mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-27 07:32:58 +08:00
be44758c1e
* [experimental] add a draft gemm_sp * [3rdparty] bump cutlass to v3.9.3 * [lint] run format.sh * [chore] rebase * [chore] use abs path * [gemm_sp] add metadata layout * [ci] add more example * [lint] run format.sh * [chore] polish * [chore] move gemm_sp to experimental * [chore] polish * [lint] run format.sh * [Enhancement] Improve bulk copy handling and update GEMM sparse tensor test * Added a warning log for unsupported non-swizzled global layouts in the bulk copy operation, ensuring fallback to normal copy. * Refactored the GEMM sparse tensor test by removing unnecessary imports and simplifying the kernel compilation process. * Updated the test to directly call the `run_gemm_sp` function, enhancing clarity and functionality. * Implement Test * [Enhancement] Update GEMM SP and SM89 templates for improved functionality * Refactored GEMM SP computation to enhance warp partitioning logic, ensuring compatibility with Hopper architecture. * Updated layout inference to support new WGMMA conditions and improved error messaging for unsupported targets. * Modified SM89 templates to utilize new MMA atom structures, enhancing performance and compatibility with fp8 types. * Added conditional inclusion for GEMM SP header based on CUDA architecture version. * lint fix * [gemm_sp] support more layout and data types * Enhancement: sync T.gemm_sp's layout inference with T.gemm * Enhancement: support more block_k in compress util * [Enhancement] enable block_k=64 * [Lint] run format.sh * [Enhancement] compressor support more dtype * Enhancement: enable block_K=32 * [Lint] format.sh * [Fixbug] fix shape * Refactor: sync gemm * [Enhancement] enable transpose * [Enhancement] enable fp8_e4m3 * [Enhancement] enable int8 * [Lint] run format.sh * [Benchmark] add gemm_sp benchmark * [Example] fix 256 threads hang * [CI] fix ci * [Chore] resolve gemini feedback * [Benchmark] increase search space * [Lint] format * [CI] skip sparse tensor core related tests as only sm90 is supported * [CI] pass local run * Update gemm_sm89.h * lint fix * lint fix * [Enhancement] Add support for sparse GEMM and initialize CUDA architecture flags - Introduced a new boolean flag `enable_sparse_gemm_` to control the inclusion of sparse GEMM functionality in CUDA code generation. - Updated the `Finish` method to conditionally include the sparse GEMM header based on the new flag. - Implemented logic in `VisitStmt_` to enable sparse GEMM when the corresponding external call is detected. - Added a function to initialize the `TORCH_CUDA_ARCH_LIST` environment variable based on the target compute version, enhancing compatibility with PyTorch. - Refactored the initialization function into the appropriate module and ensured it is called in the sparse utilities module. * Update test_compress_utils.py --------- Co-authored-by: LeiWang1999 <leiwang1999@outlook.com> Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>
252 lines
7.3 KiB
Python
252 lines
7.3 KiB
Python
from tilelang import tvm as tvm
|
|
import tilelang.testing
|
|
|
|
|
|
def matmul(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
accum_dtype,
|
|
num_stages,
|
|
threads,
|
|
):
|
|
A_shape = (K, M) if trans_A else (M, K)
|
|
B_shape = (N, K) if trans_B else (K, N)
|
|
A_shared_shape = (block_K, block_M) if trans_A else (block_M, block_K)
|
|
B_shared_shape = (block_N, block_K) if trans_B else (block_K, block_N)
|
|
|
|
import tilelang.language as T
|
|
|
|
@T.prim_func
|
|
def main(
|
|
A: T.Tensor(A_shape, in_dtype),
|
|
B: T.Tensor(B_shape, in_dtype),
|
|
C: T.Tensor((M, N), out_dtype),
|
|
):
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=threads) as (bx, by):
|
|
A_shared = T.alloc_shared(A_shared_shape, in_dtype)
|
|
B_shared = T.alloc_shared(B_shared_shape, in_dtype)
|
|
C_local = T.alloc_fragment((block_M, block_N), accum_dtype)
|
|
T.clear(C_local)
|
|
for k in T.Pipelined(T.ceildiv(K, block_K), num_stages=num_stages):
|
|
if trans_A:
|
|
T.copy(A[k * block_K, by * block_M], A_shared)
|
|
else:
|
|
T.copy(A[by * block_M, k * block_K], A_shared)
|
|
if trans_B:
|
|
T.copy(B[bx * block_N, k * block_K], B_shared)
|
|
else:
|
|
T.copy(B[k * block_K, bx * block_N], B_shared)
|
|
T.gemm(A_shared, B_shared, C_local, trans_A, trans_B)
|
|
T.copy(C_local, C[by * block_M, bx * block_N])
|
|
|
|
return main
|
|
|
|
|
|
def run_gemm(
|
|
M,
|
|
N,
|
|
K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
dtypeAccum,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
num_stages=3,
|
|
num_threads=128,
|
|
):
|
|
program = matmul(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
dtypeAccum,
|
|
num_stages,
|
|
num_threads,
|
|
)
|
|
|
|
kernel = tilelang.compile(program, out_idx=[2])
|
|
profiler = kernel.get_profiler()
|
|
|
|
def ref_program(A, B):
|
|
import torch
|
|
|
|
if trans_A:
|
|
A = A.T
|
|
if trans_B:
|
|
B = B.T
|
|
C = torch.matmul(A.to(torch.float), B.to(torch.float))
|
|
C = C.to(torch.__getattribute__(out_dtype))
|
|
return C
|
|
|
|
profiler.assert_allclose(ref_program, atol=1e-2, rtol=1e-2)
|
|
|
|
|
|
def test_gemm():
|
|
# GEMM tests for float16
|
|
run_gemm(512, 1024, 768, False, False, "float16", "float16", "float16", 128, 256, 32,
|
|
2) # f16f16f16_nn
|
|
run_gemm(512, 1024, 768, True, False, "float16", "float16", "float16", 128, 256, 32,
|
|
2) # f16f16f16_tn
|
|
run_gemm(512, 1024, 768, False, True, "float16", "float16", "float16", 128, 256, 32,
|
|
2) # f16f16f16_nt
|
|
run_gemm(512 - 8, 1024 - 32, 768 - 24, False, False, "float16", "float16", "float16", 128, 256,
|
|
32, 2) # pad_aligned_f16f16f16_nn
|
|
run_gemm(512 - 9, 1024 - 7, 768 - 5, False, False, "float16", "float16", "float16", 128, 256,
|
|
32, 2) # pad_f16f16f16_nn
|
|
|
|
# GEMM tests for mixed precision (float16 + float32)
|
|
run_gemm(512, 1024, 768, False, False, "float16", "float16", "float32", 128, 128,
|
|
16) # f16f16f32_nn
|
|
run_gemm(512, 1024, 768, False, False, "float16", "float16", "float32", 128, 128,
|
|
32) # f16f16f32_nn
|
|
run_gemm(512 + 19, 1024 + 17, 768 + 15, False, False, "float16", "float16", "float32", 128, 64,
|
|
32) # pad_f16f16f32_nn
|
|
|
|
# GEMM tests for bfloat16
|
|
run_gemm(512, 1024, 768, False, False, "bfloat16", "bfloat16", "float32", 128, 128,
|
|
32) # bf16bf16f32_nn
|
|
|
|
# GEMM tests for float32
|
|
run_gemm(512, 1024, 768, False, False, "float32", "float32", "float32", 64, 128,
|
|
32) # f32f32f32_nn
|
|
run_gemm(512, 1024, 768, False, True, "float32", "float32", "float32", 64, 128,
|
|
32) # f32f32f32_nt
|
|
run_gemm(512, 1024, 768, True, False, "float32", "float32", "float32", 64, 128,
|
|
32) # f32f32f32_tn
|
|
|
|
# GEMM tests for float64
|
|
run_gemm(512, 512, 512, False, True, "float64", "float64", "float64", 64, 32,
|
|
16) # f64f64f64_nt
|
|
|
|
# GEMM tests for int8
|
|
run_gemm(512, 1024, 768, False, False, "int8", "int8", "int32", 128, 128, 64) # i8i8i32_nn
|
|
run_gemm(512, 1024, 768, False, True, "int8", "int8", "int32", 128, 128, 64) # i8i8i32_nt
|
|
run_gemm(512, 1024, 768, True, False, "int8", "int8", "int32", 128, 128, 64) # i8i8i32_tn
|
|
|
|
|
|
def matmul_rs(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
accum_dtype,
|
|
num_stages,
|
|
threads,
|
|
):
|
|
A_shape = (K, M) if trans_A else (M, K)
|
|
B_shape = (N, K) if trans_B else (K, N)
|
|
A_shared_shape = (block_K, block_M) if trans_A else (block_M, block_K)
|
|
B_shared_shape = (block_N, block_K) if trans_B else (block_K, block_N)
|
|
A_frag_shape = A_shared_shape
|
|
|
|
import tilelang.language as T
|
|
|
|
@T.prim_func
|
|
def main(
|
|
A: T.Tensor(A_shape, in_dtype),
|
|
B: T.Tensor(B_shape, in_dtype),
|
|
C: T.Tensor((M, N), out_dtype),
|
|
):
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=threads) as (bx, by):
|
|
A_shared = T.alloc_shared(A_shared_shape, in_dtype)
|
|
B_shared = T.alloc_shared(B_shared_shape, in_dtype)
|
|
A_frag = T.alloc_fragment(A_frag_shape, in_dtype)
|
|
C_local = T.alloc_fragment((block_M, block_N), accum_dtype)
|
|
T.clear(C_local)
|
|
for k in T.Pipelined(T.ceildiv(K, block_K), num_stages=num_stages):
|
|
if trans_A:
|
|
T.copy(A[k * block_K, by * block_M], A_shared)
|
|
T.copy(A_shared, A_frag)
|
|
else:
|
|
T.copy(A[by * block_M, k * block_K], A_shared)
|
|
T.copy(A_shared, A_frag)
|
|
if trans_B:
|
|
T.copy(B[bx * block_N, k * block_K], B_shared)
|
|
else:
|
|
T.copy(B[k * block_K, bx * block_N], B_shared)
|
|
T.gemm(A_frag, B_shared, C_local, trans_A, trans_B)
|
|
T.copy(C_local, C[by * block_M, bx * block_N])
|
|
|
|
return main
|
|
|
|
|
|
def run_gemm_rs(
|
|
M,
|
|
N,
|
|
K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
dtypeAccum,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
num_stages=3,
|
|
num_threads=128,
|
|
):
|
|
program = matmul_rs(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
dtypeAccum,
|
|
num_stages,
|
|
num_threads,
|
|
)
|
|
|
|
kernel = tilelang.compile(program, out_idx=[2])
|
|
profiler = kernel.get_profiler()
|
|
|
|
def ref_program(A, B):
|
|
import torch
|
|
|
|
if trans_A:
|
|
A = A.T
|
|
if trans_B:
|
|
B = B.T
|
|
C = torch.matmul(A.to(torch.float), B.to(torch.float))
|
|
C = C.to(torch.__getattribute__(out_dtype))
|
|
return C
|
|
|
|
profiler.assert_allclose(ref_program, atol=1e-2, rtol=1e-2)
|
|
|
|
|
|
def test_gemm_rs():
|
|
# GEMM tests for float16
|
|
run_gemm_rs(512, 1024, 768, False, False, "float16", "float16", "float16", 128, 256, 32, 2)
|
|
run_gemm_rs(512, 1024, 768, False, True, "float16", "float16", "float16", 128, 256, 32, 2)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|