Skip to content

Speed up device-to-host staging reads (#22023) #22023

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
meta-codesync merged 1 commit into from
Aug 24, 2026
Merged

Speed up device-to-host staging reads (#22023) #22023

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion backends/vulkan/runtime/api/containers/StagingBuffer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -178,7 +178,7 @@ void StagingBuffer::cast_float_to_half_and_copy_to(
vulkan_buffer_.vma_allocator(),
vulkan_buffer_.allocation(),
0u,
VK_WHOLE_SIZE);
numel * sizeof(float));
const float* src = reinterpret_cast<const float*>(data());
for (size_t i = 0; i < numel; ++i) {
dst[i] = float_to_half(src[i]);
Expand Down
4 changes: 2 additions & 2 deletions backends/vulkan/runtime/api/containers/StagingBuffer.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -107,7 +107,7 @@ class StagingBuffer final {
vulkan_buffer_.vma_allocator(),
vulkan_buffer_.allocation(),
0u,
VK_WHOLE_SIZE);
nbytes);
memcpy(dst, data(), nbytes);
}

Expand All @@ -118,7 +118,7 @@ class StagingBuffer final {
vulkan_buffer_.vma_allocator(),
vulkan_buffer_.allocation(),
0u,
VK_WHOLE_SIZE);
numel * sizeof(SRC_T));
const SRC_T* src = reinterpret_cast<const SRC_T*>(data());
for (size_t i = 0; i < numel; ++i) {
dst[i] = static_cast<DST_T>(src[i]);
Expand Down
4 changes: 2 additions & 2 deletions backends/vulkan/runtime/vk_api/memory/Allocator.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,13 +177,13 @@ VulkanBuffer Allocator::create_staging_buffer(
if (direction == CopyDirection::HOST_TO_DEVICE) {
alloc_create_info.flags |=
VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
alloc_create_info.preferredFlags = VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
} else {
alloc_create_info.flags |= allocation_strategy_device_to_host_;
alloc_create_info.preferredFlags = VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
}
alloc_create_info.usage = VMA_MEMORY_USAGE_AUTO_PREFER_HOST;
alloc_create_info.requiredFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
alloc_create_info.preferredFlags =
VK_MEMORY_PROPERTY_HOST_COHERENT_BIT | VK_MEMORY_PROPERTY_HOST_CACHED_BIT;

return VulkanBuffer(allocator_, size, alloc_create_info, buffer_usage);
}
Expand Down
33 changes: 33 additions & 0 deletions backends/vulkan/test/vulkan_compute_api_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,39 @@ TEST_F(VulkanComputeAPITest, print_adapter) {
std::cout << *(context()->adapter_ptr()) << std::endl;
}

TEST_F(VulkanComputeAPITest, device_to_host_staging_prefers_cached_memory) {
vkapi::VulkanBuffer staging_buffer =
context()->adapter_ptr()->vma().create_staging_buffer(
4096, vkapi::CopyDirection::DEVICE_TO_HOST);
const VmaAllocator allocator = staging_buffer.vma_allocator();
ASSERT_NE(allocator, VK_NULL_HANDLE);

const VkPhysicalDeviceMemoryProperties* memory_properties = nullptr;
vmaGetMemoryProperties(allocator, &memory_properties);
ASSERT_NE(memory_properties, nullptr);

bool has_host_cached_memory = false;
for (uint32_t i = 0; i < memory_properties->memoryTypeCount; ++i) {
const VkMemoryPropertyFlags flags =
memory_properties->memoryTypes[i].propertyFlags;
has_host_cached_memory = has_host_cached_memory ||
((flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) &&
(flags & VK_MEMORY_PROPERTY_HOST_CACHED_BIT));
}
if (!has_host_cached_memory) {
GTEST_SKIP() << "Device does not expose host-cached visible memory";
}

const VmaAllocation allocation = staging_buffer.allocation();
ASSERT_NE(allocation, VK_NULL_HANDLE);

VkMemoryPropertyFlags selected_flags = 0;
vmaGetAllocationMemoryProperties(allocator, allocation, &selected_flags);

EXPECT_TRUE(selected_flags & VK_MEMORY_PROPERTY_HOST_CACHED_BIT)
<< "Selected memory with flags " << selected_flags;
}

#if defined(VK_KHR_pipeline_executable_properties) && \
defined(ETVK_INSPECT_PIPELINES)

Expand Down
Loading