(
M: int,
N: int,
K: int,
in_dtype: str,
out_dtype: str,
accum_dtype: str,
trans_A: bool,
trans_B: bool,
with_bias: bool = False,
n_partition: Optional[int] = 4,
reduce_thread: Optional[int] = 32,
)
| 12 | |
| 13 | |
| 14 | def gemv_simt( |
| 15 | M: int, |
| 16 | N: int, |
| 17 | K: int, |
| 18 | in_dtype: str, |
| 19 | out_dtype: str, |
| 20 | accum_dtype: str, |
| 21 | trans_A: bool, |
| 22 | trans_B: bool, |
| 23 | with_bias: bool = False, |
| 24 | n_partition: Optional[int] = 4, |
| 25 | reduce_thread: Optional[int] = 32, |
| 26 | ): |
| 27 | assert n_partition is not None, "n_partition must be provided" |
| 28 | assert reduce_thread is not None, ( |
| 29 | "reduce_thread must be provided currently, as related bitblas.gpu.gemv.GEMVsch_outer_reduction_with_config is not implemented" |
| 30 | ) |
| 31 | |
| 32 | assert isinstance(N, int) and isinstance(K, int), "Do not support dynamic N and K Currently" |
| 33 | |
| 34 | assert trans_A is False, "Dequantize only implement for trans_A=False currently" |
| 35 | assert trans_B is True, "Dequantize only implement for trans_B=TRue currently" |
| 36 | |
| 37 | MAX_TRANSACTION_SIZE_IN_BITS = 128 |
| 38 | micro_size_k = MAX_TRANSACTION_SIZE_IN_BITS // DataType(in_dtype).bits |
| 39 | |
| 40 | block_K = reduce_thread * micro_size_k |
| 41 | |
| 42 | A_shape = (M, K) |
| 43 | B_shape = (N, K) |
| 44 | Bias_shape = (N,) |
| 45 | C_shape = (M, N) |
| 46 | |
| 47 | dp4a_size = 4 |
| 48 | use_dp4a = in_dtype == T.int8 and accum_dtype == T.int32 |
| 49 | |
| 50 | @T.prim_func |
| 51 | def main( |
| 52 | A: T.Tensor(A_shape, in_dtype), |
| 53 | B: T.Tensor(B_shape, in_dtype), |
| 54 | Bias: T.Tensor(Bias_shape, out_dtype), |
| 55 | C: T.Tensor(C_shape, out_dtype), |
| 56 | ): |
| 57 | with T.Kernel(T.ceildiv(N, n_partition), M, threads=(reduce_thread, n_partition)) as ( |
| 58 | bx, |
| 59 | by, |
| 60 | ): |
| 61 | A_local = T.alloc_local((micro_size_k,), in_dtype) |
| 62 | B_local = T.alloc_local((micro_size_k,), in_dtype) |
| 63 | accum_res = T.alloc_local((1,), accum_dtype) |
| 64 | reduced_accum_res = T.alloc_local((1,), accum_dtype) |
| 65 | |
| 66 | kr = T.get_thread_binding(0) |
| 67 | ni = T.get_thread_binding(1) |
| 68 | |
| 69 | T.clear(accum_res) |
| 70 | for ko in T.serial(T.ceildiv(K, block_K)): |
| 71 | for v in T.vectorized(micro_size_k): |
no test coverage detected
searching dependent graphs…