mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-21 06:40:23 +08:00
xyyy1420
26ac2fe45e
This is the follow-up fix mentioned in #1899 for the remaining float4 compilation issue. The CPU C codegen emits vector types (e.g. float4, int32_t4) when VectorizeLoop vectorizes T.copy operations, but common.h had no definitions for these types, causing g++ compilation failures. Add a generic vec_type<T,N> template and type aliases for all combinations that PrintType() can emit (float/double/half/intN_t, lanes 2/4/8/16). Update test_matmul_with_copy_cython to verify end-to-end correctness now that both the wrapper and vector type issues are resolved.
156 lines
5.6 KiB
Python
156 lines
5.6 KiB
Python
import tilelang
|
|
import tilelang.testing
|
|
from tilelang import tvm as tvm
|
|
import tilelang.language as T
|
|
import torch
|
|
|
|
|
|
def matmul(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32):
|
|
num_stages = 0
|
|
|
|
@T.prim_func
|
|
def matmul(
|
|
A: T.Tensor((M, K), dtype),
|
|
B: T.Tensor((K, N), dtype),
|
|
C: T.Tensor((M, N), dtype),
|
|
):
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), is_cpu=True) as (bx, by):
|
|
A_local = T.alloc_local((block_M, block_K), dtype)
|
|
B_local = T.alloc_local((block_K, block_N), dtype)
|
|
C_local = T.alloc_local((block_M, block_N), accum_dtype)
|
|
|
|
T.clear(C_local)
|
|
|
|
# Apply layout optimizations or define your own layout
|
|
# (Optional).
|
|
# T.annotate_layout(
|
|
# {
|
|
# A_local: make_swizzle_layout(A_local),
|
|
# B_local: make_swizzle_layout(B_local),
|
|
# }
|
|
# )
|
|
|
|
for ko in T.Pipelined(K // block_K, num_stages=num_stages):
|
|
T.copy(A[by * block_M, ko * block_K], A_local)
|
|
|
|
# Or Copy with Parallel
|
|
for k, j in T.Parallel(block_K, block_N):
|
|
B_local[k, j] = B[ko * block_K + k, by * block_N + j]
|
|
|
|
for i, j, k in T.grid(block_M, block_N, block_K):
|
|
C_local[i, j] += A_local[i, k] * B_local[k, j]
|
|
|
|
T.copy(C_local, C[by * block_M, bx * block_N])
|
|
|
|
return matmul
|
|
|
|
|
|
def assert_matmul_codegen(M=1024, N=1024, K=1024, block_M=128, block_N=128, block_K=32):
|
|
func = matmul(M, N, K, block_M, block_N, block_K)
|
|
|
|
with tvm.target.Target("c"):
|
|
artifact = tilelang.lower(func)
|
|
|
|
code = artifact.kernel_source
|
|
|
|
assert code is not None, "Code generation failed"
|
|
|
|
|
|
def test_matmul_codegen():
|
|
assert_matmul_codegen(M=1024, N=1024, K=1024, block_M=128, block_N=128, block_K=32)
|
|
|
|
|
|
def test_matmul_compile():
|
|
def matmul_jit_test(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32):
|
|
# a simple kernel just for jit test
|
|
@T.prim_func
|
|
def matmul(
|
|
A: T.Tensor((M, K), dtype),
|
|
B: T.Tensor((K, N), dtype),
|
|
C: T.Tensor((M, N), dtype),
|
|
):
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), is_cpu=True) as (bx, by):
|
|
A_local = T.alloc_local((block_M, block_K), dtype)
|
|
B_local = T.alloc_local((block_K, block_N), dtype)
|
|
C_local = T.alloc_local((block_M, block_N), accum_dtype)
|
|
|
|
for p in T.serial(block_M):
|
|
for w in T.serial(block_N):
|
|
C_local[p, w] = 0
|
|
for ko in T.serial(K // block_K):
|
|
for i in T.serial(block_M):
|
|
for k in T.serial(block_K):
|
|
A_local[i, k] = A[by * block_M + i, ko * block_K + k]
|
|
|
|
for k in T.serial(block_K):
|
|
for j in T.serial(block_N):
|
|
B_local[k, j] = B[ko * block_K + k, bx * block_N + j]
|
|
|
|
for i in T.serial(block_M):
|
|
for j in T.serial(block_N):
|
|
for k in T.serial(block_K):
|
|
C_local[i, j] += A_local[i, k] * B_local[k, j]
|
|
|
|
for i in T.serial(block_M):
|
|
for j in T.serial(block_N):
|
|
C[by * block_M + i, bx * block_N + j] = C_local[i, j]
|
|
|
|
return matmul
|
|
|
|
M, N, K = 1024, 512, 512
|
|
block_M, block_N, block_K = M // 4, N // 4, K // 4
|
|
cpu_func = matmul_jit_test(M, N, K, block_M, block_N, block_K)
|
|
with tvm.target.Target("c"):
|
|
complied_fun = tilelang.compile(cpu_func, -1, execution_backend="cython")
|
|
|
|
in_dtype = T.float16
|
|
A = torch.randn(M, K, dtype=torch.__getattribute__(in_dtype))
|
|
B = torch.randn(K, N, dtype=torch.__getattribute__(in_dtype))
|
|
|
|
C = complied_fun(A, B)
|
|
C_torch = torch.matmul(A, B)
|
|
|
|
tilelang.testing.torch_assert_close(C, C_torch, atol=1e-2, rtol=1e-2, max_mismatched_ratio=0.05)
|
|
|
|
|
|
def test_matmul_with_copy_cython():
|
|
"""CPU kernel using T.copy with cython backend.
|
|
|
|
Verifies that T.copy works end-to-end on CPU: the vectorized copy
|
|
uses vector types (e.g. float4) defined in common.h, and the
|
|
wrapper correctly skips redundant re-lowering.
|
|
"""
|
|
M, N, K = 128, 128, 128
|
|
block_M, block_N, block_K = 32, 32, 32
|
|
|
|
@T.prim_func
|
|
def matmul(
|
|
A: T.Tensor((M, K), "float32"),
|
|
B: T.Tensor((K, N), "float32"),
|
|
C: T.Tensor((M, N), "float32"),
|
|
):
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), is_cpu=True) as (bx, by):
|
|
A_local = T.alloc_local((block_M, block_K), "float32")
|
|
B_local = T.alloc_local((block_K, block_N), "float32")
|
|
C_local = T.alloc_local((block_M, block_N), "float32")
|
|
|
|
T.clear(C_local)
|
|
for ko in T.serial(K // block_K):
|
|
T.copy(A[by * block_M, ko * block_K], A_local)
|
|
T.copy(B[ko * block_K, bx * block_N], B_local)
|
|
for i, j, k in T.grid(block_M, block_N, block_K):
|
|
C_local[i, j] += A_local[i, k] * B_local[k, j]
|
|
T.copy(C_local, C[by * block_M, bx * block_N])
|
|
|
|
compiled = tilelang.compile(matmul, target="c", out_idx=-1, execution_backend="cython")
|
|
|
|
a = torch.randn(M, K, dtype=torch.float32)
|
|
b = torch.randn(K, N, dtype=torch.float32)
|
|
c = compiled(a, b)
|
|
ref = a @ b
|
|
torch.testing.assert_close(c, ref, rtol=1e-5, atol=1e-5)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|