mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-24 07:04:00 +08:00
86 lines
2.9 KiB
Python
86 lines
2.9 KiB
Python
import tilelang
|
|
import tilelang.testing
|
|
import tilelang.language as T
|
|
import pytest
|
|
import torch
|
|
from tilelang.contrib import nvcc
|
|
|
|
|
|
# add decorator @tilelang.jit if you want to return a torch function
|
|
# @tilelang.jit
|
|
def matmul(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32):
|
|
@T.prim_func
|
|
def main(
|
|
A: T.Tensor((M, K), dtype),
|
|
B: T.Tensor((N, K), dtype),
|
|
C: T.Tensor((M, N), dtype),
|
|
):
|
|
# Initialize Kernel Context
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=128) as (bx, by):
|
|
A_shared = T.alloc_shared((block_M, block_K), dtype)
|
|
B_shared = T.alloc_shared((block_N, block_K), dtype)
|
|
C_local = T.alloc_fragment((block_M, block_N), accum_dtype)
|
|
|
|
T.clear(C_local)
|
|
|
|
for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=0):
|
|
# Copy tile of A
|
|
# This is a sugar syntax for parallelized copy
|
|
T.copy(A[by * block_M, ko * block_K], A_shared)
|
|
|
|
T.clear(A_shared)
|
|
|
|
# Demonstrate parallelized copy from global to shared for B
|
|
T.copy(B[bx * block_N, ko * block_K], B_shared)
|
|
|
|
# Perform a tile-level GEMM on the shared buffers
|
|
# Currently we dispatch to the cute/hip on Nvidia/AMD GPUs
|
|
T.gemm(A_shared, B_shared, C_local, transpose_B=True)
|
|
|
|
# Copy result back to global memory
|
|
T.copy(C_local, C[by * block_M, bx * block_N])
|
|
|
|
return main
|
|
|
|
|
|
def run_matmul(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32):
|
|
program = matmul(M, N, K, block_M, block_N, block_K, dtype, accum_dtype)
|
|
kernel = tilelang.compile(program, out_idx=[2])
|
|
import torch
|
|
|
|
a = torch.randn((M, K), dtype=dtype.as_torch()).cuda()
|
|
b = torch.randn((N, K), dtype=dtype.as_torch()).cuda()
|
|
c = kernel(a, b)
|
|
assert torch.allclose(c, torch.zeros_like(c))
|
|
|
|
|
|
def test_matmul():
|
|
run_matmul(1024, 1024, 1024, 128, 128, 32)
|
|
|
|
|
|
@tilelang.testing.requires_cuda_compute_version(10)
|
|
def test_shared_fill_cast_zero_uses_st_bulk():
|
|
if nvcc.get_cuda_version() < (12, 8):
|
|
pytest.skip("st.bulk shared lowering requires CUDA toolkit >= 12.8")
|
|
|
|
@T.prim_func
|
|
def main(out: T.Tensor((128,), T.float32)):
|
|
with T.Kernel(1, threads=128):
|
|
smem = T.alloc_shared((128,), T.float32)
|
|
for i in T.Parallel(128):
|
|
smem[i] = T.float32(1.0)
|
|
T.fill(smem, T.cast(0, T.float32))
|
|
T.copy(smem, out)
|
|
|
|
kernel = tilelang.compile(main, target="cuda")
|
|
assert "tl::st_bulk_shared<512, 0>" in kernel.get_kernel_source()
|
|
|
|
out = torch.empty((128,), dtype=torch.float32, device="cuda")
|
|
kernel(out)
|
|
torch.cuda.synchronize()
|
|
assert torch.allclose(out, torch.zeros_like(out))
|
|
|
|
|
|
if __name__ == "__main__":
|
|
test_matmul()
|