flash-attention/csrc/layer_norm/ln_parallel_fwd_1024.cu

#include "ln_parallel_residual_fwd_kernels.cuh"

// Create forward launch function and register. Macro signature:
//  HIDDEN_SIZE, WTYPE, ITYPE, RYTPE, OTYPE, CTYPE, CTAS_PER_ROW, WARPS_M, WARPS_N, BYTES_PER_LDG

REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp32, fp32, fp32, fp32, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp16, fp32, fp32, fp32, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp32, fp16, fp32, fp16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp16, fp16, fp32, fp16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp32, fp16, fp16, fp16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp32, bf16, fp32, bf16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, bf16, bf16, fp32, bf16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp32, bf16, bf16, bf16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, fp16, fp16, fp16, fp16, fp32, 1, 4, 1, 16);
REGISTER_PARALLEL_FWD_LAUNCHER( 1024, bf16, bf16, bf16, bf16, fp32, 1, 4, 1, 16);