mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-21 06:40:23 +08:00
Lei Wang
2a286ae68c
* [Feature] Add reduce_max functionality and corresponding tests * Introduced a new test file for the reduce_max operation in the tilelang language module. * Implemented the reduce_max functionality using T.prim_func, including local memory allocation and result copying. * Added tests for various input sizes and data types to ensure correctness of the reduce_max implementation. * Enhanced profiling assertions to validate the output against reference implementations. * Fix whitespace issues in reduce_max test file for improved readability * [Refactor] Update DebugOutput methods to return strings instead of void * Modified DebugOutput methods in LayoutNode, FragmentNode, and SwizzledLayoutNode to return std::string instead of void, enhancing usability for logging and debugging. * Updated corresponding header files to reflect the new return types. * Improved layout inference error messages by incorporating DebugOutput for better clarity in layout conflicts. * lint fix * Fix typo in matmul function: changed loop from T.Parallel to T.grid for correct parallel execution in webgpu code generation tests. * [Enhancement] Improve layout inference conflict handling in ParallelOp * Updated the layout inference logic in ParallelOp to better handle conflicts for local.fragment buffers. * Added checks to ensure that layout conflicts are reported only when both source and destination buffers are defined, improving clarity in error messages. * Enhanced the overall robustness of the layout inference process by addressing specific cases where conflicts may arise. * [Feature] Add IsEqual methods for layout comparison * Introduced IsEqual methods in LayoutNode, FragmentNode, and SwizzledLayoutNode to facilitate structural equality checks, allowing for optional index comparison. * Enhanced layout inference logic in Copy and ParallelOp to utilize the new IsEqual methods for better conflict detection in local.fragment layouts. * Improved error messages for layout conflicts to provide clearer guidance on potential issues.houm * [Refactor] Update profiler usage in benchmark_nsa_fwd.py and improve layout inference in elem.cc and parallel.cc * Modified the profiler call in benchmark_nsa_fwd.py to streamline latency measurement. * Updated layout inference logic in elem.cc and parallel.cc to use const pointers for FragmentNode, enhancing type safety and clarity. * Improved error messages in layout conflict checks to provide better guidance on potential issues. * [Refactor] Clean up pointer formatting in layout inference files * Standardized pointer formatting for FragmentNode in elem.cc and parallel.cc to improve code readability. * Minor adjustments to error message formatting in layout conflict checks for better clarity.
64 lines
1.7 KiB
Python
64 lines
1.7 KiB
Python
# Copyright (c) Microsoft Corporation.
|
|
# Licensed under the MIT License.
|
|
|
|
import tilelang
|
|
from tilelang import tvm as tvm
|
|
import tilelang.testing
|
|
import tilelang.language as T
|
|
|
|
|
|
def matmul(M, N, K, block_M, block_N, block_K, dtype="float16", accum_dtype="float"):
|
|
|
|
@T.prim_func
|
|
def main(
|
|
A: T.Buffer((M, K), dtype),
|
|
B: T.Buffer((K, N), dtype),
|
|
C: T.Buffer((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_K, block_N), 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):
|
|
T.copy(A[by * block_M, ko * block_K], A_shared, coalesced_width=2)
|
|
T.copy(B[ko * block_K, bx * block_N], B_shared, coalesced_width=2)
|
|
|
|
for i, j, k in T.grid(block_M, block_N, block_K):
|
|
C_local[i, j] += A_shared[i, k] * B_shared[k, j]
|
|
|
|
T.copy(C_local, C[by * block_M, bx * block_N], coalesced_width=2)
|
|
|
|
return main
|
|
|
|
|
|
def assert_gemm_codegen(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
dtype="float16",
|
|
accum_dtype="float",
|
|
):
|
|
func = matmul(M, N, K, block_M, block_N, block_K, dtype=dtype, accum_dtype=accum_dtype)
|
|
print(func)
|
|
|
|
rt_mod, _ = tilelang.lower(func, target="webgpu")
|
|
|
|
src_code = rt_mod.imported_modules[0].get_source()
|
|
|
|
assert src_code is not None
|
|
|
|
|
|
def test_gemm_codegen():
|
|
assert_gemm_codegen(1024, 1024, 1024, 16, 16, 16)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|