ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-29 19:28:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/develop' into support_engine_benchmark

Browse Source
This commit is contained in:
Yanxing-Shi
2025-05-21 09:48:23 +00:00
parent bb66c2af3e 990d645578
commit a0f1615c09
6 changed files with 100 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
example/65_gemm_multiply_multiply/CMakeLists.txt
View File

@@ -22,6 +22,11 @@ foreach(gpu IN LISTS GPU_TARGETS)
target_compile_options(example_moe_gemm1_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
target_compile_options(example_moe_gemm2_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
endif()
set(GEMM_OPTIONS)
list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
target_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
set(target 1)
endif()
endforeach()

6
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
View File

@@ -141,11 +141,11 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
AElementOp, BElementOp, CDEElementOp, GemmSpec, 256,
128, 128, 128,
16, 16,
16, 16,
8, 2,
32, 32,
4, 1,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 1, S<1, 32, 1, 8>, S<8, 8, 1>,
1, 1, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
// clang-format on

21
example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp
View File

@@ -67,13 +67,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
using TypeConfig = AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>;
using ADataType = typename TypeConfig::ADataType;
using BDataType = typename TypeConfig::BDataType;
using GammaDataType = typename TypeConfig::GammaDataType;
using XDataType = typename TypeConfig::XDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = float;
using ADataType = typename TypeConfig::ADataType;
using BDataType = typename TypeConfig::BDataType;
using GammaDataType = typename TypeConfig::GammaDataType;
using XDataType = typename TypeConfig::XDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = float;
using UnquantYDataType = ck_tile::null_type;
// host verify
ck_tile::HostTensor<ADataType> a_host({m, n}, {stride, 1});
@@ -184,6 +185,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
// Rmsnorm2d
{
ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
ck_tile::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n});
// CAUSION: kernel use ComputeDataType version of x, but we use XDataType here for
// simplicity
@@ -191,8 +193,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
GammaDataType,
ComputeDataType,
YDataType,
InvRmsDataType>(
x_host_ref, gamma_host, y_host, invRms_host_ref, epsilon);
InvRmsDataType,
UnquantYDataType>(
x_host_ref, gamma_host, y_host, invRms_host_ref, unquant_y_host_ref, epsilon);
}
// yscale

16
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp
View File

@@ -285,10 +285,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
static_for<0, KGroup, 1>{}([&](auto kg0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
a_thread_bufs(I0));
});
});
});
@@ -328,10 +328,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_block_buf.At(local_read_buf),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
a_thread_bufs(local_read_buf));
});
});
});
@@ -403,10 +403,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_block_buf.At(local_read_reg),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
a_thread_bufs(local_read_reg));
});
});
});
@@ -460,10 +460,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
a_block_buf,
a_block_buf.At(local_read_reg),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
a_thread_buf);
a_thread_bufs(local_read_reg));
});
});
});

48
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/CMakeLists.txt
View File

@@ -2,6 +2,18 @@
set(GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES)
list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance_v2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p1.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p1.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p2.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p3.cpp
@@ -9,6 +21,18 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p5.cpp
f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p6.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance_v2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p1.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p1.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p2.cpp
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p3.cpp
@@ -17,6 +41,18 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES
f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p6.cpp
)
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p4_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p5_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
@@ -24,6 +60,18 @@ set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p5.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma16x16_mn_compute_default_instance_p6.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p4_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_p5_default_instance_v2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p1.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p2.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(f8_f8_f16/device_gemm_multiply_multiply_wp_xdl_f8_f8_f16_mk_mfma16x16_mn_compute_default_instance_p3.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")

24
profiler/README.md
View File

@@ -1,5 +1,29 @@
[Back to the main page](../README.md)
# Composable Kernel profiler
## Profiler GEMM UNIVERSAL kernels
```bash
# arg1: tensor operation (gemm_universal: Universal GEMM)
# arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16->f8; 7: f8->bf16, comp f8; 8: f16@i4; 9: bf16@i4
# arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];
# 1: A[m, k] * B[n, k] = C[m, n];
# 2: A[k, m] * B[k, n] = C[m, n];
# 3: A[k, m] * B[n, k] = C[m, n])
# arg4: verification (0: no; 1: yes)
# arg5: initialization (0: no init; 1: integer value; 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0=no, 1=yes)
# arg8 to 13: M, N, K, StrideA, StrideB, StrideC
# arg14: split k into mulitiple batch
# optional:
# arg15: number of warm-up cycles (default 1)
# arg16: number of iterations (default 10)
# arg17: memory for rotating buffer (default 0, size in MB)
################ op datatype layout verify init print time M N K StrideA StrideB StrideC SplitK WarmupCycles Iterations MemoryBuffer
./bin/ckProfiler gemm_universal 1 0 1 1 0 1 4096 4096 4096 4096 4096 4096 1 1 10 0
```
## Profile GEMM kernels
```bash
#arg1: tensor operation (gemm=GEMM)