mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-26 07:22:54 +08:00
Lei Wang
f2e99180d5
* remove llvm build * [Refactor] Update kernel compilation and profiling in examples - Replaced `tilelang.lower` with `tilelang.compile` in multiple example scripts to streamline kernel compilation. - Updated profiling calls to utilize the new `get_profiler` method, enhancing performance measurement consistency. - Adjusted assertions and benchmarking methods to align with the new profiling structure across various examples, ensuring correctness and clarity in performance evaluations. * lint fix * License Update * [Refactor] Improve code formatting and documentation in CUDA header and HIP runtime files - Adjusted formatting in `cuda.h` for better readability, including alignment of comments and struct fields. - Cleaned up whitespace and improved comment clarity in `rt_mod_hip.cc` to enhance code maintainability. * [Refactor] Enhance formatting and clarity in CUDA header and HIP runtime files - Improved comment alignment and readability in `cuda.h`. - Cleaned up whitespace and formatting in `rt_mod_hip.cc` to enhance maintainability. * lint fix * lint fix * lint fix * lint fix * fix * License update * [Enhancement] Update JITKernel to use artifact for kernel source - Assigned the generated artifact to `self.artifact` for better management. - Updated kernel source references to use `artifact.kernel_source` for consistency in execution backend handling. * lint fix * Add @tilelang.testing.requires_llvm decorator to vectorization tests * Enhance setup.py and env.py for library management - Added functionality to remove original files after copying in CMakeBuild. - Updated TVM_LIBRARY_PATH in env.py to include the PyPI build library path for better integration. * Refactor TVM_LIBRARY_PATH assignment for improved readability in env.py * Refactor CMakeBuild file handling in setup.py - Added a check to ensure the target library directory exists before copying .so files. - Improved the logic for creating the target directory and copying files to enhance robustness. * bugfix * Rename BuildTLDebug to BuildTileLangCUDAWithoutCompile and update registration. Add @tilelang.testing.requires_llvm decorator to multiple tests for LLVM requirement. * lint fix * Enhance TileLang code generation by adding support for device code generation without compilation. Updated `host_codegen` and `device_codegen` functions to include new transformations and registration for `tilelang_hip_without_compile`. Refactored JIT kernel adapters to accommodate host and device modules, improving overall integration and flexibility. * lint fix * Add support for C target in device code generation - Updated `device_codegen_without_compile` to include handling for the C target by registering the `tilelang_cpp` function. * [Enhancement] Implement auto-clear cache feature based on environment variable * Added TILELANG_CLEAR_CACHE environment variable to control cache clearing. * Updated CI workflow to set TILELANG_CLEAR_CACHE during testing. * Modified cache initialization to clear cache if TILELANG_CLEAR_CACHE is set to true. * [Refactor] Update kernel invocation and import paths in tests and cache * Changed kernel invocation in `test_tilelang_kernel_dequantize_gemm.py` to return the result. * Updated import statements in `test_tilelang_kernel_int4_gemm_mma.py` to use `bitblas` instead of `tilelang`. * Refactored paths for artifact and parameters in `kernel_cache.py` for better maintainability. * [Refactor] Clean up whitespace and improve code formatting in kernel_cache.py * Removed unnecessary blank lines and adjusted spacing for better readability in the KernelCache class. * Enhanced overall code formatting to align with project standards. * [Enhancement] Add bfloat16 test case and improve kernel caching logic * Introduced a new test case for bfloat16 matrix multiplication in `test_tilelang_kernel_gemm_mma_intrinsic.py`. * Updated `KernelCache` to handle multiple kernel source files and improve error handling during saving and loading. * Refactored `JITKernel` to support instantiation from a database, enhancing flexibility in kernel management. * Adjusted `CtypesKernelAdapter` and `CythonKernelAdapter` to utilize the new kernel loading mechanism from the database. * Improved code formatting and readability across several files. * lint fix * Update bfloat16 matrix multiplication test case to use larger dimensions for improved coverage
404 lines
13 KiB
Python
404 lines
13 KiB
Python
import torch
|
|
import torch.backends
|
|
import tilelang
|
|
from tilelang import tvm as tvm
|
|
import tilelang.testing
|
|
import tilelang.language as T
|
|
from tilelang.intrinsics import (
|
|
make_mma_swizzle_layout as make_swizzle_layout,)
|
|
|
|
from tilelang.intrinsics.mma_macro_generator import (
|
|
INT4TensorCoreIntrinEmitter,
|
|
INT4TensorCoreIntrinEmitterWithLadderTransform,
|
|
)
|
|
from tilelang.transform import simplify_prim_func
|
|
|
|
tilelang.testing.set_random_seed(0)
|
|
|
|
|
|
@simplify_prim_func
|
|
def tl_matmul(
|
|
M,
|
|
N,
|
|
K,
|
|
in_dtype,
|
|
out_dtype,
|
|
accum_dtype,
|
|
):
|
|
assert in_dtype in [
|
|
"float16",
|
|
"int8",
|
|
], "Currently only float16 and int8 are supported"
|
|
assert out_dtype in [
|
|
"float16",
|
|
"float32",
|
|
"int32",
|
|
], "Currently only float16, float32 and int32 are supported"
|
|
|
|
K = K // 2
|
|
|
|
micro_size_x = micro_size_y = micro_size_k = 16
|
|
|
|
if accum_dtype == "int32":
|
|
micro_size_k = 32
|
|
|
|
# This is a debug config
|
|
block_row_warps = 2
|
|
block_col_warps = 2
|
|
warp_row_tiles = 64
|
|
warp_col_tiles = 64
|
|
chunk = 32 if in_dtype == "float16" else 64
|
|
shared_scope = "shared.dyn"
|
|
|
|
# Pipeline Stage
|
|
stage = 2
|
|
|
|
block_M = block_row_warps * warp_row_tiles
|
|
block_N = block_col_warps * warp_col_tiles
|
|
block_K = chunk
|
|
|
|
A_shape = (M, K) # int8 storage represents int4*2
|
|
B_shape = (N, K) # int8 storage represents int4*2
|
|
A_shared_shape = (block_M, block_K)
|
|
B_shared_shape = (block_N, block_K)
|
|
C_shared_shape = (
|
|
block_M // micro_size_x,
|
|
block_N // micro_size_y,
|
|
micro_size_x,
|
|
micro_size_y,
|
|
)
|
|
|
|
warp_size = 32
|
|
threads = warp_size * (block_row_warps * block_col_warps)
|
|
local_size_a = (micro_size_x * micro_size_k) // warp_size
|
|
local_size_b = (micro_size_y * micro_size_k) // warp_size
|
|
local_size_c = (micro_size_x * micro_size_y) // warp_size
|
|
warp_rows = warp_row_tiles // micro_size_x
|
|
warp_cols = warp_col_tiles // micro_size_y
|
|
|
|
# MMA Wrapper to Auto Generate Code for MMA
|
|
mma_emitter = INT4TensorCoreIntrinEmitter(
|
|
a_dtype=in_dtype,
|
|
b_dtype=in_dtype,
|
|
accum_dtype=accum_dtype,
|
|
a_transposed=False,
|
|
b_transposed=True,
|
|
block_row_warps=block_row_warps,
|
|
block_col_warps=block_col_warps,
|
|
warp_row_tiles=warp_row_tiles,
|
|
warp_col_tiles=warp_col_tiles,
|
|
chunk=chunk,
|
|
)
|
|
|
|
@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, scope=shared_scope)
|
|
B_shared = T.alloc_shared(B_shared_shape, in_dtype, scope=shared_scope)
|
|
C_shared = T.alloc_shared(C_shared_shape, out_dtype, scope=shared_scope)
|
|
A_local = T.alloc_local((warp_rows * local_size_a), in_dtype)
|
|
B_local = T.alloc_local((warp_cols * local_size_b), in_dtype)
|
|
C_local = T.alloc_local((warp_rows * warp_cols * local_size_c), accum_dtype)
|
|
|
|
T.annotate_layout({
|
|
A_shared: make_swizzle_layout(A_shared),
|
|
B_shared: make_swizzle_layout(B_shared),
|
|
})
|
|
|
|
# Improve L2 Cache
|
|
T.use_swizzle(panel_size=10)
|
|
|
|
T.clear(C_local)
|
|
|
|
for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=stage):
|
|
|
|
# Load A into shared memory
|
|
for i, k in T.Parallel(block_M, block_K):
|
|
A_shared[i, k] = A[by * block_M + i, ko * block_K + k]
|
|
|
|
# Load B into shared memory
|
|
for j, k in T.Parallel(block_N, block_K):
|
|
B_shared[j, k] = B[bx * block_N + j, ko * block_K + k]
|
|
|
|
for ki in T.serial(0, (block_K // micro_size_k)):
|
|
|
|
# Load A into fragment
|
|
mma_emitter.ldmatrix_a(
|
|
A_local,
|
|
A_shared,
|
|
ki,
|
|
)
|
|
|
|
# Load B into fragment
|
|
mma_emitter.ldmatrix_b(
|
|
B_local,
|
|
B_shared,
|
|
ki,
|
|
)
|
|
|
|
# Perform Matrix Multiplication
|
|
mma_emitter.mma(A_local, B_local, C_local)
|
|
|
|
# Perform STMatrix
|
|
mma_emitter.stmatrix(
|
|
C_local,
|
|
C_shared,
|
|
)
|
|
|
|
# Store shared into global
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
C[by * block_M + i, bx * block_N + j] = C_shared[
|
|
i // micro_size_x,
|
|
j // micro_size_y,
|
|
i % micro_size_x,
|
|
j % micro_size_y,
|
|
]
|
|
|
|
return main
|
|
|
|
|
|
def assert_tl_matmul_correctness(M, N, K, in_dtype, out_dtype, accum_dtype):
|
|
matmul = tl_matmul(M, N, K, in_dtype, out_dtype, accum_dtype)
|
|
kernel = tilelang.compile(matmul, out_idx=[2])
|
|
profiler = kernel.get_profiler()
|
|
|
|
src_code = kernel.get_kernel_source()
|
|
# src_code is the generated cuda source
|
|
assert src_code is not None
|
|
|
|
A = torch.randint(0, 4, (M, K), device="cuda", dtype=getattr(torch, in_dtype))
|
|
B = torch.randint(0, 4, (N, K), device="cuda", dtype=getattr(torch, in_dtype))
|
|
|
|
compressed_A = (A[:, ::2] & 0x0F) + ((A[:, 1::2] & 0x0F) << 4)
|
|
compressed_B = (B[:, ::2] & 0x0F) + ((B[:, 1::2] & 0x0F) << 4)
|
|
C = kernel(compressed_A, compressed_B)
|
|
print(C)
|
|
latency = profiler.do_bench()
|
|
print(latency)
|
|
# Ensure that the latency is not None
|
|
assert latency is not None
|
|
|
|
# Get Reference Result
|
|
ref_c = torch.matmul(A.to(torch.float32), B.T.to(torch.float32)).to(getattr(torch, accum_dtype))
|
|
|
|
print(ref_c)
|
|
torch.testing.assert_close(C, ref_c, rtol=1e-2, atol=1e-2)
|
|
|
|
|
|
def test_assert_tl_matmul_correctness():
|
|
assert_tl_matmul_correctness(128, 128, 128, "int8", "int32", "int32")
|
|
assert_tl_matmul_correctness(128, 128, 64, "int8", "int32", "int32")
|
|
|
|
|
|
@simplify_prim_func
|
|
def tl_matmul_weight_only_transform(
|
|
M,
|
|
N,
|
|
K,
|
|
in_dtype,
|
|
out_dtype,
|
|
accum_dtype,
|
|
):
|
|
K = K // 2
|
|
assert in_dtype in [
|
|
"float16",
|
|
"int8",
|
|
], "Currently only float16 and int8 are supported"
|
|
assert out_dtype in [
|
|
"float16",
|
|
"float32",
|
|
"int32",
|
|
], "Currently only float16, float32 and int32 are supported"
|
|
|
|
micro_size_x = micro_size_y = micro_size_k = 16
|
|
|
|
if out_dtype == "int32":
|
|
micro_size_k = 32
|
|
|
|
transform_b = 3
|
|
|
|
# This is a debug config
|
|
block_row_warps = 2
|
|
block_col_warps = 2
|
|
warp_row_tiles = 64
|
|
warp_col_tiles = 64
|
|
chunk = 32 if in_dtype == "float16" else 64
|
|
shared_scope = "shared.dyn"
|
|
|
|
# Pipeline Stage
|
|
stage = 2
|
|
|
|
block_M = block_row_warps * warp_row_tiles
|
|
block_N = block_col_warps * warp_col_tiles
|
|
block_K = chunk
|
|
|
|
A_shape = (M, K)
|
|
B_shape = (N // micro_size_y, K // micro_size_k, micro_size_y, micro_size_k)
|
|
A_shared_shape = (
|
|
block_M,
|
|
block_K,
|
|
)
|
|
B_shared_shape = (
|
|
block_N // micro_size_y,
|
|
block_K // micro_size_k,
|
|
micro_size_y,
|
|
micro_size_k,
|
|
)
|
|
C_shared_shape = (
|
|
block_M // micro_size_x,
|
|
block_N // micro_size_y,
|
|
micro_size_x,
|
|
micro_size_y,
|
|
)
|
|
warp_size = 32
|
|
threads = warp_size * (block_row_warps * block_col_warps)
|
|
local_size_a = (micro_size_x * micro_size_k) // warp_size
|
|
local_size_b = (micro_size_y * micro_size_k) // warp_size
|
|
local_size_c = (micro_size_x * micro_size_y) // warp_size
|
|
warp_rows = warp_row_tiles // micro_size_x
|
|
warp_cols = warp_col_tiles // micro_size_y
|
|
|
|
# MMA Wrapper to Auto Generate Code for MMA
|
|
mma_emitter = INT4TensorCoreIntrinEmitterWithLadderTransform(
|
|
a_dtype=in_dtype,
|
|
b_dtype=in_dtype,
|
|
accum_dtype=accum_dtype,
|
|
a_transposed=False,
|
|
b_transposed=True,
|
|
block_row_warps=block_row_warps,
|
|
block_col_warps=block_col_warps,
|
|
warp_row_tiles=warp_row_tiles,
|
|
warp_col_tiles=warp_col_tiles,
|
|
chunk=chunk,
|
|
transform_kind_b=transform_b,
|
|
)
|
|
|
|
@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, scope=shared_scope)
|
|
B_shared = T.alloc_shared(B_shared_shape, in_dtype, scope=shared_scope)
|
|
C_shared = T.alloc_shared(C_shared_shape, out_dtype, scope=shared_scope)
|
|
A_local = T.alloc_local((warp_rows * local_size_a), in_dtype)
|
|
B_local = T.alloc_local((warp_cols * local_size_b), in_dtype)
|
|
C_local = T.alloc_local((warp_rows * warp_cols * local_size_c), accum_dtype)
|
|
|
|
T.annotate_layout({
|
|
A_shared: make_swizzle_layout(A_shared),
|
|
B_shared: make_swizzle_layout(B_shared),
|
|
})
|
|
|
|
# Improve L2 Cache
|
|
T.use_swizzle(panel_size=10)
|
|
|
|
T.clear(C_local)
|
|
|
|
for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=stage):
|
|
|
|
# Load A into shared memory
|
|
for i, k in T.Parallel(block_M, block_K):
|
|
A_shared[i, k] = A[by * block_M + i, ko * block_K + k]
|
|
|
|
# Load B into shared memory
|
|
for j, k, jj, kk in T.Parallel(block_N // micro_size_y, block_K // micro_size_k,
|
|
micro_size_y, micro_size_k):
|
|
B_shared[j, k, jj, kk] = B[bx * (block_N // micro_size_y) + j,
|
|
ko * (block_K // micro_size_k) + k, jj, kk]
|
|
|
|
for ki in T.serial(0, (block_K // micro_size_k)):
|
|
|
|
# Load A into fragment
|
|
mma_emitter.ldmatrix_a(
|
|
A_local,
|
|
A_shared,
|
|
ki,
|
|
)
|
|
|
|
# Load B into fragment
|
|
mma_emitter.ldmatrix_b(
|
|
B_local,
|
|
B_shared,
|
|
ki,
|
|
)
|
|
|
|
# Perform Matrix Multiplication
|
|
mma_emitter.mma(A_local, B_local, C_local)
|
|
|
|
# Perform STMatrix
|
|
mma_emitter.stmatrix(
|
|
C_local,
|
|
C_shared,
|
|
)
|
|
|
|
# Store shared into global
|
|
for i, j in T.Parallel(block_M, block_N):
|
|
C[by * block_M + i, bx * block_N + j] = C_shared[
|
|
i // micro_size_x,
|
|
j // micro_size_y,
|
|
i % micro_size_x,
|
|
j % micro_size_y,
|
|
]
|
|
|
|
return main
|
|
|
|
|
|
def assert_tl_matmul_weight_only_transform_correctness(M, N, K, in_dtype, out_dtype, accum_dtype):
|
|
import bitblas
|
|
matmul = tl_matmul_weight_only_transform(M, N, K, in_dtype, out_dtype, accum_dtype)
|
|
kernel = tilelang.compile(matmul, out_idx=[2])
|
|
profiler = kernel.get_profiler()
|
|
|
|
src_code = kernel.get_kernel_source()
|
|
# src_code is the generated cuda source
|
|
assert src_code is not None
|
|
transform_b = 3
|
|
|
|
A = torch.randint(0, 4, (M, K), device="cuda", dtype=getattr(torch, in_dtype))
|
|
B = torch.randint(0, 4, (N, K), device="cuda", dtype=getattr(torch, in_dtype))
|
|
compressed_A = (A[:, ::2] & 0x0F) + ((A[:, 1::2] & 0x0F) << 4)
|
|
compressed_B = (B[:, ::2] & 0x0F) + ((B[:, 1::2] & 0x0F) << 4)
|
|
|
|
ladder_permutate_config = bitblas.ops.LadderPermutateConfig(
|
|
M=N,
|
|
N=(K // 2),
|
|
datatype="int8",
|
|
storage_dtype="int8",
|
|
transform_kind=transform_b,
|
|
transpose_matrix=True,
|
|
)
|
|
|
|
ladder_permutate = bitblas.ops.LadderPermutate(ladder_permutate_config)
|
|
LB = ladder_permutate(compressed_B.cpu()).cuda()
|
|
C = kernel(compressed_A, LB)
|
|
|
|
latency = profiler.do_bench()
|
|
print(f"Latency: {latency}")
|
|
# Ensure that the latency is not None
|
|
assert latency is not None
|
|
|
|
# Get Reference Result
|
|
ref_c = torch.matmul(A.to(torch.float32), B.T.to(torch.float32)).to(getattr(torch, accum_dtype))
|
|
print(C)
|
|
print(ref_c)
|
|
torch.testing.assert_close(C, ref_c, rtol=1e-2, atol=1e-2)
|
|
|
|
|
|
@tilelang.testing.requires_package("bitblas")
|
|
def test_assert_tl_matmul_weight_only_transform():
|
|
assert_tl_matmul_weight_only_transform_correctness(128, 128, 128, "int8", "int32", "int32")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
# tilelang.testing.main()
|
|
test_assert_tl_matmul_weight_only_transform()
|