fused_embedding_eltwise_layernorm_op and skip_layernorm_op support fp16

[yuanlehome]

PR types

Others

PR changes

Others

Describe

该PR效果：fused_embedding_eltwise_layernorm_op 和 skip_layernorm_op 算子添加对fp16的支持

测试模型：

import paddle
from paddle.static import InputSpec
import numpy as np
import os
import paddle.inference as inference

class EmbEwLnNet(paddle.nn.Layer):
    def __init__(self):
        super(EmbEwLnNet, self).__init__()
        self.embedding_layer1 = paddle.nn.Embedding(1024, 256, sparse=True)
        self.embedding_layer2 = paddle.nn.Embedding(1024, 256, sparse=True)
        self.embedding_layer3 = paddle.nn.Embedding(1024, 256, sparse=True)
        self.layer_norm = paddle.nn.LayerNorm(256)

    def forward(self, x1, x2, x3):
        x = self.embedding_layer1(x1) + self.embedding_layer2(x2)
        x = x + self.embedding_layer3(x3)
        x = self.layer_norm(x)
        return x


class SkipLnNet(paddle.nn.Layer):
    def __init__(self):
        super(SkipLnNet, self).__init__()
        self.layer_norm = paddle.nn.LayerNorm(256)

    def forward(self, x1, x2):
        x = x1 + x2
        x = self.layer_norm(x)
        return x

测试环境： GPU Compute Capability: 7.5, Driver API Version: 11.4, Runtime API Version: 11.2, cuDNN Version: 8.1. with warmup: 100 and repeats: 10000

fused_embedding_eltwise_layernorm_op算子

gpu原生

model	gpu
float	0.336549 ms
float16	0.165548 ms(原生没有cast)

trt加速

model	trt32	trt16
float	0.336706 ms	0.346024 ms
float16	0.338106 ms	0.284455 ms(trt加了cast)

skip_layernorm_op算子

gpu原生

model	gpu
float	0.288717 ms
float16	0.161595 ms(原生没有cast)

trt加速

model	trt32	trt16
float	0.293951 ms	-
float16	-	0.270939 ms(trt加了cast)

[paddle-bot[bot]]

你的PR提交成功，感谢你对开源项目的贡献! 请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。 Your PR has been submitted. Thanks for your contribution! Please wait for the result of CI firstly. See Paddle CI Manual for details.

[CLAassistant]

All committers have signed the CLA.