ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-07 23:57:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

revert back to v1

Browse Source
This commit is contained in:
coderfeli
2025-02-25 03:06:55 +00:00
parent 6934ac0466
commit d87ddebb30
3 changed files with 18 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
example/65_gemm_multiply_multiply/moe_gemm1.cpp
View File

@@ -139,7 +139,7 @@ static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t MNPerXDL = 32;
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
static constexpr ck::index_t Nswizzle = false;
static constexpr ck::index_t Nswizzle = true;
static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType);
static constexpr ck::index_t EVec = 16 / sizeof(EDataType);
@@ -175,9 +175,10 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
4, 1, S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, Nswizzle, true, A0DataType>;
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, Nswizzle, true, A0DataType>;
// kernel 2: 128->32x128x128
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, EDataType>;
// DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, F16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 128, 128, 16, 16, 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, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F8>,
// clang-format on
@@ -257,8 +258,10 @@ int main(int argc, char* argv[])
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1 + sorted_tile_num}));
// max_token_id.mData = {valid_size, 2, 2, 1, 1, 2, 2, 2,2, 2, 2, 2, 2,1,0,0,0};
max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
int eids[] = {0, 0,1, 2,3, 3, 4,4, 5, 5, 6, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
// max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
// int eids[] = {0, 0,1, 2,3, 3, 4,4, 5, 5, 6, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
max_token_id.mData = {valid_size, 0, 2, 4, 6, 8, 10, 12, 14, 16};
int eids[] = {0, 0,1, 1, 2,2, 3,3, 4,4, 5, 5, 6, 6, 7,7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
for (int i = 0; i < sorted_tile_num; i++) {
expert_ids.mData[i] = eids[i];
}