rocWMMA is a C++ library for accelerating mixed-precision matrix multiply-accumulate (MMA) operations leveraging AMD GPU hardware. rocWMMA makes it easier to break down MMA problems into fragments and distribute block-wise MMA operations in parallel across GPU wavefronts. Our API consists of a header library, that you can use to compile MMA acceleration directly into GPU kernel device code. This can benefit from compiler optimization in the generation of kernel assembly, and doesn't incur additional overhead costs of linking to external runtime libraries or having to launch separate kernels.
The rocWMMA API is designed from a wave perspective, where block-wise API calls are processed by GPU threads coordinated together as a group, or a wave. Importantly, individual threads may only access a portion of a fragment, and there is no guaranteed layout or locality of this data. Thread access to fragment data is analogous to the vector register access of individual threads. Fragments and block-wise operations can therefore be thought of as a wave of grouped data buffers and operations.
Thread blocks that contain multiple waves have the added benefit of resource sharing and
cooperation. For rocWMMA, this is especially useful for moving data that may be shared across
multiple waves. The rocWMMA Coop API facilitates cooperation between multiple waves when
loading and storing opaque data fragments. The order and locality of fragment data contribution per
wave is not guaranteed, and can be thought of as a distributed wave. Typically, waves cooperate in
moving data from global memory to shared memory (LDS), loading their own full fragment copies
from LDS.
Memory addresses are treated as 1D arrays. The rocWMMA API can opaquely handle moving data between both global and shared memory address locations. The library code includes utilities for mapping 2D grid and matrix coordinates to 1D array coordinates, and supports either row-major or column-major data layouts. Block-wise dimensions (BlockM,N,K) familiar to block-wise general matrix multiply (GEMM) product algorithms are supported directly through the availability of matrix instructions for the target architecture. Likewise, you can vary mixed-precision data types for input, output, and accumulation fragments (as listed in the supported configurations section).
rocWMMA is a header library that includes test and sample projects to validate and demonstrate API use. GEMM is used as primary validation given the heavy precedent for the library. However, we are expanding our portfolio to demonstrate more diverse rocWMMA uses.
You can use rocWMMA with the following hardware:
- AMD CDNA class GPU featuring matrix core support: gfx908, gfx90a as 'gfx9'
- AMD RDNA3 class GPU featuring AI acceleration support: gfx1100, gfx1101, gfx1102 as 'gfx11'
Double precision FP64 data type support requires gfx90a
rocWMMA software requirements include:
- ROCm stack minimum version 5.4
- rocm-cmake minimum version 0.8.0 for ROCm 5.3
- C++ 17
- CMake >=3.6
- OpenMP
Optional:
- rocBLAS minimum version 2.46.0 for ROCm 5.4 (for rocBLAS validation and benchmarks)
- Doxygen (for building documentation)
To build our documentation locally, use the following commands.
cd docs
pip3 install -r .sphinx/requirements.txt
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html- Wave Size: Wave32 (gfx11), Wave64 (gfx9)
- Matrix Layout <LayoutA, LayoutB, Layout C, LayoutD> (N = col major, T = row major)
<N, N, N, N>, <N, N, T, T>
<N, T, N, N>, <N, T, T, T>
<T, N, N, N>, <T, N, T, T>
<T, T, N, N>, <T, T, T, T>- Thread Block Sizes <TBlockX, TBlockY>
TBlockX must be a multiple of WaveSize
For GEMM, rocWMMA focuses on thread blocks of up to 4 waves for optimal resource usage and occupancy. Larger thread block sizes are possible but are not officially supported.
WS = Wave Size
<WS, 1>, <WS, 2>, <WS, 4>
<WS * 2, 1>, <WS * 2, 2>
<WS * 4, 1>-
Data Types <Ti / To / Tc> = <Input type / Output Type / Compute Type>
Input Type = Matrix A/B
Output Type = Matrix C/D
Compute Type = math / accumulation type
gfx11 only supports BlockM/N = 16
| Ti / To / Tc | BlockM | BlockN | BlockK Range (powers of 2) |
Notes |
|---|---|---|---|---|
| i8 / i32 / i32 | 16 | 16 | [16, 256] | |
| 32 | 32 | [8, 128] | ||
| i8 / i8 / i32 | 16 | 16 | [16, 256] | |
| 32 | 32 | [8, 128] | ||
| f16 / f32 / f32 | 16 | 16 | [16, 256] | |
| 32 | 32 | [8, 128] | ||
| f16 / f16 / f32 | 16 | 16 | [16, 256] | |
| 32 | 32 | [8, 128] | ||
| f16 / f16 / f16* | 16 | 16 | [16, 256] | *= CDNA native f32 accumulation downcasted to fp16 |
| 32 | 32 | [8, 128] | ||
| __half / f32 / f32 | 16 | 16 | [16, 256] | |
| 32 | 32 | [8, 128] | ||
| __half / __half / f32 | 16 | 16 | [16, 256] | |
| 32 | 32 | [8, 128] | ||
| __half / __half / __half* | 16 | 16 | [16, 256] | *= CDNA native f32 accumulation downcasted to __half |
| 32 | 32 | [8, 128] | ||
| bf16 / f32 / f32 | 16 | 16 | [8, 256] | |
| 32 | 32 | [4, 128] | ||
| bf16 / bf16 / f32 | 16 | 16 | [8, 256] | |
| 32 | 32 | [4, 128] | ||
| bf16 / bf16 / bf16* | 16 | 16 | [8, 256] | *= CDNA native f32 accumulation downcasted to bf16 |
| 32 | 32 | [4, 128] | ||
| f32 / f32 / f32* | 16 | 16 | [4, 256] | *= Supported only on gfx9 |
| 32 | 32 | [2, 128] | ||
| f64 / f64 / f64* | 16 | 16 | [4, 256] | *= Supported only on gfx90a + |
Broadcasts a desired value to all elements in the fragment.
Loads data from memory according to the matrix layout.
- Matrix A layout: Loads and stores matrix columns in the K direction (Matrix A = M x K, fragA = BlockM x BlockK)
- Matrix B layout: Loads and stores matrix rows in the K direction (Matrix B = K x N, fragB = BlockK x BlockN)
- Matrix C layout: Loads and stores matrix rows in a vector width of 4 (Matrix C = M x N, fragAcc = BlockM x BlockN)
Fragments are stored in packed registers in optimal load and store patterns. In-register elements have no guaranteed order, which optimizes loading and storing efficiency.
The MMA operation is performed on fragment data. The outer product of Fragment A elements with Fragment B elements is added to the accumulator fragment.
Flow control for synchronization across multiple wavefronts in a workgroup. It also ensures the synchronization of shared and global memory accesses across wavefronts.
Loads data from memory according to the matrix layout. Splits operation among wave members of a workgroup.
- Matrix A: Layout loads and stores matrix columns in the K direction (Matrix A = M x K, fragA = BlockM x BlockK)
- Matrix B: Layout loads and stores matrix rows in the K direction (Matrix B = K x N, fragB = BlockK x BlockN)
- Matrix C: Layout loads and stores matrix rows in vector width of 4 (Matrix C = M x N, fragAcc = BlockM x BlockN)
Each contributing wave handles partial operation data, so split parameters should be consistent across subsequent operations. For example, cooperatively moving data from global to shared memory should use the same split parameters for the global load and subsequent local store.
-
Create and track a rocWMMA fork.
-
Clone your fork:
git clone -b develop https://github.com/<your_fork>/rocWMMA.git . .githooks/install git checkout -b <new_branch> ... git add <new_work> git commit -m "What was changed" git push origin <new_branch> ...
-
Create a pull request to the ROCmSoftwarePlatform/rocWMMA develop branch.
-
Await CI and approval feedback.
-
Once approved, merge.
You must install GitHooks because there are triggers for Clang formatting in commits.
| Option | Description | Default value |
|---|---|---|
| AMDGPU_TARGETS | Build code for specific GPU target(s) | gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx1100;gfx1101;gfx1102 |
| ROCWMMA_BUILD_TESTS | Build Tests | ON |
| ROCWMMA_BUILD_SAMPLES | Build Samples | ON |
| ROCWMMA_BUILD_DOCS | Build doxygen documentation from code | OFF |
| ROCWMMA_BUILD_ASSEMBLY | Generate assembly files | OFF |
| ROCWMMA_BUILD_VALIDATION_TESTS | Build validation tests | ON (requires ROCWMMA_BUILD_TESTS=ON) |
| ROCWMMA_BUILD_BENCHMARK_TESTS | Build benchmark tests | OFF (requires ROCWMMA_BUILD_TESTS=ON) |
| ROCWMMA_BUILD_EXTENDED_TESTS | Build extended testing coverage | OFF (requires ROCWMMA_BUILD_TESTS=ON) |
| ROCWMMA_VALIDATE_WITH_ROCBLAS | Use rocBLAS for validation tests | ON (requires ROCWMMA_BUILD_VALIDATION_TESTS=ON) |
| ROCWMMA_BENCHMARK_WITH_ROCBLAS | Include rocBLAS benchmarking data | OFF (requires ROCWMMA_BUILD_BENCHMARK_TESTS=ON) |
By default, the project is configured in release mode and is linked against rocBLAS for validating results. Here are some configuration examples:
| Configuration | Command |
|---|---|
| Basic | CC=hipcc CXX=hipcc cmake -B<build_dir> . |
| Targeting gfx908 | CC=hipcc CXX=hipcc cmake -B<build_dir> . -DAMDGPU_TARGETS=gfx908:xnack- |
| Debug build | CC=hipcc CXX=hipcc cmake -B<build_dir> . -DCMAKE_BUILD_TYPE=Debug |
| Build without rocBLAS (default on) | CC=hipcc CXX=hipcc cmake -B<build_dir> . -DROCWMMA_VALIDATE_WITH_ROCBLAS=OFF -DROCWMMA_BENCHMARK_WITH_ROCBLAS=OFF |
After configuration, build with cmake --build <build_dir> -- -j<nproc>
The build time for all projects can take several minutes.
-
Target a specific GPU (e.g.,
gfx908:xnack-) -
Use lots of threads (e.g.,
-j64) -
Select
ROCWMMA_BUILD_ASSEMBLY=OFF -
Select
ROCWMMA_BUILD_DOCS=OFF -
Select
ROCWMMA_BUILD_EXTENDED_TESTS=OFF -
Specify
ROCWMMA_BUILD_VALIDATION_TESTSorROCWMMA_BUILD_BENCHMARK_TESTSas ON, and the other as OFF -
For investigating particular kernels, build the ad hoc test with the parameters you are interested in
-
Manually build specific tests:
cd <build_dir> make <target_name> -j64
Where
<target_name>is one of the following:<target_name>Description rocwmma_unit_testsBuild all rocWMMA unit tests rocwmma_gemm_tests_validateBuild all GEMM validation tests rocwmma_gemm_tests_benchBuild all GEMM benchmark tests rocwmma_dlrm_tests_validateBuild all deep learning recommendation model (DLRM) validation tests rocwmma_dlrm_tests_benchBuild all DLRM benchmark tests rocwmma_samplesBuild all rocWMMA samples Individual target name ( contamination_test,simple_sgemm, etcetera)Build individual rocWMMA test or sample -
Manually reduce tests or test case coverage
rocWMMA library features are showcased, validated, and benchmarked if applicable in GoogleTest applications.
With CTest enabled, you can run the entire test suite by running CTest in the build folder:
cd <build_dir>
ctest --output-on-failureOtherwise, individual tests can be run as follows:
Unit tests for loading and storing APIs to verify that data boundaries are not crossed and pristine data remain untouched.
Run the validation:
<build_dir>/test/unit/contamination_testUnit tests for vector cross-lane operations.
Run the validation:
<build_dir>/test/unit/cross_lane_ops_testTests the rocwmma::fill_fragment API function for all supported configurations. Tests broadcasting of a
desired value to all elements in the fragment.
Run the validation:
<build_dir>/test/unit/fill_fragment_testUnit test for I/O shape meta-data generation on host machine
Run the validation:
<build_dir>/test/unit/io_shape_testUnit test for I/O traits metadata generation on host machine.
Run the validation:
<build_dir>/test/unit/io_traits_testUnit tests for the internal collect and scatter matrix element to register mapping transforms.
Run the validation:
<build_dir>/test/unit/layout_testTests the rocwmma::load_matrix_sync and rocwmma::store_matrix_sync API functions for all
supported configurations. Tests proper emplacement of data during loads and stores.
Run the validation:
<build_dir>/test/unit/load_store_matrix_sync_test
<build_dir>/test/unit/load_store_matrix_coop_sync_testUnit tests for the utility class used to calculate transforms and offsets between grid, matrix, and data coordinate systems.
Run the validation:
<build_dir>/test/unit/map_util_testUnit tests for internal vector iteration and navigation during access and storage.
Run the validation:
<build_dir>/test/unit/vector_iterator_testUnit tests for internal vector storage.
Run the validation:
<build_dir>/test/unit/vector_testImplements a GEMM blocking algorithm using rocWMMA for all supported parametric configurations. Validates on a CPU algorithm, or rocBLAS if available. Validation runs are performed on a reduced subset of matrix sizes. Benchmark runs are performed on a larger set of matrix sizes. Extended tests provide comprehensive parameter coverage and larger matrix sizes.
As of rocWMMA v0.8, we've introduced GEMM kernel naming nomenclature to allow compact representation and quick identification of kernel implementation features. The rocWMMA GEMM test library includes kernels that support the following features:
PGR# - Prefetch Global Read lookup stages. PGR0 = no global read prefetch. PGR1 = 1 stage global read prefetch.
LB# - Lds buffer count. LB0 = no lds usage, LB2 = 2 Lds buffers used for swap.
MP# - MFMA instruction priority. MP0 = default MFMA instruction priority of 0. MP1 = raise MFMA instruction priority to 1.
MB - Multiple output blocks targeted per wave
SB - Single output block target per wave
NC - Non-Cooperative load / store
CP - Cooperative load / store
BLK - Cooperative load / store per block tile
WV - Cooperative load / store per wave tile
WG - Cooperative load / store per macro tile-
gemm_PGR0_LB0_MP0_SB_NC: The simplest blocked GEMM example, which targets one output block of matrix multiplication per wave. No prefetch, no lDs usage, default MFMA prioritization, single block output and non-collaborative. -
gemm_PGR0_LB0_MP0_MB_NC: Implements a multi-block GEMM where each wave is responsible for a BlocksX x BlocksY grid of output blocks. No prefetch, no lDs usage, default MFMA prioritization, multiple blocks output, and non-collaborative. -
gemm_PGR1_LB2_MP0_MB_CP_BLK: Implements a multi-block GEMM where each wave is responsible for a BlocksX x BlocksY grid of output blocks. This kernel leverages shared memory to implement a data prefetching pipeline and collaborates with other waves to improve performance. Implements single stage prefetch, double lDs buffer, default MFMA prioritization, multiple blocks output, and is block-tile collaborative in global read and local write. -
gemm_PGR1_LB2_MP0_MB_CP_WV: Implements a multi-block GEMM where each wave is responsible for a BlocksX x BlocksY grid of output blocks. This kernel leverages shared memory to implement a data prefetching pipeline and collaborates with other waves to improve performance. Implements single stage prefetch, double lDs buffer, default MFMA prioritization, multiple blocks output, and is wave-tile collaborative in global read and local write. -
gemm_PGR1_LB2_MP0_MB_CP_WG: Implements a multi-block GEMM where each wave is responsible for a BlocksX x BlocksY grid of output blocks. This kernel leverages shared memory to implement a data prefetching pipeline and collaborates with other waves to improve performance. Implements single stage prefetch, double lDs buffer, default MFMA prioritization, multiple blocks output and is macro-tile collaborative in global read and local write. -
Ad Hoc Test: An executable that focuses on a specific set of kernel parameters. This is used as a quick mock-up of a situational investigation of a particular GEMM kernel.
Validation tests are postfixed with -validate. Benchmark tests are postfixed with -bench.
Run the validation tests:
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_SB_NC-validate
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_MB_NC-validate
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_BLK-validate
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_WV-validate
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_WG-validateRun the benchmark only:
Note that benchmark runs can take several hours to complete.
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_SB_NC-bench
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_MB_NC-bench
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_BLK-bench
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_WV-bench
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_WG-benchRun the ad hoc test:
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_SB_NC_ad_hoc-validate
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_MB_NC_ad_hoc-validate
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_ad_hoc-validate
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_SB_NC_ad_hoc-bench
<build_dir>/test/gemm/gemm_PGR0_LB0_MP0_MB_NC_ad_hoc-bench
<build_dir>/test/gemm/gemm_PGR1_LB2_MP0_MB_CP_ad_hoc-bench| Compact | Verbose | Description |
|---|---|---|
| -os <output_file>.csv | --output_stream <output_file>.csv | stream GEMM testing output to CSV file |
| --omit | omits certain outputs : 1 = SKIPPED tests 2 - FAILED tests 4 - PASSED tests 8 - All non-gtest output |
- Use GoogleTest filters, target specific test names:
<test_exe> --gtest_filter=*name_filter*- Manually adjust the test cases coverage
- Use ad hoc tests to focus in on specific parameters
- Select
ROCWMMA_BUILD_EXTENDED_TESTS=OFF
These are standalone, practical use cases for the rocWMMA API. They have minimal dependencies and represent a targeted application with a fixed set of parameters.
Use MMA to demonstrate rocWMMA API usage in the context of wave-level GEMM computation, in both simplified and optimized versions.
Simple GEMM algorithm demonstration without LDS memory usage and no transpose.
simple_sgemm calculates D = Alpha * A x B + Beta * C with fp32 inputs and output.
simple_dgemm calculates D = Alpha * A x B + Beta * C with fp64 inputs and output.
simple_hgemm calculates D = Alpha * A x B + Beta * C with fp16 inputs and output.
Includes a simple CPU validation and benchmark.
Run a simple gemm sample:
<build_dir>/samples/simple_sgemm
<build_dir>/samples/simple_dgemm
<build_dir>/samples/simple_hgemmTo implement and measure performance of Matrix Multiply-Accumulate(D = Alpha * A x B + Beta * C) with user-defined configurations on a GPU.
It contains the best performant version of the multi-block GEMM algorithm with LDS memory, macro-tile collaboration, data reuse, and optimized pipeline, configured with the finest parameters for larger sizes (1K and greater).
perf_sgemm calculates D = Alpha * A x B + Beta * C with fp32 inputs and output.
perf_dgemm calculates D = Alpha * A x B + Beta * C with fp64 inputs and output.
perf_hgemm calculates D = Alpha * A x B + Beta * C with fp16 inputs and output.
Includes a simple CPU validation and benchmark.
Run a perf gemm sample:
<build_dir>/samples/perf_sgemm
<build_dir>/samples/perf_dgemm
<build_dir>/samples/perf_hgemmSimple MMA with a vector demonstration, without LDS or transpose.
Calculates Y = alpha * (A) * X + beta * Y with fp32 inputs and output.
Includes a simple CPU validation and benchmark.
A = Matrix of size m * k (col-major)
X = Vector of size k * 1 (col-major)
Y = accumulator of size m * 1 (col-major)
Run the sgemv sample:
<build_dir>/samples/simple_sgemvSimple MMA with a vector demonstration, without LDS or transpose.
Calculates Y = alpha * (A) * X + beta * Y with fp64 inputs and output.
Includes a simple CPU validation and benchmark.
A = Matrix of size m * k (row-major)
X = Vector of size k * 1 (col-major)
Y = accumulator of size m * 1 (row-major)
Run the dgemv sample:
<build_dir>/samples/simple_dgemvSimple deep learning recommendation model (DLRM) for machine learning. Implements both forward and backward passes on fp16 inputs and outputs.
Includes a simple CPU validation and benchmark.
Run the simple_dlrm sample:
<build_dir>/samples/simple_dlrmThe HIP runtime compilation (hipRTC) environment allows simultaneous compilation, loading, and running of device code on AMD GPUs. The rocWMMA library is compatible with hipRTC, so you can leverage it for runtime-generated kernels. A simple GEMM sample is included to demonstrate compatibility. For more information, refer to the HIP API reference.
The `rocwmma::bfloat16_t` data type is not currently supported in hipRTC.
A simple GEMM algorithm demonstrating runtime compilation (hipRTC) compatibility. Calculates D = Alpha * A x B + Beta * C with mixed-precision fp16 inputs and fp32 output. Includes code compilation, module loading, and kernel launch with the hipRTC API, with simple CPU validation and benchmarking.
Run the hipRTC_gemm sample:
<build_dir>/samples/hipRTC_gemm