Replace EigenBroadcast with ElementwiseBroadcast in ReduceGrad

[AnnaTrainingG]

PR types

Performance optimization

PR changes

OPs

Describe

Reduce EigenBroadcastcase with ElementwiseBroadcast in ReduceGrad
为扩大KP算子覆盖率，统一将Reduce_sum/mean 反向的Eigen适配代码替换为ElementwiseBroadcast Kernel
ReduceGrad 性能统计对比：

op name	case	axise	dtype	优化前 （us)	优化后 （us)	speed up
reduce_sum_grad	[-1L, 2048L, 33L, 33L]	[2,3]	float32	626	160.73	3.89
reduce_sum_grad	[-1L, 2048L, 33L, 33L]	[2,3]	float16	633	83.64	7.57
reduce_sum_grad	[-1L, 8L, 128L]	[1]	float32	2.44	1.71	1.43
reduce_sum_grad	[-1L, 8L, 128L]	[1]	float16	2.84	1.7	1.67
reduce_sum_grad	[30522L, 1024L]	[]	float32	150.38	138.79	1.08
reduce_sum_grad	[30522L, 1024L]	[]	float16	89.81	43.44	2.07
op name	case	axise	dtype	优化前 （us)	优化后 （us)	speed up
reduce_mean_grad	[-1L, 2048L, 33L, 33L]	[2,3]	float32	672	160.71	4.18
reduce_mean_grad	[-1L, 2048L, 33L, 33L]	[2,3]	float16	681	83.1	8.19
reduce_mean_grad	[-1L, 8L, 128L]	[1]	float32	3.115	1.72	1.81
reduce_mean_grad	[-1L, 8L, 128L]	[1]	float16	3.171	1.66	1.91
reduce_mean_grad	[30522L, 1024L]	[]	float32	152.78	138.83	1.10
reduce_mean_grad	[30522L, 1024L]	[]	float16	134.96	69.835	1.93

benchmark 异常说明，与本次PR修改无关，本地测试无性能影响：

op name	shape	dtype	dev (us)	new (us)	speed up
p_norm_1 backward	[300, 128, 128], axis = -1 porder = 3.0	float32	178.26	178.06	1.00112322
matmul_3_backward	[paddle][p_norm] p_norm {   run_tf: True   run_torch: True   axis: -1   porder: 3.0   keepdim: False   x_shape: [300, 128, 128]   x_dtype: float32   atol: 1e-06 [paddle][p_norm] p_norm {   run_tf: True   run_torch: True   axis: -1   porder: 3.0   keepdim: False   x_shape: [300, 128, 128]   x_dtype: float32   atol: 1e-06	float32	12.7867	12.865	0.99391372
matmul_9_forward	[paddle][matmul] matmul {   run_tf: True   run_torch: True   atol: 1.0   transpose_x: False   transpose_y: False   x_shape: [4, 12, 64, 85]   x_dtype: float16   y_shape: [4, 12, 85, 512]   y_dtype: float16 }	float32	26.165	26.146	1.00072669

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhiqiu]

LGTM for shareDataWith

[ZzSean]

LGTM for ci op benchmark