mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-22 06:50:25 +08:00
Chennes
4f442a395e
T.atomic_load and T.atomic_store previously accepted every order in the shared memory-order map, including operation-illegal combinations that trigger device-side assertions in libcu++. Validate memory orders against operation-specific allow lists before lowering, report invalid values with a frontend ValueError, and update the documentation and regression coverage accordingly. Fixes #2574
902 lines
32 KiB
Python
902 lines
32 KiB
Python
import pytest
|
|
import tilelang.testing
|
|
import tilelang.layout
|
|
import tilelang.language as T
|
|
import torch
|
|
|
|
|
|
# ======================= Thread-level atomic add =======================
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_add_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_add(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
T.atomic_add(B[bx * block_M + i, by * block_N + j], A_shared[i, j])
|
|
|
|
return atomic_add
|
|
|
|
|
|
def run_atomic_add(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_add_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B + A.sum(dim=0)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_memory_order_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_with_memory_order(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
T.atomic_add(B[bx * block_M + i, by * block_N + j], A_shared[i, j], memory_order="relaxed")
|
|
|
|
return atomic_with_memory_order
|
|
|
|
|
|
def run_atomic_memory_order(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_memory_order_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B + A.sum(dim=0)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_addx2_program(M, N, block_M, block_N, dtype=T.float16):
|
|
@T.prim_func
|
|
def atomic_addx2(A: T.Tensor((M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
for i, j in T.Parallel(block_M, block_N // 2):
|
|
idx_i = bx * block_M + i
|
|
idx_j = by * block_N + j * 2
|
|
T.atomic_addx2(B[idx_i, idx_j], A[idx_i, idx_j])
|
|
|
|
return atomic_addx2
|
|
|
|
|
|
def run_atomic_addx2(M, N, block_M, block_N, dtype=T.float16):
|
|
kernel = atomic_addx2_program(M, N, block_M, block_N, dtype=dtype)
|
|
|
|
import torch
|
|
|
|
A = torch.randn(M, N, dtype=torch.float32).cuda().to(getattr(torch, dtype))
|
|
B = torch.zeros(M, N, dtype=torch.float32).cuda().to(getattr(torch, dtype))
|
|
ref_B = B.clone()
|
|
|
|
for i in range(M):
|
|
for j in range(0, N - 1, 2):
|
|
ref_B[i, j] += A[i, j]
|
|
ref_B[i, j + 1] += A[i, j + 1]
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_add_mixed_dtype_program(N, src_dtype, dst_dtype):
|
|
@T.prim_func
|
|
def atomic_add(Src: T.Tensor((N,), src_dtype), Out: T.Tensor((N,), dst_dtype)):
|
|
with T.Kernel(threads=1):
|
|
frag = T.alloc_fragment((N,), src_dtype)
|
|
for i in T.Parallel(N):
|
|
frag[i] = Src[i]
|
|
for i in T.Parallel(N):
|
|
T.atomic_add(Out[i], frag[i])
|
|
|
|
return atomic_add
|
|
|
|
|
|
def run_atomic_add_mixed_dtype(N, src_dtype, dst_dtype):
|
|
kernel = atomic_add_mixed_dtype_program(N, src_dtype, dst_dtype)
|
|
assert "AtomicAddx2" in kernel.get_kernel_source()
|
|
|
|
src = torch.arange(1, N + 1, dtype=getattr(torch, src_dtype)).cuda()
|
|
out = torch.zeros(N, dtype=getattr(torch, dst_dtype)).cuda()
|
|
kernel(src, out)
|
|
torch.testing.assert_close(out, src.to(getattr(torch, dst_dtype)), atol=1e-2, rtol=1e-2)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_addx2_mixed_dtype_program(M, N, block_M, block_N, src_dtype, dst_dtype):
|
|
@T.prim_func
|
|
def atomic_addx2(A: T.Tensor((M, N), src_dtype), B: T.Tensor((M, N), dst_dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
for i, j in T.Parallel(block_M, block_N // 2):
|
|
idx_i = bx * block_M + i
|
|
idx_j = by * block_N + j * 2
|
|
T.atomic_addx2(B[idx_i, idx_j], A[idx_i, idx_j])
|
|
|
|
return atomic_addx2
|
|
|
|
|
|
def run_atomic_addx2_mixed_dtype(M, N, block_M, block_N, src_dtype, dst_dtype):
|
|
kernel = atomic_addx2_mixed_dtype_program(M, N, block_M, block_N, src_dtype, dst_dtype)
|
|
assert "AtomicAddx2" in kernel.get_kernel_source()
|
|
|
|
A = torch.randn(M, N, dtype=getattr(torch, src_dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dst_dtype)).cuda()
|
|
ref_B = A.to(getattr(torch, dst_dtype))
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-2, rtol=1e-2)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_different_memory_orders_program(M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_different_orders(
|
|
A: T.Tensor((M, N), dtype), B: T.Tensor((M, N), dtype), C: T.Tensor((M, N), dtype), D: T.Tensor((M, N), dtype)
|
|
):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
idx_i = bx * block_M + i
|
|
idx_j = by * block_N + j
|
|
if idx_i < M and idx_j < N:
|
|
val = A[idx_i, idx_j]
|
|
T.atomic_add(B[idx_i, idx_j], val, memory_order="release")
|
|
T.atomic_max(C[idx_i, idx_j], val, memory_order="relaxed")
|
|
T.atomic_min(D[idx_i, idx_j], val, memory_order="relaxed")
|
|
|
|
return atomic_different_orders
|
|
|
|
|
|
def run_atomic_different_memory_orders(M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_different_memory_orders_program(M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
A = torch.randn(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
C = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
D = torch.full((M, N), float("inf"), dtype=getattr(torch, dtype)).cuda()
|
|
|
|
kernel(A, B, C, D)
|
|
|
|
torch.testing.assert_close(B, A, atol=1e-3, rtol=1e-3)
|
|
torch.testing.assert_close(C, torch.maximum(torch.zeros_like(A), A))
|
|
torch.testing.assert_close(D, torch.minimum(torch.full_like(A, float("inf")), A))
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_addx4_program(M, N, block_M, block_N):
|
|
@T.prim_func
|
|
def atomic_addx4(A: T.Tensor((M, N), T.float32), B: T.Tensor((M, N), T.float32)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
for i, j in T.Parallel(block_M, block_N // 4):
|
|
idx_i = bx * block_M + i
|
|
idx_j = by * block_N + j * 4
|
|
T.atomic_addx4(B[idx_i, idx_j], A[idx_i, idx_j])
|
|
|
|
return atomic_addx4
|
|
|
|
|
|
def run_atomic_addx4(M, N, block_M, block_N):
|
|
kernel = atomic_addx4_program(M, N, block_M, block_N)
|
|
import torch
|
|
|
|
A = torch.randn(M, N, dtype=torch.float32).cuda()
|
|
B = torch.zeros(M, N, dtype=torch.float32).cuda()
|
|
ref_B = B.clone()
|
|
|
|
for i in range(M):
|
|
for j in range(0, N - 3, 4):
|
|
ref_B[i, j] += A[i, j]
|
|
ref_B[i, j + 1] += A[i, j + 1]
|
|
ref_B[i, j + 2] += A[i, j + 2]
|
|
ref_B[i, j + 3] += A[i, j + 3]
|
|
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_addx4_16bit_program(dtype, offset, nthreads):
|
|
@T.prim_func
|
|
def atomic_addx4(A: T.Tensor((16,), dtype), B: T.Tensor((16,), dtype)):
|
|
with T.Kernel(1, threads=nthreads):
|
|
T.atomic_addx4(B[offset], A[offset])
|
|
|
|
return atomic_addx4
|
|
|
|
|
|
def run_atomic_addx4_16bit(dtype, offset, nthreads):
|
|
kernel = atomic_addx4_16bit_program(dtype, offset, nthreads)
|
|
source = kernel.get_kernel_source()
|
|
assert "AtomicAddx4" in source
|
|
|
|
torch_dtype = getattr(torch, str(dtype))
|
|
A = torch.zeros(16, dtype=torch_dtype, device="cuda")
|
|
B_init = torch.zeros(16, dtype=torch_dtype, device="cuda")
|
|
A[offset : offset + 4] = torch.tensor([1, 2, 3, 4], dtype=torch_dtype, device="cuda")
|
|
B_init[offset : offset + 4] = torch.tensor([10, 20, 30, 40], dtype=torch_dtype, device="cuda")
|
|
B = B_init.clone()
|
|
|
|
ref_B = B_init.float()
|
|
ref_B[offset : offset + 4] += nthreads * A[offset : offset + 4].float()
|
|
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B.float(), ref_B, atol=0, rtol=0)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_return_prev_program(M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_with_return_prev(A: T.Tensor((M, N), dtype), B: T.Tensor((M, N), dtype), old_vals: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
idx_i = bx * block_M + i
|
|
idx_j = by * block_N + j
|
|
if idx_i < M and idx_j < N:
|
|
old_vals[idx_i, idx_j] = T.atomic_add(B[idx_i, idx_j], A[idx_i, idx_j], return_prev=True)
|
|
|
|
return atomic_with_return_prev
|
|
|
|
|
|
def run_atomic_return_prev(M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_return_prev_program(M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
A = torch.ones(M, N, dtype=getattr(torch, dtype)).cuda() * 5.0
|
|
B = torch.ones(M, N, dtype=getattr(torch, dtype)).cuda() * 2.0
|
|
old_vals = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
|
|
initial_B = B.clone()
|
|
kernel(A, B, old_vals)
|
|
|
|
torch.testing.assert_close(old_vals, initial_B, atol=1e-3, rtol=1e-3)
|
|
torch.testing.assert_close(B, initial_B + A, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def tma_atomic_add_program(out, explicit_swizzle=False):
|
|
out: T.Tensor[(16, 16), T.float32]
|
|
|
|
with T.Kernel(
|
|
1,
|
|
):
|
|
out_shared = T.alloc_shared((16, 16), dtype=T.float32)
|
|
if explicit_swizzle:
|
|
T.annotate_layout({out_shared: tilelang.layout.make_swizzled_layout(out_shared)})
|
|
T.fill(out_shared, 1)
|
|
for _ in range(16):
|
|
T.atomic_add(out, out_shared, use_tma=True)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_tma_atomic_add():
|
|
out = torch.zeros((16, 16), dtype=torch.float32, device="cuda")
|
|
tma_atomic_add_program(out)
|
|
torch.testing.assert_close(out, torch.ones((16, 16), dtype=torch.float32, device="cuda") * 16)
|
|
|
|
kernel = tma_atomic_add_program.compile(out=T.Tensor[(16, 16), T.float32])
|
|
assert "tma_store_add" in kernel.get_kernel_source()
|
|
assert "desc" in kernel.get_kernel_source() # Ensure using cp.reduce.async.bulk.tensor
|
|
|
|
kernel_with_explicit_swizzle = tma_atomic_add_program.compile(out=T.Tensor[(16, 16), T.float32], explicit_swizzle=True)
|
|
# Ensure auto swizzled layout is applied
|
|
assert kernel.get_kernel_source() == kernel_with_explicit_swizzle.get_kernel_source()
|
|
|
|
|
|
def run_atomic_add_auto_vectorized(K, M, N, block_M, block_N, dtype=T.float32):
|
|
tilelang.disable_cache()
|
|
kernel = atomic_add_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
assert "AtomicAddx4" in kernel.get_kernel_source()
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_add_auto_vectorized_unit_test(vec_size: int, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_addx2(A: T.Tensor((vec_size,), dtype)):
|
|
with T.Kernel(threads=1):
|
|
A_local = T.alloc_fragment((vec_size,), dtype)
|
|
for i in T.Parallel(vec_size):
|
|
T.atomic_add(A[i], A_local[i])
|
|
|
|
return atomic_addx2
|
|
|
|
|
|
def run_atomic_add_auto_vectorized_unit_test(vec_size: int, dtype=T.float32):
|
|
kernel = atomic_add_auto_vectorized_unit_test(vec_size, dtype)
|
|
assert f"AtomicAddx{vec_size}" in kernel.get_kernel_source()
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_add_complicated_parallel_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_add(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
value = A_shared[i, j]
|
|
T.atomic_add(B[bx * block_M + i, by * block_N + j], value)
|
|
|
|
return atomic_add
|
|
|
|
|
|
def run_atomic_add_complicated_parallel(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_add_complicated_parallel_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
assert "float4 value" in kernel.get_kernel_source()
|
|
assert "AtomicAddx4" in kernel.get_kernel_source()
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_memory_order():
|
|
run_atomic_memory_order(4, 64, 64, 16, 16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx2_half():
|
|
run_atomic_addx2(32, 64, 8, 16, dtype=T.float16)
|
|
|
|
|
|
def test_atomic_addx2_float():
|
|
run_atomic_addx2(32, 64, 8, 16, dtype=T.float32)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_add_mixed_dtype_fp16():
|
|
run_atomic_add_mixed_dtype(8, T.float32, T.float16)
|
|
run_atomic_addx2_mixed_dtype(32, 64, 8, 16, T.float32, T.float16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
@tilelang.testing.requires_cuda_compute_version_ge(8, 0)
|
|
def test_atomic_add_mixed_dtype_bf16():
|
|
run_atomic_add_mixed_dtype(8, T.float32, T.bfloat16)
|
|
run_atomic_addx2_mixed_dtype(32, 64, 8, 16, T.float32, T.bfloat16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_different_memory_orders():
|
|
run_atomic_different_memory_orders(32, 32, 8, 8, dtype=T.float32)
|
|
run_atomic_different_memory_orders(32, 32, 8, 8, dtype=T.float16)
|
|
run_atomic_different_memory_orders(32, 32, 8, 8, dtype=T.bfloat16)
|
|
|
|
|
|
def test_atomic_addx4():
|
|
run_atomic_addx4(16, 64, 4, 4)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
@tilelang.testing.requires_cuda_compute_version_ge(8, 0)
|
|
def test_atomic_addx4_16bit():
|
|
for dtype in (T.float16, T.bfloat16):
|
|
for offset in (0, 4):
|
|
run_atomic_addx4_16bit(dtype, offset=offset, nthreads=2)
|
|
|
|
|
|
def test_atomic_return_prev():
|
|
run_atomic_return_prev(32, 32, 8, 8)
|
|
|
|
|
|
def test_atomic_add():
|
|
run_atomic_add(8, 128, 128, 32, 32)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
@tilelang.testing.requires_cuda_compute_version_ge(9, 0)
|
|
def test_atomic_add_auto_vectorized():
|
|
run_atomic_add_auto_vectorized(8, 128, 128, 32, 32, dtype=T.float32)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
@tilelang.testing.requires_cuda_compute_version_ge(9, 0)
|
|
def test_atomic_add_auto_vectorized_unit_test():
|
|
run_atomic_add_auto_vectorized_unit_test(2, dtype=T.float32)
|
|
run_atomic_add_auto_vectorized_unit_test(4, dtype=T.float32)
|
|
run_atomic_add_auto_vectorized_unit_test(2, dtype=T.float16)
|
|
run_atomic_add_auto_vectorized_unit_test(2, dtype=T.bfloat16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda_compute_version_ge(9, 0)
|
|
def test_atomic_add_complicated_parallel():
|
|
run_atomic_add_complicated_parallel(8, 128, 128, 32, 32, dtype=T.float32)
|
|
|
|
|
|
# ======================= Tile-level atomic add =======================
|
|
|
|
|
|
@tilelang.jit
|
|
def tile_atomic_add_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_add(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
T.atomic_add(B[bx * block_M, by * block_N], A_shared)
|
|
|
|
return atomic_add
|
|
|
|
|
|
def run_tile_atomic_add(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = tile_atomic_add_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B + A.sum(dim=0)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def tile_atomic_add_expr_program(M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_add(A: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
T.atomic_add(A[bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], 1.0)
|
|
|
|
return atomic_add
|
|
|
|
|
|
def run_tile_atomic_add_expr(M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = tile_atomic_add_expr_program(M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
def ref_program(A):
|
|
for i in range(M):
|
|
for j in range(N):
|
|
A[i, j] += 1
|
|
|
|
A = torch.zeros(M, N, dtype=torch.float32).cuda()
|
|
ref_A = A.clone()
|
|
ref_program(ref_A)
|
|
kernel(A)
|
|
torch.testing.assert_close(A, ref_A, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def tile_atomic_add_scalar_program(dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_add(A: T.Tensor((1), dtype), B: T.Tensor((1), dtype)):
|
|
with T.Kernel(
|
|
1,
|
|
) as _:
|
|
A_local = T.alloc_local([1], dtype)
|
|
T.copy(A, A_local)
|
|
T.clear(B)
|
|
T.atomic_add(B, A_local)
|
|
T.atomic_add(B, 1)
|
|
|
|
return atomic_add
|
|
|
|
|
|
def run_tile_atomic_add_scalar(dtype=T.float32):
|
|
kernel = tile_atomic_add_scalar_program(dtype=dtype)
|
|
import torch
|
|
|
|
def ref_program(A, B):
|
|
B[0] = A[0] + 1
|
|
|
|
A = torch.randn(1, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(1, dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B.clone()
|
|
ref_program(A, ref_B)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
def test_tile_atomic_add():
|
|
run_tile_atomic_add(8, 128, 128, 32, 32)
|
|
|
|
|
|
def test_tile_atomic_add_expr():
|
|
run_tile_atomic_add_expr(128, 128, 32, 32)
|
|
|
|
|
|
def test_tile_atomic_add_scalar():
|
|
run_tile_atomic_add_scalar()
|
|
|
|
|
|
# ======================= Thread-level atomic max/min/load store =======================
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_max_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_max(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
T.atomic_max(B[bx * block_M + i, by * block_N + j], A_shared[i, j])
|
|
|
|
return atomic_max
|
|
|
|
|
|
def run_atomic_max(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_max_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
def ref_program(A, B):
|
|
for k in range(K):
|
|
for i in range(M):
|
|
for j in range(N):
|
|
B[i, j] = max(B[i, j], A[k, i, j])
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B.clone()
|
|
ref_program(A, ref_B)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_min_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_min(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
T.atomic_min(B[bx * block_M + i, by * block_N + j], A_shared[i, j])
|
|
|
|
return atomic_min
|
|
|
|
|
|
def run_atomic_min(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_min_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
def ref_program(A, B):
|
|
for k in range(K):
|
|
for i in range(M):
|
|
for j in range(N):
|
|
B[i, j] = min(B[i, j], A[k, i, j])
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.full((M, N), float("inf"), dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B.clone()
|
|
ref_program(A, ref_B)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def atomic_load_store_program(M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_load_store(A: T.Tensor((M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
idx_i = bx * block_M + i
|
|
idx_j = by * block_N + j
|
|
if idx_i < M and idx_j < N:
|
|
val = T.atomic_load(A[idx_i, idx_j])
|
|
T.atomic_store(B[idx_i, idx_j], val)
|
|
|
|
return atomic_load_store
|
|
|
|
|
|
def run_atomic_load_store(M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = atomic_load_store_program(M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
A = torch.randn(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.zeros(M, N, dtype=getattr(torch, dtype)).cuda()
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, A, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
def _build_atomic_load_with_memory_order(memory_order):
|
|
@T.prim_func
|
|
def kernel(source: T.Tensor((1,), T.int32), output: T.Tensor((1,), T.int32)):
|
|
with T.Kernel(1, threads=1):
|
|
output[0] = T.atomic_load(source[0], memory_order=memory_order)
|
|
|
|
return kernel
|
|
|
|
|
|
def _build_atomic_store_with_memory_order(memory_order):
|
|
@T.prim_func
|
|
def kernel(destination: T.Tensor((1,), T.int32)):
|
|
with T.Kernel(1, threads=1):
|
|
T.atomic_store(destination[0], 1, memory_order=memory_order)
|
|
|
|
return kernel
|
|
|
|
|
|
@pytest.mark.parametrize("memory_order", ["relaxed", "consume", "acquire", "seq_cst"])
|
|
def test_atomic_load_accepts_valid_memory_orders(memory_order):
|
|
assert _build_atomic_load_with_memory_order(memory_order) is not None
|
|
|
|
|
|
@pytest.mark.parametrize("memory_order", ["release", "acq_rel", "invalid"])
|
|
def test_atomic_load_rejects_unsupported_memory_orders(memory_order):
|
|
with pytest.raises(ValueError, match="atomic_load does not support memory_order"):
|
|
_build_atomic_load_with_memory_order(memory_order)
|
|
|
|
|
|
@pytest.mark.parametrize("memory_order", ["relaxed", "release", "seq_cst"])
|
|
def test_atomic_store_accepts_valid_memory_orders(memory_order):
|
|
assert _build_atomic_store_with_memory_order(memory_order) is not None
|
|
|
|
|
|
@pytest.mark.parametrize("memory_order", ["consume", "acquire", "acq_rel", "invalid"])
|
|
def test_atomic_store_rejects_unsupported_memory_orders(memory_order):
|
|
with pytest.raises(ValueError, match="atomic_store does not support memory_order"):
|
|
_build_atomic_store_with_memory_order(memory_order)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_or_codegen():
|
|
@T.prim_func
|
|
def atomic_or_kernel(A: T.Tensor((1,), T.int32), mask: T.int32):
|
|
with T.Kernel(1, threads=32):
|
|
T.atomic_or(A[0], mask, memory_order="release")
|
|
|
|
kernel = tilelang.compile(atomic_or_kernel, out_idx=[0], target="cuda")
|
|
source = kernel.get_kernel_source()
|
|
assert "AtomicOr" in source
|
|
|
|
|
|
def test_atomic_max():
|
|
run_atomic_max(4, 64, 64, 16, 16)
|
|
|
|
|
|
def test_atomic_min():
|
|
run_atomic_min(4, 64, 64, 16, 16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_load_store():
|
|
run_atomic_load_store(64, 64, 16, 16)
|
|
|
|
|
|
# ======================= Tile-level atomic max/min =======================
|
|
|
|
|
|
@tilelang.jit
|
|
def tile_atomic_max_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def tile_atomic_max(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
T.atomic_max(B[bx * block_M, by * block_N], A_shared)
|
|
|
|
return tile_atomic_max
|
|
|
|
|
|
def run_tile_atomic_max(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = tile_atomic_max_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
|
|
def ref_program(A, B):
|
|
for k in range(K):
|
|
for i in range(M):
|
|
for j in range(N):
|
|
B[i, j] = max(B[i, j], A[k, i, j])
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.full((M, N), float("-inf"), dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B.clone()
|
|
ref_program(A, ref_B)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def tile_atomic_min_program(K, M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def tile_atomic_min(A: T.Tensor((K, M, N), dtype), B: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), K, threads=32) as (bx, by, bz):
|
|
A_shared = T.alloc_shared((block_M, block_N), dtype)
|
|
|
|
T.copy(A[bz, bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], A_shared)
|
|
|
|
T.atomic_min(B[bx * block_M, by * block_N], A_shared)
|
|
|
|
return tile_atomic_min
|
|
|
|
|
|
def run_tile_atomic_min(K, M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = tile_atomic_min_program(K, M, N, block_M, block_N, dtype=dtype)
|
|
|
|
def ref_program(A, B):
|
|
for k in range(K):
|
|
for i in range(M):
|
|
for j in range(N):
|
|
B[i, j] = min(B[i, j], A[k, i, j])
|
|
|
|
A = torch.randn(K, M, N, dtype=getattr(torch, dtype)).cuda()
|
|
B = torch.full((M, N), float("inf"), dtype=getattr(torch, dtype)).cuda()
|
|
ref_B = B.clone()
|
|
ref_program(A, ref_B)
|
|
kernel(A, B)
|
|
torch.testing.assert_close(B, ref_B, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
@tilelang.jit
|
|
def tile_atomic_max_expr_program(M, N, block_M, block_N, dtype=T.float32):
|
|
@T.prim_func
|
|
def atomic_max(A: T.Tensor((M, N), dtype)):
|
|
with T.Kernel(T.ceildiv(M, block_M), T.ceildiv(N, block_N), threads=32) as (bx, by):
|
|
T.atomic_max(A[bx * block_M : (bx + 1) * block_M, by * block_N : (by + 1) * block_N], 0.5)
|
|
|
|
return atomic_max
|
|
|
|
|
|
def run_tile_atomic_max_expr(M, N, block_M, block_N, dtype=T.float32):
|
|
kernel = tile_atomic_max_expr_program(M, N, block_M, block_N, dtype=dtype)
|
|
import torch
|
|
|
|
def ref_program(A):
|
|
for i in range(M):
|
|
for j in range(N):
|
|
A[i, j] = max(A[i, j], 0.5)
|
|
|
|
A = torch.randn(M, N, dtype=torch.float32).cuda()
|
|
ref_A = A.clone()
|
|
ref_program(ref_A)
|
|
kernel(A)
|
|
torch.testing.assert_close(A, ref_A, atol=1e-3, rtol=1e-3)
|
|
|
|
|
|
def test_tile_atomic_max():
|
|
run_tile_atomic_max(8, 128, 128, 32, 32)
|
|
|
|
|
|
def test_tile_atomic_min():
|
|
run_tile_atomic_min(8, 128, 128, 32, 32)
|
|
|
|
|
|
def test_tile_atomic_max_expr():
|
|
run_tile_atomic_max_expr(128, 128, 32, 32)
|
|
|
|
|
|
def atomic_scalar_return_prev_program(op_name, val):
|
|
atom = getattr(T, op_name)
|
|
|
|
@T.prim_func
|
|
def main(Dst: T.Tensor((1,), "float32"), Prev: T.Tensor((1,), "float32")):
|
|
with T.Kernel(1, threads=1):
|
|
Prev[0] = atom(Dst[0], val, return_prev=True)
|
|
|
|
return main
|
|
|
|
|
|
def run_atomic_scalar_return_prev(op_name):
|
|
# Pick val so the op actually changes dst (else a silent no-op would still
|
|
# pass): max needs val > dst, min needs val < dst.
|
|
val = 5.0 if op_name == "atomic_max" else 1.0
|
|
kernel = tilelang.compile(atomic_scalar_return_prev_program(op_name, val))
|
|
dst = torch.tensor([3.0], dtype=torch.float32).cuda()
|
|
prev = torch.zeros(1, dtype=torch.float32).cuda()
|
|
kernel(dst, prev)
|
|
assert prev.item() == 3.0, f"{op_name} return_prev should be the old value"
|
|
expected = max(3.0, val) if op_name == "atomic_max" else min(3.0, val)
|
|
assert dst.item() == expected, f"{op_name} should still update Dst"
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_scalar_return_prev():
|
|
run_atomic_scalar_return_prev("atomic_max")
|
|
run_atomic_scalar_return_prev("atomic_min")
|
|
|
|
|
|
def atomic_addx2_return_prev_program(dtype=T.float32):
|
|
@T.prim_func
|
|
def main(Dst: T.Tensor((2,), dtype), Val: T.Tensor((2,), dtype), Prev: T.Tensor((2,), dtype)):
|
|
with T.Kernel(1, threads=1):
|
|
Prev[0:2] = T.atomic_addx2(Dst[0], Val[0], return_prev=True)
|
|
|
|
return main
|
|
|
|
|
|
def atomic_addx4_return_prev_program(dtype=T.float32):
|
|
@T.prim_func
|
|
def main(Dst: T.Tensor((4,), dtype), Val: T.Tensor((4,), dtype), Prev: T.Tensor((4,), dtype)):
|
|
with T.Kernel(1, threads=1):
|
|
Prev[0:4] = T.atomic_addx4(Dst[0], Val[0], return_prev=True)
|
|
|
|
return main
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx2_return_prev():
|
|
kernel = tilelang.compile(atomic_addx2_return_prev_program(T.float32))
|
|
assert "AtomicAddx2Ret" in kernel.get_kernel_source()
|
|
dst = torch.tensor([1.0, 2.0], dtype=torch.float32).cuda()
|
|
val = torch.tensor([10.0, 20.0], dtype=torch.float32).cuda()
|
|
prev = torch.zeros(2, dtype=torch.float32).cuda()
|
|
kernel(dst, val, prev)
|
|
torch.testing.assert_close(prev, torch.tensor([1.0, 2.0], device="cuda"))
|
|
torch.testing.assert_close(dst, torch.tensor([11.0, 22.0], device="cuda"))
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx4_return_prev():
|
|
kernel = tilelang.compile(atomic_addx4_return_prev_program(T.float32))
|
|
assert "AtomicAddx4Ret" in kernel.get_kernel_source()
|
|
dst = torch.tensor([1.0, 2.0, 3.0, 4.0], dtype=torch.float32).cuda()
|
|
val = torch.tensor([10.0, 20.0, 30.0, 40.0], dtype=torch.float32).cuda()
|
|
prev = torch.zeros(4, dtype=torch.float32).cuda()
|
|
kernel(dst, val, prev)
|
|
torch.testing.assert_close(prev, torch.tensor([1.0, 2.0, 3.0, 4.0], device="cuda"))
|
|
torch.testing.assert_close(dst, torch.tensor([11.0, 22.0, 33.0, 44.0], device="cuda"))
|
|
|
|
|
|
def run_atomic_addx2_return_prev_16bit(dtype, torch_dtype):
|
|
# The 16-bit packed 2-lane vector is stored as uint1 in codegen, while
|
|
# AtomicAddx2Ret returns the native __half2 / __nv_bfloat162. The ret path
|
|
# must bridge with tl::to_uint1 so the store LHS (uint1) matches; otherwise
|
|
# nvcc fails with `no operator "=" ... uint1 = half2`.
|
|
kernel = tilelang.compile(atomic_addx2_return_prev_program(dtype))
|
|
src = kernel.get_kernel_source()
|
|
assert "tl::to_uint1(AtomicAddx2Ret" in src, src
|
|
dst = torch.tensor([1.0, 2.0], dtype=torch_dtype).cuda()
|
|
val = torch.tensor([10.0, 20.0], dtype=torch_dtype).cuda()
|
|
prev = torch.zeros(2, dtype=torch_dtype).cuda()
|
|
kernel(dst, val, prev)
|
|
torch.testing.assert_close(prev, torch.tensor([1.0, 2.0], dtype=torch_dtype, device="cuda"))
|
|
torch.testing.assert_close(dst, torch.tensor([11.0, 22.0], dtype=torch_dtype, device="cuda"))
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx2_return_prev_fp16():
|
|
run_atomic_addx2_return_prev_16bit(T.float16, torch.float16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx2_return_prev_bf16():
|
|
run_atomic_addx2_return_prev_16bit(T.bfloat16, torch.bfloat16)
|
|
|
|
|
|
def run_atomic_addx4_return_prev_16bit(dtype, torch_dtype):
|
|
# There is NO single-atomic fp16x4/bf16x4 add in hardware (PTX vector
|
|
# atomic .v4 tops out at .f16x2 for 16-bit types). The x4 ret path realizes
|
|
# the quad as two per-pair AtomicAddx2Ret calls and returns them packed as
|
|
# a uint2 (a half4/bf16x4 is stored as uint2), matching the store LHS. This
|
|
# is per-pair atomic, not whole-quad -- the same contract as the fp32-x4
|
|
# scalar fallback.
|
|
kernel = tilelang.compile(atomic_addx4_return_prev_program(dtype))
|
|
src = kernel.get_kernel_source()
|
|
assert "AtomicAddx4Ret" in src, src
|
|
dst = torch.tensor([1.0, 2.0, 3.0, 4.0], dtype=torch_dtype).cuda()
|
|
val = torch.tensor([10.0, 20.0, 30.0, 40.0], dtype=torch_dtype).cuda()
|
|
prev = torch.zeros(4, dtype=torch_dtype).cuda()
|
|
kernel(dst, val, prev)
|
|
torch.testing.assert_close(prev, torch.tensor([1.0, 2.0, 3.0, 4.0], dtype=torch_dtype, device="cuda"))
|
|
torch.testing.assert_close(dst, torch.tensor([11.0, 22.0, 33.0, 44.0], dtype=torch_dtype, device="cuda"))
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx4_return_prev_fp16():
|
|
run_atomic_addx4_return_prev_16bit(T.float16, torch.float16)
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
def test_atomic_addx4_return_prev_bf16():
|
|
run_atomic_addx4_return_prev_16bit(T.bfloat16, torch.bfloat16)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|