mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-07-11 17:51:40 +00:00
7a710c3542
feat(ck): Extend and optimize Quant Gemm Kernel for Aiter a8w8 (#8423) ## Motivation JIRA ID : ROCM-20362 JIRA ID : ROCM-26313 The main purpose of this PR is to allow using CK and CK Tile GEMM for Aiter a8w8 (row-col quantization) and improve its performance. ## Technical Details ### Multiple D for Aiter a8w8 with bias * Support multiple D (bias) in Quant GEMM kernel * Extend CShuffleEpilogue * support int8 -> int32 WarpGemms with row-col quantization * allow shuffling in fp32 before applying multiple D to prevent precision loss ### Large tensors support * Support large tensors in the Quant GEMM kernel by offsetting pointers of matrices A, D and C. This feature can be used when M is large, N and K are relatively small and layout is RCR, it's currently enabled only row-col quantization. * Allow broadcasting of D column vectors in the old CK's `DeviceGemmMultiD_Xdl_CShuffle_V3` with large tensors, this case is used to implement row-col quant scales in Aiter. ### Optimization and workarounds * Use literal 0 as scales for unscaled 16x16x128 and 32x32x64 mfma: llvm uses v_mfma for `__builtin_amdgcn_mfma_scale_f32_..._f8f6f4` instead of v_mfma_scale only if scales are literal 0 values. These instruction don't need loading scales and save vector registers. See https://github.com/ROCm/llvm-project/blob/therock-7.13/llvm/lib/Target/AMDGPU/SIInstrInfo.td#L317-L327 * Add workaround for inefficient buffer_load to lds on 7.2 The 3rd argument of buffer_load_dwordx4 is a scalar register. But the compiler generates a waterfall loop as if lanes can have a different value, even though the original values comes from as scalar register. * Use buffer_store_dwordx4 to store 8 bf16 values in epilogue instead of two buffer_store_dwordx2 * Optimize eight waves pipeline: * Improve instruction scheduling * Remove unneeded barriers * Use nontemporal store/load for C and D matrices in the Quant GEMM kernel (they are used once per block but may consume cache that can be better used for matrices A and B) * Use more efficient padding in epilogue with CTransposed ## Test Plan A new test is added for multiple D Quant Gemm (`TestCkTileGemmRowColQuantMultiD/*.RowColQuantMultiDTest`): ``` ninja test_tile_gemm_quant_rowcol && bin/test_tile_gemm_quant_rowcol ``` Testing the large tensor support is not feasible with the current testing infrastructure because of reference calculations on CPU - it takes several minutes to run a single test case. Such cases are tested manually in Aiter. ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
CK Tile Testing Guide
This document describes the test organization and available test targets for CK Tile operations.
Overview
CK Tile tests are organized with multiple levels of granularity to support different development workflows:
- Global test labels - Run tests across all operations
- Operation-specific umbrella targets - Run all tests for a specific operation
- Individual test executables - Run specific tests
Global Test Labels
These targets run tests across all CK operations (not just CK Tile):
ninja smoke
Run fast smoke tests (tests that complete within ~30 seconds on gfx90a).
ninja smoke
ninja regression
Run slower, more comprehensive regression tests.
ninja regression
ninja check
Run ALL available tests in the entire codebase.
ninja check
Operation-Specific Umbrella Targets
These targets allow you to run all tests for a specific CK Tile operation. This is useful when making changes to a particular operation and wanting to validate all related tests without running the entire test suite.
GEMM Operations
ck_tile_gemm_tests
Run all basic GEMM pipeline tests (memory, compute variants, persistent, etc.)
ninja ck_tile_gemm_tests
Test executables included:
test_ck_tile_gemm_pipeline_memtest_ck_tile_gemm_pipeline_compv3test_ck_tile_gemm_pipeline_compv4test_ck_tile_gemm_pipeline_persistenttest_ck_tile_gemm_pipeline_compv6test_ck_tile_gemm_pipeline_comp_async(gfx95 only)test_ck_tile_gemm_pipeline_*_wmmavariants (gfx11/gfx12 only)
ck_tile_gemm_block_scale_tests
Run all GEMM tests with block-scale quantization (AQuant, BQuant, ABQuant, etc.)
ninja ck_tile_gemm_block_scale_tests
Test executables included: 29 test executables covering:
- AQuant tests (memory pipelines, base layouts, prefill, preshuffle, transpose)
- ABQuant tests (base, padding, preshuffle)
- BQuant tests (1D/2D variants, transpose)
- BQuant with PreshuffleB (decode/prefill, 1D/2D)
- BQuant with PreshuffleQuant (decode/prefill, 1D/2D)
- RowColQuant and TensorQuant tests
ck_tile_gemm_streamk_tests
Run all GEMM StreamK tests (tile partitioner, reduction, smoke, extended)
ninja ck_tile_gemm_streamk_tests
Test executables included:
test_ck_tile_streamk_tile_partitionertest_ck_tile_streamk_reductiontest_ck_tile_streamk_smoketest_ck_tile_streamk_extended
ck_tile_grouped_gemm_quant_tests
Run all grouped GEMM quantization tests
ninja ck_tile_grouped_gemm_quant_tests
Test executables included:
test_ck_tile_grouped_gemm_quant_rowcoltest_ck_tile_grouped_gemm_quant_tensortest_ck_tile_grouped_gemm_quant_aquanttest_ck_tile_grouped_gemm_quant_bquanttest_ck_tile_grouped_gemm_quant_bquant_preshuffleb
Other Operations
ck_tile_fmha_tests
Run all FMHA (Flash Multi-Head Attention) tests
ninja ck_tile_fmha_tests
Test executables included: Forward and backward tests for fp16, bf16, fp8bf16, fp32
ck_tile_reduce_tests
Run all reduce operation tests
ninja ck_tile_reduce_tests
Test executables included:
test_ck_tile_reduce2dtest_ck_tile_multi_reduce2d_threadwisetest_ck_tile_multi_reduce2d_multiblock
Individual Test Executables
You can also build and run individual test executables:
Build a specific test
ninja test_ck_tile_gemm_pipeline_mem
Run a specific test directly
./build/bin/test_ck_tile_gemm_pipeline_mem
Run a specific test through ctest
ctest -R test_ck_tile_gemm_pipeline_mem --output-on-failure