mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-20 06:30:07 +08:00
Lei WangandUbuntu
7b74bb0128
* refactor code * enhance tutorial * Enhance error handling and code generation in CUDA and TileLang components This commit introduces several improvements across multiple files: - Added more informative error messages in GEMM layout checks - Updated CUDA codegen to support more flexible function signature generation - Improved TMA descriptor initialization and kernel dispatch logic - Refined library generation and source code parsing utilities - Enhanced error handling in various adapter and wrapper classes * Add thread tag validation for warp specialization Introduce a ThreadTagChecker to validate that a PrimFunc only uses threadIdx.x before applying warp specialization. This prevents unintended transformations on kernels with complex thread binding and provides a clear warning to users about potential issues with warp specialization. * Update TileLang Profiling and Compilation in Flash Decoding Examples Refactor the profiling and compilation workflow in two flash decoding example scripts: - Replace `tilelang.lower()` and `tilelang.Profiler()` with `tilelang.compile()` - Simplify profiler initialization using `get_profiler()` - Update method calls to use the new profiler and compiled kernel objects - Maintain existing performance benchmarking and validation logic * Refactor and clean up code formatting in TileLang testing and adapter modules This commit includes several code style and formatting improvements: - Adjust whitespace and line breaks in test files - Improve code formatting in CUDA source wrapper and adapter utilities - Enhance readability of function calls and argument handling - Remove unnecessary whitespace and standardize indentation - Simplify function signatures and argument parsing * Refactor CUDA codegen and improve code formatting This commit includes several improvements to CUDA code generation and formatting: - Enhance function signature generation in CodeGenTileLangCUDA - Improve code formatting and readability in CUDA-related files - Simplify parameter handling and type annotations - Clean up whitespace and line breaks in codegen and layout files --------- Co-authored-by: Ubuntu <dlisuser@h100testl730RPS.xu5snccwrbtejcqqalluoku5hb.xx.internal.cloudapp.net>
300 lines
5.3 KiB
Python
300 lines
5.3 KiB
Python
# Copyright (c) Microsoft Corporation.
|
|
# Licensed under the MIT License.
|
|
|
|
from tilelang import tvm as tvm
|
|
import tilelang.testing
|
|
import tilelang as tl
|
|
|
|
|
|
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.Buffer(A_shape, in_dtype),
|
|
B: T.Buffer(B_shape, in_dtype),
|
|
C: T.Buffer((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,
|
|
)
|
|
|
|
mod, params = tl.lower(program)
|
|
mod = tl.Profiler(mod, params, [2], tl.TensorSupplyType.Integer)
|
|
|
|
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
|
|
|
|
mod.assert_allclose(ref_program, atol=1e-2, rtol=1e-2)
|
|
|
|
|
|
def test_gemm_f16f16f16_nn():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float16",
|
|
128,
|
|
256,
|
|
32,
|
|
2,
|
|
)
|
|
|
|
|
|
def test_gemm_f16f16f32_nn():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float32",
|
|
128,
|
|
128,
|
|
32,
|
|
)
|
|
|
|
|
|
def test_gemm_bf16bf16f32_nn():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
False,
|
|
"bfloat16",
|
|
"bfloat16",
|
|
"float32",
|
|
128,
|
|
128,
|
|
32,
|
|
)
|
|
|
|
|
|
def test_gemm_f32f32f32_nn():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
False,
|
|
"float32",
|
|
"float32",
|
|
"float32",
|
|
64,
|
|
128,
|
|
32,
|
|
)
|
|
|
|
|
|
def test_gemm_f16f16f16_tn():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
True,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float16",
|
|
128,
|
|
256,
|
|
32,
|
|
2,
|
|
)
|
|
|
|
|
|
def test_gemm_f16f16f16_nt():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
True,
|
|
"float16",
|
|
"float16",
|
|
"float16",
|
|
128,
|
|
256,
|
|
32,
|
|
2,
|
|
)
|
|
|
|
|
|
def test_gemm_i8i8i32_nt():
|
|
run_gemm(512, 1024, 768, False, True, "int8", "int8", "int32", 128, 128, 64)
|
|
|
|
|
|
def test_gemm_i8i8i32_tn():
|
|
run_gemm(512, 1024, 768, True, False, "int8", "int8", "int32", 128, 128, 64)
|
|
|
|
|
|
def test_gemm_f64f64f64_nt():
|
|
run_gemm(512, 512, 512, False, True, "float64", "float64", "float64", 64, 32, 16)
|
|
|
|
|
|
def test_gemm_f32f32f32_nt():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
True,
|
|
"float32",
|
|
"float32",
|
|
"float32",
|
|
64,
|
|
128,
|
|
32,
|
|
)
|
|
|
|
|
|
def test_gemm_f32f32f32_tn():
|
|
run_gemm(
|
|
512,
|
|
1024,
|
|
768,
|
|
True,
|
|
False,
|
|
"float32",
|
|
"float32",
|
|
"float32",
|
|
64,
|
|
128,
|
|
32,
|
|
)
|
|
|
|
|
|
def test_pad_aligned_f16f16f16_nn():
|
|
run_gemm(
|
|
512 - 8,
|
|
1024 - 32,
|
|
768 - 24,
|
|
False,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float16",
|
|
128,
|
|
256,
|
|
32,
|
|
2,
|
|
)
|
|
|
|
|
|
def test_pad_f16f16f16_nn():
|
|
run_gemm(
|
|
512 - 9,
|
|
1024 - 7,
|
|
768 - 5,
|
|
False,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float16",
|
|
128,
|
|
256,
|
|
32,
|
|
2,
|
|
)
|
|
|
|
|
|
def test_pad_f16f16f32_nn():
|
|
run_gemm(
|
|
512 + 19,
|
|
1024 + 17,
|
|
768 + 15,
|
|
False,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float32",
|
|
128,
|
|
64,
|
|
32,
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|