1# mypy: allow-untyped-defs 2import logging 3from typing import List, TYPE_CHECKING 4 5from ..select_algorithm import autotune_select_algorithm, TritonTemplate 6from .mm_common import mm_args, mm_configs, mm_grid, mm_options 7 8 9if TYPE_CHECKING: 10 from ..ir import ChoiceCaller 11 12log = logging.getLogger(__name__) 13 14uint4x2_mixed_mm_template = TritonTemplate( 15 name="uint4x2_mixed_mm", 16 grid=mm_grid, 17 source=r""" 18{{def_kernel("A", "B")}} 19 M = {{size("A", 0)}} 20 N = {{size("B", 1)}} 21 K = {{size("A", 1)}} 22 stride_am = {{stride("A", 0)}} 23 stride_ak = {{stride("A", 1)}} 24 stride_bk = {{stride("B", 0)}} 25 stride_bn = {{stride("B", 1)}} 26 27 # based on triton.ops.matmul 28 pid = tl.program_id(0) 29 grid_m = (M + BLOCK_M - 1) // BLOCK_M 30 grid_n = (N + BLOCK_N - 1) // BLOCK_N 31 32 # re-order program ID for better L2 performance 33 width = GROUP_M * grid_n 34 group_id = pid // width 35 group_size = min(grid_m - group_id * GROUP_M, GROUP_M) 36 pid_m = group_id * GROUP_M + (pid % group_size) 37 pid_n = (pid % width) // (group_size) 38 39 rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M) 40 rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N) 41 ram = tl.max_contiguous(tl.multiple_of(rm % M, BLOCK_M), BLOCK_M) 42 rbn = tl.max_contiguous(tl.multiple_of(rn % N, BLOCK_N), BLOCK_N) 43 rk = tl.arange(0, BLOCK_K) 44 A = A + (ram[:, None] * stride_am + rk[None, :] * stride_ak) 45 B = B + (rk[:, None]//2 * stride_bk + rbn[None, :] * stride_bn) 46 b_shifts = 4*(rk%2) 47 b_subs = 8*(1-(rk%2)) 48 49 acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=ACC_TYPE) 50 for k in range(K, 0, -BLOCK_K): 51 if EVEN_K: 52 a = tl.load(A) 53 b = tl.load(B) 54 else: 55 a = tl.load(A, mask=rk[None, :] < k, other=0.) 56 b = tl.load(B, mask=rk[:, None] < k, other=0.) 57 b = ((b >> b_shifts[:, None]) & 0xF) - 8 58 b = b.to(B_PROLOGUE_CAST_TYPE) 59 acc += tl.dot(a, b, allow_tf32=ALLOW_TF32) 60 A += BLOCK_K * stride_ak 61 B += BLOCK_K//2 * stride_bk 62 63 # rematerialize rm and rn to save registers 64 rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M) 65 rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N) 66 idx_m = rm[:, None] 67 idx_n = rn[None, :] 68 mask = (idx_m < M) & (idx_n < N) 69 70 # inductor generates a suffix 71 {{store_output(("idx_m", "idx_n"), "acc", "mask")}} 72""", 73) 74 75 76def tuned_uint4x2_mixed_mm(mat1, mat2, mat2_mm_shape, mat2_dtype): 77 m, n, k, layout, mat1, mat2 = mm_args(mat1, mat2, layout=None, use_4x2_dim=True) 78 choices: List[ChoiceCaller] = [] 79 b_prologue_cast_type = f"tl.{mat2_dtype}".replace("torch.", "") 80 for config in mm_configs(m, n, k): 81 uint4x2_mixed_mm_template.maybe_append_choice( 82 choices, 83 input_nodes=(mat1, mat2), 84 layout=layout, 85 **mm_options(config, m, n, k, layout, b_prologue_cast_type), 86 ) 87 return autotune_select_algorithm("uint4x2_mixed_mm", choices, [mat1, mat2], layout) 88