@@ -7,6 +7,7 @@ Please describe the motivation behind the pull request, whether it enables a new
...
@@ -7,6 +7,7 @@ Please describe the motivation behind the pull request, whether it enables a new
Please put an `x` into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.
Please put an `x` into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and the unit tests are passing locally
- [ ] I have added tests relevant to the introduced functionality, and the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers with understanding the motivation
- [ ] I have added inline documentation which enables the maintainers with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant after this pull request
- [ ] I have removed the stale documentation which is no longer relevant after this pull request
- [ ] (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
- [ ] (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
@@ -121,6 +121,15 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa
...
@@ -121,6 +121,15 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa
You can find instructions for running each individual example in [example](/example).
You can find instructions for running each individual example in [example](/example).
* Build and run smoke/regression examples and tests:
```bash
make -j smoke # tests and examples that run for < 30 seconds each
```
```bash
make -j regression # tests and examples that run for >= 30 seconds each
```
* Build ckProfiler:
* Build ckProfiler:
```bash
```bash
...
@@ -153,6 +162,9 @@ Additional cmake flags can be used to significantly speed-up the build:
...
@@ -153,6 +162,9 @@ Additional cmake flags can be used to significantly speed-up the build:
`batched_gemm_multi_d_dl`. These instances are useful on architectures like the NAVI2x, as most
`batched_gemm_multi_d_dl`. These instances are useful on architectures like the NAVI2x, as most
other platforms have faster instances, such as `xdl` or `wmma`, available.
other platforms have faster instances, such as `xdl` or `wmma`, available.
*`DPP_KERNELS` (default is OFF) must be set to ON in order to build instances, such as `gemm_dpp`.
These instances are useful on architectures like the NAVI2x, as most other platforms have faster instances, such as `xdl` or `wmma`, available.
*`CK_USE_FP8_ON_UNSUPPORTED_ARCH` (default is OFF) must be set to ON in order to build instances,
*`CK_USE_FP8_ON_UNSUPPORTED_ARCH` (default is OFF) must be set to ON in order to build instances,
such as `gemm_universal`, `gemm_universal_streamk` and `gemm_multiply_multiply` for fp8 data type for GPU targets which do not have native support for fp8 data type, such as gfx908 or gfx90a. These instances are useful on
such as `gemm_universal`, `gemm_universal_streamk` and `gemm_multiply_multiply` for fp8 data type for GPU targets which do not have native support for fp8 data type, such as gfx908 or gfx90a. These instances are useful on
architectures like the MI100/MI200 for the functional support only.
architectures like the MI100/MI200 for the functional support only.
[Back to supported operations](../../../include/ck/README.md)
# Composable Kernel GEMM
## GEMM
General matrix multiplications operation. In CK GEMM operation is called as `DeviceGemm` and requires following types as template parameters:
***ALayout** - A matrix layout (RowMajor/ColumnMajor).
***BLayout** - B matrix layout (RowMajor/ColumnMajor).
***CLayout** - B matrix layout (RowMajor/ColumnMajor).
***ADataType** - A matrix data type.
***BDataType** - B matrix data type.
***CDataType** - B matrix data type.
***AElementwiseOperation** - Fused operation on tensor A before GEMM.
***BElementwiseOperation** - Fused operation on tensor B before GEMM.
***CElementwiseOperation** - Fused operation on tensor C after GEMM.
For matrices with large K dimension `DeviceGemmSplitK` implementation is available. This implementation allows user to split K dimension between work groups. This implementation uses `AtomicAdd` operation on global memory, thus need to zero-out output buffer for correct results.
For fused operations with additional tensor there are `DeviceGemmMultipleABD` or `DeviceGemmMultipleD` operation which require following parameters:
***DsLayout** - layouts for additional tensors for fused operations.
***DsDataType** - data types for additional tensors for fused operations.
For `DeviceGemmMultipleABD`**ALayout**, **BLayout**, **ADataType** and **BDataType** user should pass a tuple.
List of the device operations in CK:
***DeviceGemmDl** - Device operation with DL instructions.
***DeviceGemmDpp** - Device operation with DL instructions with DPP instructions during data load.
***DeviceGemmWmma_CShuffle** - Device operation with WMMA instructions with CShuffle optimization for more optimized data store.
***DeviceGemm_Xdl_CShuffle_LdsDirectLoad** - Device operation with XDL instructions and CShuffle optimization for more optimized data store and direct load from global memory to shared memory.
***DeviceGemm_Xdl_CShuffle** - Device operation with XDL instructions with CShuffle optimization for more optimized data store.
***DeviceGemm_Xdl_CShuffleV2** - Device operation with XDL instructions with CShuffle optimization for more optimized data store. GEMM pipeline has been optimized compared to **DeviceGemm_Xdl_CShuffle**.
***DeviceGemmXdlSkipBLds** - Device operation with XDL instructions. Load to shared memory has been skiped for B matrix.
***DeviceGemm_Xdl_WaveletModel_CShuffle** - Device operation with XDL instructions with CShuffle optimization for more optimized data store. Producer and consumer scheme cooperation between waves in workgroup.
***DeviceGemmXdl** - Device operation with XDL instructions.
Table of supported cases by instance factory with XDL instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
| |Is supported|
|-------|---|
|bf16|✓|
|fp16|✓|
|fp32|✓|
|int8|✓|
|fp8 |✓|
Table of supported cases by instance factory with WMMA instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
| |Is supported|
|-------|---|
|bf16|✓|
|fp16|✓|
|fp32|✗|
|int8|✓|
|fp8 |✗|
Table of supported cases by instance factory with DL instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
| |Is supported|
|-------|---|
|bf16|✗|
|fp16|✓|
|fp32|✓|
|int8|✓|
|fp8 |✗|
Table of supported cases by instance factory with fused output elementwise operation:
***B Matrix Multiply + Add + Gelu** - bf16 (int8 for B matrix)
***B Matrix Multiply + Add** - bf16 (int8 for B matrix)
***B Matrix Multiply + Gelu** - bf16 (int8 for B matrix)
***B Matrix Multiply** - bf16 (int8 for B matrix)
***Add + Add + Gelu** - fp16
***Add + Gelu** - fp16, bf16 (int8 for B matrix) for Row/Column/Row
***Multiply** - fp16
***Add + Multiply** - fp16
***Add + Relu** - fp16 (int8 for B matrix) for Row/Column/Row, bf16 (int8 for B matrix) for Row/Column/Row
***Add + Silu** - fp16 (int8 for B matrix) for Row/Column/Row, bf16 (int8 for B matrix) for Row/Column/Row
***Add** - fp16 (int8 for B matrix) for Row/Column/Row, bf16 (int8 for B matrix) for Row/Column/Row
***Bilinear** - fp16, int8
***Gelu** - fp16
***Multiply + Add** - fp16 for Row/Column/Row and Row/Row/Row, fp16 (int8 for B matrix, fp32 for Bias) for Row/Column/Row and Row/Row/Row,
***Quantization** - int8
## GEMM V2 (Universal GEMM)
General matrix multiplications operation optimized for MI300 series. Operation is called as `DeviceGemmV2` and requires following types as template parameters:
***ALayout** - A matrix layout (RowMajor/ColumnMajor).
***BLayout** - B matrix layout (RowMajor/ColumnMajor).
***CLayout** - B matrix layout (RowMajor/ColumnMajor).
***ADataType** - A matrix data type.
***BDataType** - B matrix data type.
***CDataType** - B matrix data type.
***AElementwiseOperation** - Fused operation on tensor A before GEMM.
***BElementwiseOperation** - Fused operation on tensor B before GEMM.
***CElementwiseOperation** - Fused operation on tensor C after GEMM.
This implementation allows user to split K dimension between work groups. This implementation requires AtomicAdd operation on global memory (output buffer must be set to zeroes if splitK parameter is larger than one).
List of the device operations for in CK:
***DeviceGemm_Xdl_CShuffleV3** - Device operation with XDL instructions with CShuffle optimization for more optimized data store.
***DeviceGemm_Xdl_CShuffleV3R1** - Device operation with XDL instructions with CShuffle optimization for more optimized data store. This implementation perform reduction on splitted K dimension after GEMM instead of AtomicAdd instruction.
Table of supported cases by instance factory with XDL instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
| |Is supported|
|-------|---|
|bf16|✓|
|fp16|✓|
|fp32|✗|
|int8|✗|
|fp8 (C bf16)|✓|
|fp16 (A fp8)|✓|
|fp16 (B fp8)|✓|
## Others
***DeviceGemm_dequantB** - GEMM with dequantization (implemented with WMMA instructions).
***DeviceGemmMultipleD_ABScale** - GEMM with scale for A and B matrix.
***DeviceGemmMultipleDLayernorm** - GEMM fused with layernorm.
***DeviceGemmMultipleDMultipleR** - GEMM fused with reductions and custom global reductions operators.
***DeviceGemmReduce** - GEMM fused with reduction.
***DeviceGemm_Streamk_V2** - GEMM stream K implementation. Implementation allows to use reduction instead of AtomicAdd.
***DeviceGemmStreamK** - GEMM stream K implementation using AtomicAdd.
[Back to supported operations](../../../include/ck/README.md)
# Composable Kernel Grouped Convolution
## Grouped Convolution Forward
Grouped convolution operation for 1D, 2D or 3D spatial dimensions. Convolution utilizes GEMM kernel after tensor coordinate transform. In CK Grouped Convolution Forward operation is called as `DeviceGroupedConvFwdMultipleABD` and requires following types as template parameters:
***NumDimSpatial** - number of spatial dimensions (1D, 2D, 3D).
***ADataType** - input data type. Pass tuple if there is fused operation with input.
***BDataType** - weight data type. Pass tuple if there is fused operation with weight.
***DsDataType** - data types for additional tensors for fused operations.
***EDataType** - Output data type.
***AElementwiseOperation** - fused operation on tensor A (input).
***BElementwiseOperation** - fused operation on tensor B (weight).
***CDEElementwiseOperation** - fused operation on tensor C (output).
***AComputeType** - compute data type of tensor A for mfma instruction (ADataType by default).
***BComputeType** - compute data type of tensor B for mfma instruction (AComputeType by default).
Grouped convolution forward support tensors larger than 2GB.
List of the device operations for grouped convolution forward in CK:
***DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3** - Device operation with XDL instructions. Optimized for AMD Instinct MI300 series.
***DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle** - Device operation with XDL instructions and support of fused operations to input, weight and output.
***DeviceGroupedConvFwdMultipleD_Wmma_CShuffle** - Device operation with WMMA instructions.
***DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK** - Device operation with DL instructions.
Table of supported cases by instance factory with XDL instruction:
[Back to supported operations](../../../include/ck/README.md)
# Composable Kernel Grouped Convolution
## Grouped Convolution Backward Data
Grouped convolution operation for 1D, 2D or 3D spatial dimensions. Convolution utilizes GEMM kernel after tensor coordinate transform. In CK Grouped Convolution Backward Data operation is called as `DeviceGroupedConvBwdDataMultipleD` and requires following types as template parameters:
***NumDimSpatial** - number of spatial dimensions (1D, 2D, 3D).
[Back to supported operations](../../../include/ck/README.md)
# Composable Kernel Grouped Convolution
## Grouped Convolution Backward Weight
Grouped convolution operation for 1D, 2D or 3D spatial dimensions. Convolution utilizes GEMM kernel after tensor coordinate transform. Backward weight version uses splitK feature (due to large GEMM K dimension). In CK Grouped Convolution Backward Weight operation is called as `DeviceGroupedConvBwdWeight` and requires following types as template parameters:
***NumDimSpatial** - number of spatial dimensions (1D, 2D, 3D).
