED_IMAGE_BIT_EXT;
mapping.source = VK_DESCRIPTOR_MAPPING_SOURCE_HEAP_WITH_PUSH_INDEX_EXT;
mapping.sourceData.pushIndex.pushOffset = push_offset + heap.push_inline_offsets[set_index];
mapping.sourceData.pushIndex.heapOffset = 0;
mapping.sourceData.pushIndex.heapArrayStride = device->get_device_features().resource_heap_resource_desc_size;
mapping.sourceData.pushIndex.heapIndexStride = device->get_device_features().resource_heap_resource_desc_size;
if ((desc_set.sampled_image_mask & (1u << bit)) != 0)
{
mapping.sourceData.pushIndex.useCombinedImageSamplerIndex = VK_TRUE;
mapping.sourceData.pushIndex.samplerHeapArrayStride = sampler_desc_size;
mapping.sourceData.pushIndex.samplerHeapIndexStride = sampler_desc_size;
}
if ((desc_set.sampler_mask & (1u << bit)) == 0)
heap.mappings.push_back(mapping);
mapping.resourceMask = VK_SPIRV_RESOURCE_TYPE_SAMPLER_BIT_EXT;
mapping.sourceData.pushIndex = {};
mapping.sourceData.pushIndex.pushOffset = push_offset + heap.push_inline_offsets[set_index];
mapping.sourceData.pushIndex.heapArrayStride = sampler_desc_size;
mapping.sourceData.pushIndex.heapIndexStride = sampler_desc_size;
if ((desc_set.sampler_mask & (1u << bit)) != 0)
heap.mappings.push_back(mapping);
push_offset += sizeof(uint32_t);
});
break;
case DescriptorStrategy::HeapSlice:
Util::for_each_bit(image_sampler_mask, [&](unsigned bit)
{
slice_offset = align(slice_offset, image_desc_size);
auto mapping = image_template;
mapping.descriptorSet = set_index;
mapping.firstBinding = bit;
mapping.bindingCount = bindless ? 1 : desc_set.meta[bit].array_size;
mapping.source = VK_DESCRIPTOR_MAPPING_SOURCE_HEAP_WITH_PUSH_INDEX_EXT;
// HeapSlice is not compatible with sampler and combined image sampler.
mapping.sourceData.pushIndex.pushOffset = heap.push_image_offsets[set_index];
mapping.sourceData.pushIndex.heapArrayStride = image_desc_size;
mapping.sourceData.pushIndex.heapIndexStride = device->get_device_features().resource_heap_resource_desc_size;
mapping.sourceData.pushIndex.heapOffset = slice_offset;
if (!bindless)
{
for (unsigned i = 0; i < mapping.bindingCount; i++)
{
heap.desc_offsets[set_index][bit + i] = slice_offset;
slice_offset += image_desc_size;
}
}
heap.mappings.push_back(mapping);
});
break;
case DescriptorStrategy::IndirectTable:
Util::for_each_bit(image_sampler_mask, [&](unsigned bit)
{
table_offset = align(table_offset, sizeof(uint32_t));
heap.desc_offsets[set_index][bit] = table_offset;
auto mapping = image_template;
mapping.descriptorSet = set_index;
mapping.firstBinding = bit;
mapping.bindingCount = desc_set.meta[bit].array_size;
mapping.source = VK_DESCRIPTOR_MAPPING_SOURCE_HEAP_WITH_INDIRECT_INDEX_ARRAY_EXT;
mapping.resourceMask |= VK_SPIRV_RESOURCE_TYPE_COMBINED_SAMPLED_IMAGE_BIT_EXT;
mapping.sourceData.indirectIndexArray.pushOffset = heap.push_image_offsets[set_index];
mapping.sourceData.indirectIndexArray.heapOffset = 0;
mapping.sourceData.indirectIndexArray.samplerHeapOffset = 0;
mapping.sourceData.indirectIndexArray.addressOffset = table_offset;
mapping.sourceData.indirectIndexArray.heapIndexStride = device->get_device_features().resource_heap_resource_desc_size;
if ((desc_set.sampled_image_mask & (1u << bit)) != 0)
{
mapping.sourceData.indirectIndexArray.useCombinedImageSamplerIndex = VK_TRUE;
mapping.sourceData.indirectIndexArray.samplerHeapIndexStride = sampler_desc_size;
}
if ((desc_set.sampler_mask & (1u << bit)) == 0)
heap.mappings.push_back(mapping);
mapping.resourceMask = VK_SPIRV_RESOURCE_TYPE_SAMPLER_BIT_EXT;
mapping.sourceData.indirectIndexArray = {};
mapping.sourceData.indirectIndexArray.pushOffset = heap.push_image_offsets[set_index];
mapping.sourceData.indirectIndexArray.heapOffset = 0;
mapping.sourceData.indirectIndexArray.addressOffset = table_offset;
mapping.sourceData.indirectIndexArray.heapIndexStride = sampler_desc_size;
if ((desc_set.sampler_mask & (1u << bit)) != 0)
heap.mappings.push_back(mapping);
for (unsigned i = 0; i < mapping.bindingCount; i++)
{
heap.desc_offsets[set_index][bit + i] = table_offset;
table_offset += sizeof(uint32_t);
}
});
break;
default:
break;
}
VK_ASSERT(push_offset <= MaxInlineSizePerSet);
VK_ASSERT(push_offset == heap.push_inline_size[set_index]);
VK_ASSERT(table_offset == heap.heap_table_size[set_index]);
VK_ASSERT(slice_offset == heap.heap_slice_size[set_index]);
}
void PipelineLayout::init_heap(uint32_t set_index)
{
init_heap_buffers(set_index);
init_heap_image(set_index);
init_heap_offsets(set_index);
}
void PipelineLayout::init_heap()
{
uint32_t push_data_offset = layout.push_constant_range.offset + layout.push_constant_range.size;
heap.push_data_size = push_data_offset;
for (unsigned i = 0; i < VULKAN_NUM_DESCRIPTOR_SETS; i++)
{
if ((layout.descriptor_set_mask & (1u << i)) != 0)
{
init_heap(i);
// Only used to track when we need to invalidate sets.
set_allocators[i] = device->request_descriptor_set_allocator(
layout.sets[i], layout.stages_for_bindings[i], nullptr);
}
}
VK_ASSERT(heap.push_data_size <= VULKAN_PUSH_DATA_SIZE);
}
void PipelineLayout::init_legacy(const ImmutableSamplerBank *immutable_samplers)
{
VkDescriptorSetLayout layouts[VULKAN_NUM_DESCRIPTOR_SETS] = {};
unsigned num_sets = 0;
for (unsigned i = 0; i < VULKAN_NUM_DESCRIPTOR_SETS; i++)
{
set_allocators[i] = device->request_descriptor_set_allocator(layout.sets[i], layout.stages_for_bindings[i],
immutable_samplers ? immutable_samplers->samplers[i] : nullptr);
layouts[i] = set_allocators[i]->get_layout_for_pool();
if (layout.descriptor_set_mask & (1u << i))
{
num_sets = i + 1;
// Assume the last set index in layout is the highest frequency update one, make that push descriptor if possible.
// Only one descriptor set can be push descriptor.
bool has_push_layout = set_allocators[i]->get_layout_for_push() != VK_NULL_HANDLE;
if (has_push_layout)
push_set_index = i;
}
}
if (push_set_index != UINT32_MAX)
layouts[push_set_index] = set_allocators[push_set_index]->get_layout_for_push();
if (num_sets > VULKAN_NUM_DESCRIPTOR_SETS)
LOGE("Number of sets %u exceeds limit of %u.\n", num_sets, VULKAN_NUM_DESCRIPTOR_SETS);
VkPipelineLayoutCreateInfo info = { VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO };
if (num_sets)
{
info.setLayoutCount = num_sets;
info.pSetLayouts = layouts;
}
if (layout.push_constant_range.stageFlags != 0)
{
info.pushConstantRangeCount = 1;
info.pPushConstantRanges = &layout.push_constant_range;
}
#ifdef VULKAN_DEBUG
LOGI("Creating pipeline layout.\n");
#endif
auto &table = device->get_device_table();
if (table.vkCreatePipelineLayout(device->get_device(), &info, nullptr, &pipe_layout) != VK_SUCCESS)
LOGE("Failed to create pipeline layout.\n");
#ifdef GRANITE_VULKAN_FOSSILIZE
device->register_pipeline_layout(pipe_layout, get_hash(), info);
#endif
if (!device->get_device_features().descriptor_buffer_features.descriptorBuffer)
create_update_templates();
}
PipelineLayout::PipelineLayout(Hash hash, Device *device_, const CombinedResourceLayout &layout_,
const ImmutableSamplerBank *immutable_samplers)
: IntrusiveHashMapEnabled<PipelineLayout>(hash)
, device(device_)
, layout(layout_)
{
if (device->get_device_features().descriptor_heap_features.descriptorHeap)
init_heap();
else
init_legacy(immutable_samplers);
}
void PipelineLayout::create_update_templates()
{
auto &table = device->get_device_table();
for (unsigned desc_set = 0; desc_set < VULKAN_NUM_DESCRIPTOR_SETS; desc_set++)
{
if ((layout.descriptor_set_mask & (1u << desc_set)) == 0)
continue;
if ((layout.bindless_descriptor_set_mask & (1u << desc_set)) != 0)
continue;
VkDescriptorUpdateTemplateEntry update_entries[VULKAN_NUM_BINDINGS];
uint32_t update_count = 0;
auto &set_layout = layout.sets[desc_set];
for_each_bit(set_layout.uniform_buffer_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
// Work around a RenderDoc capture bug where descriptorCount > 1 is not handled correctly.
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, buffer) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.storage_buffer_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, buffer) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.rtas_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_ACCELERATION_STRUCTURE_KHR;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, rtas) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.sampled_texel_buffer_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, buffer_view.handle) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.storage_texel_buffer_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, buffer_view.handle) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.sampled_image_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
if (set_layout.fp_mask & (1u << binding))
entry.offset = offsetof(ResourceBinding, image.fp) + sizeof(ResourceBinding) * (binding + i);
else
entry.offset = offsetof(ResourceBinding, image.integer) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.separate_image_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
if (set_layout.fp_mask & (1u << binding))
entry.offset = offsetof(ResourceBinding, image.fp) + sizeof(ResourceBinding) * (binding + i);
else
entry.offset = offsetof(ResourceBinding, image.integer) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.sampler_mask & ~set_layout.immutable_sampler_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_SAMPLER;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, image.fp) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.storage_image_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
entry.offset = offsetof(ResourceBinding, image.fp) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
for_each_bit(set_layout.input_attachment_mask, [&](uint32_t binding) {
unsigned array_size = set_layout.meta[binding].array_size;
VK_ASSERT(update_count < VULKAN_NUM_BINDINGS);
for (unsigned i = 0; i < array_size; i++)
{
auto &entry = update_entries[update_count++];
entry.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
entry.dstBinding = binding;
entry.dstArrayElement = i;
entry.descriptorCount = 1;
if (set_layout.fp_mask & (1u << binding))
entry.offset = offsetof(ResourceBinding, image.fp) + sizeof(ResourceBinding) * (binding + i);
else
entry.offset = offsetof(ResourceBinding, image.integer) + sizeof(ResourceBinding) * (binding + i);
entry.stride = sizeof(ResourceBinding);
}
});
VkDescriptorUpdateTemplateCreateInfo info = { VK_STRUCTURE_TYPE_DESCRIPTOR_UPDATE_TEMPLATE_CREATE_INFO };
info.pipelineLayout = pipe_layout;
if (desc_set == push_set_index)
{
info.descriptorSetLayout = set_allocators[desc_set]->get_layout_for_push();
info.templateType = VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS;
}
else
{
info.descriptorSetLayout = set_allocators[desc_set]->get_layout_for_pool();
info.templateType = VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET;
}
info.set = desc_set;
info.descriptorUpdateEntryCount = update_count;
info.pDescriptorUpdateEntries = update_entries;
info.pipelineBindPoint = (layout.stages_for_sets[desc_set] & VK_SHADER_STAGE_COMPUTE_BIT) ?
VK_PIPELINE_BIND_POINT_COMPUTE : VK_PIPELINE_BIND_POINT_GRAPHICS;
if (table.vkCreateDescriptorUpdateTemplate(device->get_device(), &info, nullptr,
&update_template[desc_set]) != VK_SUCCESS)
{
LOGE("Failed to create descriptor update template.\n");
}
}
}
PipelineLayout::~PipelineLayout()
{
auto &table = device->get_device_table();
if (pipe_layout != VK_NULL_HANDLE)
table.vkDestroyPipelineLayout(device->get_device(), pipe_layout, nullptr);
for (auto &update : update_template)
if (update != VK_NULL_HANDLE)
table.vkDestroyDescriptorUpdateTemplate(device->get_device(), update, nullptr);
}
const char *Shader::stage_to_name(ShaderStage stage)
{
switch (stage)
{
case ShaderStage::Compute:
return "compute";
case ShaderStage::Vertex:
return "vertex";
case ShaderStage::Fragment:
return "fragment";
case ShaderStage::Task:
return "task";
case ShaderStage::Mesh:
return "mesh";
default:
return "unknown";
}
}
// Implicitly also checks for endian issues.
static const uint16_t reflection_magic[] = { 'G', 'R', 'A', ResourceLayout::Version };
size_t ResourceLayout::serialization_size()
{
return sizeof(ResourceLayout) + sizeof(reflection_magic);
}
bool ResourceLayout::serialize(uint8_t *data, size_t size) const
{
if (size != serialization_size())
return false;
// Cannot serialize externally defined immutable samplers.
for (auto &set : sets)
if (set.immutable_sampler_mask != 0)
return false;
memcpy(data, reflection_magic, sizeof(reflection_magic));
memcpy(data + sizeof(reflection_magic), this, sizeof(*this));
return true;
}
bool ResourceLayout::unserialize(const uint8_t *data, size_t size)
{
if (size != sizeof(*this) + sizeof(reflection_magic))
{
LOGE("Reflection size mismatch.\n");
return false;
}
if (memcmp(data, reflection_magic, sizeof(reflection_magic)) != 0)
{
LOGE("Magic mismatch.\n");
return false;
}
memcpy(this, data + sizeof(reflection_magic), sizeof(*this));
return true;
}
Util::Hash Shader::hash(const uint32_t *data, size_t size)
{
Util::Hasher hasher;
hasher.data(data, size);
return hasher.get();
}
#ifdef GRANITE_VULKAN_SPIRV_CROSS
static void update_array_info(ResourceLayout &layout, const SPIRType &type, unsigned set, unsigned binding)
{
auto &meta = layout.sets[set].meta[binding];
if (!type.array.empty())
{
if (type.array.size() != 1)
LOGE("Array dimension must be 1.\n");
else if (!type.array_size_literal.front())
LOGE("Array dimension must be a literal.\n");
else
{
if (type.array.front() == 0)
{
if (binding != 0)
LOGE("Bindless textures can only be used with binding = 0 in a set.\n");
if (type.basetype != SPIRType::Image || type.image.dim == spv::DimBuffer)
{
LOGE("Can only use bindless for sampled images.\n");
}
else
{
layout.bindless_set_mask |= 1u << set;
// Ignore fp_mask for bindless since we can mix and match.
layout.sets[set].fp_mask = 0;
}
meta.array_size = DescriptorSetLayout::UNSIZED_ARRAY;
}
else if (meta.array_size && meta.array_size != type.array.front())
LOGE("Array dimension for (%u, %u) is inconsistent.\n", set, binding);
else if (type.array.front() + binding > VULKAN_NUM_BINDINGS)
LOGE("Binding array will go out of bounds.\n");
else
meta.array_size = uint8_t(type.array.front());
}
}
else
{
if (meta.array_size && meta.array_size != 1)
LOGE("Array dimension for (%u, %u) is inconsistent.\n", set, binding);
meta.array_size = 1;
}
}
bool Shader::reflect_resource_layout(ResourceLayout &layout, const uint32_t *data, size_t size)
{
Compiler compiler(data, size / sizeof(uint32_t));
#ifdef VULKAN_DEBUG
LOGI("Reflecting shader layout.\n");
#endif
bool has_array_length = false;
auto &ir = compiler.get_ir();
ir.for_each_typed_id<SPIRBlock>([&](uint32_t, const SPIRBlock &block)
{
if (has_array_length)
return;
for (auto &op : block.ops)
{
auto spvop = spv::Op(op.op);
if (spvop == spv::OpArrayLength)
{
has_array_length = true;
return;
}
}
});
auto resources = compiler.get_shader_resources();
for (auto &image : resources.sampled_images)
{
auto set = compiler.get_decoration(image.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(image.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
auto &type = compiler.get_type(image.type_id);
if (type.image.dim == spv::DimBuffer)
layout.sets[set].sampled_texel_buffer_mask |= 1u << binding;
else
layout.sets[set].sampled_image_mask |= 1u << binding;
if (compiler.get_type(type.image.type).basetype == SPIRType::BaseType::Float)
layout.sets[set].fp_mask |= 1u << binding;
update_array_info(layout, type, set, binding);
}
for (auto &image : resources.subpass_inputs)
{
auto set = compiler.get_decoration(image.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(image.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
layout.sets[set].input_attachment_mask |= 1u << binding;
auto &type = compiler.get_type(image.type_id);
if (compiler.get_type(type.image.type).basetype == SPIRType::BaseType::Float)
layout.sets[set].fp_mask |= 1u << binding;
update_array_info(layout, type, set, binding);
}
for (auto &image : resources.separate_images)
{
auto set = compiler.get_decoration(image.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(image.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
auto &type = compiler.get_type(image.type_id);
if (compiler.get_type(type.image.type).basetype == SPIRType::BaseType::Float)
layout.sets[set].fp_mask |= 1u << binding;
if (type.image.dim == spv::DimBuffer)
layout.sets[set].sampled_texel_buffer_mask |= 1u << binding;
else
layout.sets[set].separate_image_mask |= 1u << binding;
update_array_info(layout, type, set, binding);
}
for (auto &image : resources.separate_samplers)
{
auto set = compiler.get_decoration(image.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(image.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
layout.sets[set].sampler_mask |= 1u << binding;
update_array_info(layout, compiler.get_type(image.type_id), set, binding);
}
for (auto &image : resources.storage_images)
{
auto set = compiler.get_decoration(image.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(image.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
auto &type = compiler.get_type(image.type_id);
if (type.image.dim == spv::DimBuffer)
layout.sets[set].storage_texel_buffer_mask |= 1u << binding;
else
layout.sets[set].storage_image_mask |= 1u << binding;
if (compiler.get_type(type.image.type).basetype == SPIRType::BaseType::Float)
layout.sets[set].fp_mask |= 1u << binding;
update_array_info(layout, type, set, binding);
}
for (auto &buffer : resources.uniform_buffers)
{
auto set = compiler.get_decoration(buffer.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(buffer.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
layout.sets[set].uniform_buffer_mask |= 1u << binding;
update_array_info(layout, compiler.get_type(buffer.type_id), set, binding);
}
for (auto &buffer : resources.storage_buffers)
{
auto set = compiler.get_decoration(buffer.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(buffer.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
layout.sets[set].storage_buffer_mask |= 1u << binding;
update_array_info(layout, compiler.get_type(buffer.type_id), set, binding);
if (has_array_length)
layout.sets[set].meta[binding].requires_descriptor_size = 1;
}
for (auto &buffer : resources.acceleration_structures)
{
auto set = compiler.get_decoration(buffer.id, spv::DecorationDescriptorSet);
auto binding = compiler.get_decoration(buffer.id, spv::DecorationBinding);
VK_ASSERT(set < VULKAN_NUM_DESCRIPTOR_SETS);
VK_ASSERT(binding < VULKAN_NUM_BINDINGS);
layout.sets[set].rtas_mask |= 1u << binding;
update_array_info(layout, compiler.get_type(buffer.type_id), set, binding);
}
for (auto &attrib : resources.stage_inputs)
{
auto location = compiler.get_decoration(attrib.id, spv::DecorationLocation);
layout.input_mask |= 1u << location;
}
for (auto &attrib : resources.stage_outputs)
{
auto location = compiler.get_decoration(attrib.id, spv::DecorationLocation);
layout.output_mask |= 1u << location;
}
if (!resources.push_constant_buffers.empty())
{
// Don't bother trying to extract which part of a push constant block we're using.
// Just assume we're accessing everything. At least on older validation layers,
// it did not do a static analysis to determine similar information, so we got a lot
// of false positives.
layout.push_constant_size =
compiler.get_declared_struct_size(compiler.get_type(resources.push_constant_buffers.front().base_type_id));
}
auto spec_constants = compiler.get_specialization_constants();
for (auto &c : spec_constants)
{
if (c.constant_id >= VULKAN_NUM_TOTAL_SPEC_CONSTANTS)
{
LOGE("Spec constant ID: %u is out of range, will be ignored.\n", c.constant_id);
continue;
}
layout.spec_constant_mask |= 1u << c.constant_id;
}
return true;
}
#else
bool Shader::reflect_resource_layout(ResourceLayout &, const uint32_t *, size_t)
{
return false;
}
#endif
Shader::Shader(Hash hash, Device *device_, const uint32_t *data, size_t size,
const ResourceLayout *resource_layout)
: IntrusiveHashMapEnabled<Shader>(hash)
, device(device_)
{
VkShaderModuleCreateInfo info = { VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO };
info.codeSize = size;
info.pCode = data;
#ifdef VULKAN_DEBUG
LOGI("Creating shader module.\n");
#endif
auto &table = device->get_device_table();
if (table.vkCreateShaderModule(device->get_device(), &info, nullptr, &module) != VK_SUCCESS)
LOGE("Failed to create shader module.\n");
#ifdef HAVE_GRANITE_VULKAN_POST_MORTEM
PostMortem::register_shader(data, size);
#endif
#ifdef GRANITE_VULKAN_FOSSILIZE
device->register_shader_module(module, get_hash(), info);
#endif
if (resource_layout)
layout = *resource_layout;
#ifdef GRANITE_VULKAN_SPIRV_CROSS
else if (!reflect_resource_layout(layout, data, size))
LOGE("Failed to reflect resource layout.\n");
#endif
if (layout.bindless_set_mask != 0 && !device->get_device_features().vk12_features.descriptorIndexing)
LOGE("Sufficient features for descriptor indexing is not supported on this device.\n");
}
Shader::~Shader()
{
auto &table = device->get_device_table();
if (module)
table.vkDestroyShaderModule(device->get_device(), module, nullptr);
}
void Program::set_shader(ShaderStage stage, Shader *handle)
{
shaders[Util::ecast(stage)] = handle;
}
Program::Program(Device *device_, Shader *vertex, Shader *fragment, const ImmutableSamplerBank *sampler_bank)
: device(device_)
{
set_shader(ShaderStage::Vertex, vertex);
set_shader(ShaderStage::Fragment, fragment);
device->bake_program(*this, sampler_bank);
}
Program::Program(Device *device_, Shader *task, Shader *mesh, Shader *fragment, const ImmutableSamplerBank *sampler_bank)
: device(device_)
{
if (task)
set_shader(ShaderStage::Task, task);
set_shader(ShaderStage::Mesh, mesh);
set_shader(ShaderStage::Fragment, fragment);
device->bake_program(*this, sampler_bank);
}
Program::Program(Device *device_, Shader *compute_shader, const ImmutableSamplerBank *sampler_bank)
: device(device_)
{
set_shader(ShaderStage::Compute, compute_shader);
device->bake_program(*this, sampler_bank);
}
Pipeline Program::get_pipeline(Hash hash) const
{
auto *ret = pipelines.find(hash);
return ret ? ret->get() : Pipeline{};
}
Pipeline Program::add_pipeline(Hash hash, const Pipeline &pipeline)
{
return pipelines.emplace_yield(hash, pipeline)->get();
}
void Program::destroy_pipeline(const Pipeline &pipeline)
{
device->get_device_table().vkDestroyPipeline(device->get_device(), pipeline.pipeline, nullptr);
}
void Program::promote_read_write_to_read_only()
{
pipelines.move_to_read_only();
}
Program::~Program()
{
for (auto &pipe : pipelines.get_read_only())
destroy_pipeline(pipe.get());
for (auto &pipe : pipelines.get_read_write())
destroy_pipeline(pipe.get());
}
}