GPU compute API abstraction on top of WebGPU #3238

daljit46 · 2025-11-26T13:01:00Z

This PR supersedes #3096.

It introduces a new GPU compute abstraction built on top of WebGPU. See #3096 for general motivation and design philosophy. A notable change from that PR is now shaders are now required to be written in Slang, a new programming language created by NVidia (now under the umbrella of the Khronos group). This choice has been motivated by the fact that Slang provides many useful features like modules and generics for better code reusability and modularisation.

Some issues still need to be resolved (in future PRs):

Deal with the issue raised in Add option for canonical direct I/O layout #3108.
Add more tests, especially for the upload/download of MR::Image instances to/from GPU (currently the API only supports MR::Image<float>.
Create a real-world example that illustrates the use of the API.

One additional thing that this PR introduces is the addition of a new tcb::span class (see #3219); however, unlike for other third-party dependencies, the class has been directly added to the codebase rather than fetched via CMake. The hope is that this class will become redundant once we have access to C++20's std::span.

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 83. Check the log or trigger a new build to see more.

cpp/core/gpu/gpu.cpp

github-actions · 2025-11-26T13:07:54Z

cpp/core/gpu/gpu.h

+  Texture new_empty_texture(const TextureSpec &textureSpec) const;
+
+  [[nodiscard]]
+  Texture new_texture_from_host_memory(const TextureSpec &texture_desc, tcb::span<const float> src_memory_region) const;


warning: function 'MR::GPU::ComputeContext::new_texture_from_host_memory' has a definition with different parameter names [readability-inconsistent-declaration-parameter-name]

Texture new_texture_from_host_memory(const TextureSpec &texture_desc, tcb::span<const float> src_memory_region) const; ^

Additional context

cpp/core/gpu/gpu.cpp:416: the definition seen here

Texture ComputeContext::new_texture_from_host_memory(const TextureSpec &texture_desc, ^

cpp/core/gpu/gpu.h:295: differing parameters are named here: ('src_memory_region'), in definition: ('srcMemoryRegion')

Texture new_texture_from_host_memory(const TextureSpec &texture_desc, tcb::span<const float> src_memory_region) const; ^

cpp/core/span.h

testing/unit_tests/gputests.cpp

github-actions · 2025-11-26T13:07:56Z

testing/unit_tests/gputests.cpp

+  inline static std::unique_ptr<MR::GPU::ComputeContext> shared_context;
+
+  // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
+  MR::GPU::ComputeContext &context;


warning: member 'context' of type 'MR::GPU::ComputeContext &' is a reference [cppcoreguidelines-avoid-const-or-ref-data-members]

MR::GPU::ComputeContext &context; ^

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 58. Check the log or trigger a new build to see more.

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 33. Check the log or trigger a new build to see more.

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 33. Check the log or trigger a new build to see more.

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 33. Check the log or trigger a new build to see more.

testing/unit_tests/gputests.cpp

github-actions · 2025-11-26T13:40:16Z

testing/unit_tests/gputests.cpp

+    )slang";
+
+  const std::vector<float> host_data = {1.0F, 2.0F, 3.0F, 4.0F};
+  const std::vector<float> expected_data = {3.0F, 6.0F, 9.0F, 12.0F};


warning: variable 'expected_data' is not initialized [cppcoreguidelines-init-variables]

Suggested change

const std::vector<float> expected_data = {3.0F, 6.0F, 9.0F, 12.0F};

const std::vector<float> expected_data = 0 = {3.0F, 6.0F, 9.0F, 12.0F};

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

github-actions · 2025-11-26T17:28:45Z

cpp/core/gpu/gpu.h

+  Texture new_empty_texture(const TextureSpec &textureSpec) const;
+
+  [[nodiscard]]
+  Texture new_texture_from_host_memory(const TextureSpec &texture_desc, tcb::span<const float> src_memory_region) const;


warning: function 'MR::GPU::ComputeContext::new_texture_from_host_memory' has a definition with different parameter names [readability-inconsistent-declaration-parameter-name]

Texture new_texture_from_host_memory(const TextureSpec &texture_desc, tcb::span<const float> src_memory_region) const; ^

Additional context

cpp/core/gpu/gpu.cpp:419: the definition seen here

Texture ComputeContext::new_texture_from_host_memory(const TextureSpec &texture_desc, ^

cpp/core/gpu/gpu.h:295: differing parameters are named here: ('src_memory_region'), in definition: ('srcMemoryRegion')

Texture new_texture_from_host_memory(const TextureSpec &texture_desc, tcb::span<const float> src_memory_region) const; ^

github-actions

clang-tidy made some suggestions

cpp/core/gpu/gpu.h

github-actions

clang-tidy made some suggestions

cpp/core/gpu/gpu.h

github-actions

clang-tidy made some suggestions

github-actions · 2025-11-27T12:20:20Z

cpp/core/gpu/slangcodegen.cpp

+void find_bindings_in_type_layout(slang::TypeLayoutReflection *typeLayout,
+                                  std::unordered_map<std::string, ReflectedBindingInfo> &bindings);
+
+void find_bindings_in_variable_layout(slang::VariableLayoutReflection *varLayout,


warning: function 'find_bindings_in_variable_layout' is within a recursive call chain [misc-no-recursion]

void find_bindings_in_variable_layout(slang::VariableLayoutReflection *varLayout, ^

Additional context

cpp/core/gpu/slangcodegen.cpp:124: example recursive call chain, starting from function 'find_bindings_in_type_layout'

void find_bindings_in_type_layout(slang::TypeLayoutReflection *typeLayout, ^

cpp/core/gpu/slangcodegen.cpp:133: Frame #1: function 'find_bindings_in_type_layout' calls function 'find_bindings_in_variable_layout' here:

find_bindings_in_variable_layout(typeLayout->getFieldByIndex(i), bindings); ^

cpp/core/gpu/slangcodegen.cpp:119: Frame #2: function 'find_bindings_in_variable_layout' calls function 'find_bindings_in_type_layout' here:

find_bindings_in_type_layout(varLayout->getTypeLayout(), bindings); ^

cpp/core/gpu/slangcodegen.cpp:119: ... which was the starting point of the recursive call chain; there may be other cycles

find_bindings_in_type_layout(varLayout->getTypeLayout(), bindings); ^

cpp/core/gpu/slangcodegen.h

This a generic GPGPU compute abstraction built on top of WebGPU. To run operations on the GPU, shaders need to be written using the Slang programming language. See https://dawn.googlesource.com/dawn See https://shader-slang.org/

DOWNLOAD_EXTRACT_TIMESTAMP is only available on CMake >=3.24. This change make CMake ignore the option on older versions.

We also set CMP0135 policy behaviour to NEW to fix warnings.

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 40. Check the log or trigger a new build to see more.

github-actions · 2025-11-27T14:56:34Z

testing/unit_tests/gputests.cpp

+  inline static std::unique_ptr<MR::GPU::ComputeContext> shared_context;
+
+  // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
+  MR::GPU::ComputeContext &context;


warning: member 'context' of type 'int &' is a reference [cppcoreguidelines-avoid-const-or-ref-data-members]

MR::GPU::ComputeContext &context; ^

github-actions · 2025-11-27T14:56:34Z

testing/unit_tests/gputests.cpp

+
+TEST_F(GPUTest, MakeEmptyBuffer) {
+  const size_t buffer_element_count = 1024;
+  const Buffer<uint32_t> buffer = context.new_empty_buffer<uint32_t>(buffer_element_count);


warning: variable 'buffer' is not initialized [cppcoreguidelines-init-variables]

Suggested change

const Buffer<uint32_t> buffer = context.new_empty_buffer<uint32_t>(buffer_element_count);

const Buffer<uint32_t> buffer = 0 = context.new_empty_buffer<uint32_t>(buffer_element_count);

github-actions · 2025-11-27T14:56:34Z

testing/unit_tests/gputests.cpp

+TEST_F(GPUTest, BufferFromHostMemory) {
+  std::vector<int32_t> host_data = {1, 2, 3, 4, 5};
+
+  const Buffer<int32_t> buffer = context.new_buffer_from_host_memory<int32_t>(tcb::span<const int32_t>(host_data));


warning: variable 'buffer' is not initialized [cppcoreguidelines-init-variables]

Suggested change

const Buffer<int32_t> buffer = context.new_buffer_from_host_memory<int32_t>(tcb::span<const int32_t>(host_data));

const Buffer<int32_t> buffer = 0 = context.new_buffer_from_host_memory<int32_t>(tcb::span<const int32_t>(host_data));

github-actions · 2025-11-27T14:56:34Z

testing/unit_tests/gputests.cpp

+
+TEST_F(GPUTest, BufferFromHostMemoryVoidPtr) {
+  std::vector<float> host_data = {1.0F, 2.5F, -3.0F};
+  const Buffer<float> buffer =


warning: variable 'buffer' is not initialized [cppcoreguidelines-init-variables]

Suggested change

const Buffer<float> buffer =

const Buffer<float> buffer = 0 =

github-actions · 2025-11-27T14:56:34Z

testing/unit_tests/gputests.cpp

+  std::vector<uint32_t> region3 = {6, 7, 8, 9};
+
+  const std::vector<tcb::span<const uint32_t>> regions = {region1, region2, region3};
+  const Buffer<uint32_t> buffer = context.new_buffer_from_host_memory<uint32_t>(regions);


warning: variable 'buffer' is not initialized [cppcoreguidelines-init-variables]

Suggested change

const Buffer<uint32_t> buffer = context.new_buffer_from_host_memory<uint32_t>(regions);

const Buffer<uint32_t> buffer = 0 = context.new_buffer_from_host_memory<uint32_t>(regions);

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 25 out of 40. Check the log or trigger a new build to see more.

testing/unit_tests/gputests.cpp

github-actions

clang-tidy made some suggestions

cpp/core/gpu/gpu.h

Lestropie · 2025-11-30T23:47:22Z

$ ./build/bin/mrtrix-unit-tests.exe --gtest_filter=GPUTest*
Running main() from C:/msys64/home/rob/src/mrtrix3/build/_deps/googletest-src/googletest/src/gtest_main.cc
Note: Google Test filter = GPUTest*
[==========] Running 16 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 16 tests from GPUTest
[ RUN      ] GPUTest.MakeEmptyBuffer
[       OK ] GPUTest.MakeEmptyBuffer (0 ms)
[ RUN      ] GPUTest.BufferFromHostMemory
[       OK ] GPUTest.BufferFromHostMemory (0 ms)
[ RUN      ] GPUTest.BufferFromHostMemoryVoidPtr
[       OK ] GPUTest.BufferFromHostMemoryVoidPtr (0 ms)
[ RUN      ] GPUTest.BufferFromHostMemoryMultipleRegions
[       OK ] GPUTest.BufferFromHostMemoryMultipleRegions (0 ms)
[ RUN      ] GPUTest.WriteToBuffer
[       OK ] GPUTest.WriteToBuffer (0 ms)
[ RUN      ] GPUTest.WriteToBufferWithOffset
[       OK ] GPUTest.WriteToBufferWithOffset (0 ms)
[ RUN      ] GPUTest.EmptyTexture
[       OK ] GPUTest.EmptyTexture (0 ms)
[ RUN      ] GPUTest.KernelWithInlineShader
[       OK ] GPUTest.KernelWithInlineShader (38 ms)
[ RUN      ] GPUTest.ShaderConstants
[       OK ] GPUTest.ShaderConstants (11 ms)
[ RUN      ] GPUTest.ShaderEntryPointArgs
[       OK ] GPUTest.ShaderEntryPointArgs (21 ms)
[ RUN      ] GPUTest.CopyBufferToBuffer_Full
[       OK ] GPUTest.CopyBufferToBuffer_Full (0 ms)
[ RUN      ] GPUTest.CopyBufferToBuffer_Partial
[       OK ] GPUTest.CopyBufferToBuffer_Partial (0 ms)
[ RUN      ] GPUTest.CopyBufferToBuffer_SourceOutOfRangeThrows
[       OK ] GPUTest.CopyBufferToBuffer_SourceOutOfRangeThrows (0 ms)
[ RUN      ] GPUTest.CopyBufferToBuffer_DestinationOutOfRangeThrows
[       OK ] GPUTest.CopyBufferToBuffer_DestinationOutOfRangeThrows (0 ms)
[ RUN      ] GPUTest.ClearBuffer
[       OK ] GPUTest.ClearBuffer (0 ms)
[ RUN      ] GPUTest.DownloadBufferAsVector
[       OK ] GPUTest.DownloadBufferAsVector (0 ms)
[----------] 16 tests from GPUTest (131 ms total)

[----------] Global test environment tear-down
[==========] 16 tests from 1 test suite ran. (657 ms total)
[  PASSED  ] 16 tests.

daljit46 · 2025-12-01T13:54:23Z

Thanks, that deepens the mystery of why the CI tests are not executing.

daljit46 self-assigned this Nov 26, 2025

daljit46 mentioned this pull request Nov 26, 2025

Experimental GPU-compute abstraction using WebGPU #3096

Closed

daljit46 requested review from a team and removed request for a team November 26, 2025 13:02

github-actions bot reviewed Nov 26, 2025

View reviewed changes

daljit46 force-pushed the webgpu branch from 6ecaa3f to 962afeb Compare November 26, 2025 13:36

github-actions bot reviewed Nov 26, 2025

View reviewed changes

cpp/core/gpu/gpu.h Outdated Show resolved Hide resolved

github-actions bot reviewed Nov 26, 2025

View reviewed changes

cpp/core/gpu/gpu.h Outdated Show resolved Hide resolved

daljit46 force-pushed the webgpu branch from aad3aec to 13e5410 Compare November 26, 2025 21:54

Add Slang + Dawn dependencies

69855b0

daljit46 force-pushed the webgpu branch from 97d59d7 to 3b4b99a Compare November 27, 2025 12:04

github-actions bot reviewed Nov 27, 2025

View reviewed changes

daljit46 force-pushed the webgpu branch 2 times, most recently from 4e205c0 to 24b7532 Compare November 27, 2025 14:25

daljit46 added 10 commits November 27, 2025 14:34

Add new GPU compute API

e89112d

This a generic GPGPU compute abstraction built on top of WebGPU. To run operations on the GPU, shaders need to be written using the Slang programming language. See https://dawn.googlesource.com/dawn See https://shader-slang.org/

Add unit tests for GPU API

d7c70cc

Reorder arguments for FetchContent_Declare

c24a724

DOWNLOAD_EXTRACT_TIMESTAMP is only available on CMake >=3.24. This change make CMake ignore the option on older versions.

Fix typos

9725022

Remove DOWNLOAD_EXTRACT_TIMESTAMP

f63dd43

We also set CMP0135 policy behaviour to NEW to fix warnings.

Don't run unit tests in parallel

0428640

Add Threads package requirement for Dawn integration

9d9111e

Update Dawn library path handling for Linux systems

1c58428

Update required cmake version in readme

3e763d5

Add mesa-vulkan-drivers to dependency installation for Linux builds

13a684b

daljit46 added 9 commits November 27, 2025 14:34

Fix instance descriptor initialization in ComputeContext constructor

50712c6

Use supported device limits for storage buffer sizes

444d16a

Update Slang directory path handling for cross-platform compatibility

887c710

Fix filesystem path handling in ComputeContext constructor

f9a43fb

Add post-build command to copy imported dependencies

2308693

Add mesa package to MinGW dependencies for Windows build

95232f4

Catch exceptions during ComputeContext creation for gpu tests

c543f2b

Use std::string_view where required by check_syntax

764fa3a

Add workaround for loading vulkan dll on MSYS2

c5691db

daljit46 force-pushed the webgpu branch from 24b7532 to 17054af Compare November 27, 2025 14:46

daljit46 added 2 commits November 27, 2025 14:52

Fetch tcb::span with CMake

d3c008b

Introduce MRTRIX_USE_SYSTEM options for Slang, Dawn and tcb::span

9738f59

daljit46 force-pushed the webgpu branch from 17054af to 9738f59 Compare November 27, 2025 14:53

github-actions bot reviewed Nov 27, 2025

View reviewed changes

Use FetchContent_Populate directly to avoid building tcb::span tests

365f00d

daljit46 force-pushed the webgpu branch from 58423ce to 365f00d Compare November 27, 2025 15:11

daljit46 added 2 commits November 27, 2025 15:18

Add copyright notice

dc37898

Apply clang-format-all

bfe549d

daljit46 force-pushed the webgpu branch from 045f5cd to bfe549d Compare November 28, 2025 14:04

github-actions bot reviewed Nov 28, 2025

View reviewed changes

cpp/core/gpu/gpu.h Show resolved Hide resolved

	const std::vector<float> expected_data = {3.0F, 6.0F, 9.0F, 12.0F};
	const std::vector<float> expected_data = 0 = {3.0F, 6.0F, 9.0F, 12.0F};

	const Buffer<uint32_t> buffer = context.new_empty_buffer<uint32_t>(buffer_element_count);
	const Buffer<uint32_t> buffer = 0 = context.new_empty_buffer<uint32_t>(buffer_element_count);

	const Buffer<int32_t> buffer = context.new_buffer_from_host_memory<int32_t>(tcb::span<const int32_t>(host_data));
	const Buffer<int32_t> buffer = 0 = context.new_buffer_from_host_memory<int32_t>(tcb::span<const int32_t>(host_data));

	const Buffer<float> buffer =
	const Buffer<float> buffer = 0 =

GPU compute API abstraction on top of WebGPU #3238

Are you sure you want to change the base?

GPU compute API abstraction on top of WebGPU #3238

Uh oh!

Conversation

daljit46 commented Nov 26, 2025

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot Nov 26, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot Nov 26, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot Nov 26, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment