import torch import torch.backends from bitblas import tvm as tvm from tvm import DataType from tvm import tl as TL import tvm.tl.language as T from bitblas.tl.utils import get_swizzle_layout from bitblas.tl.mma_macro_generator import ( TensorCoreIntrinEmitter,) from bitblas.base import simplify_prim_func torch.manual_seed(0) def make_swizzle_layout(shared_buf): dtype = shared_buf.dtype shape = shared_buf.shape can_swizzle = shape[-1] * DataType(dtype).bits == 512 if not can_swizzle: return T.Layout(shape, lambda *args: args) def transform_func(i, j): new_warp_i, new_warp_j = get_swizzle_layout(i, j, shape[-1], dtype) return [new_warp_i, new_warp_j] return T.Layout(shape, transform_func) @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" micro_size_x = micro_size_y = micro_size_k = 16 if out_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) B_shape = (N, K) 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 = TensorCoreIntrinEmitter( 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) thread_bindings = T.thread_binding(0, threads, "threadIdx.x") 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((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, thread_bindings=thread_bindings, ) # Load B into fragment mma_emitter.ldmatrix_b( B_local, B_shared, ki, thread_bindings=thread_bindings, ) # Perform Matrix Multiplication mma_emitter.mma(A_local, B_local, C_local) # Perform STMatrix mma_emitter.stmatrix( C_local, C_shared, thread_bindings=thread_bindings, ) # 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) mod, params = TL.lower(matmul) src_code = mod.imported_modules[0].get_source() # src_code is the generated cuda source assert src_code is not None print(src_code) if in_dtype == "int8": A = torch.randint(-7, 7, (M, K), device="cuda", dtype=torch.int8) B = torch.randint(-7, 7, (N, K), device="cuda", dtype=torch.int8) else: A = torch.rand(M, K, device="cuda", dtype=getattr(torch, in_dtype)) B = torch.rand(N, K, device="cuda", dtype=getattr(torch, in_dtype)) C = torch.zeros(M, N, device="cuda", dtype=getattr(torch, accum_dtype)) mod = TL.Profiler(mod, params, [], TL.TensorSupplyType.Integer) mod(A, B, C) latency = mod.do_bench(mod.func, warmup=25) 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) def test_assert_tl_matmul(): assert_tl_matmul_correctness(128, 128, 128, "float16", "float16", "float16") assert_tl_matmul_correctness(128, 256, 256, "float16", "float32", "float32") if __name__ == "__main__": # bitblas.testing.main() # assert_tl_matmul_correctness(128, 128, 128, "float16", "float16", "float16") # assert_tl_matmul_correctness(128, 128, 128, "int8", "int32", "int32") assert_tl_matmul_correctness(16384, 16384, 16384, "int8", "int32", "int32")