[CINN] preload scalar tensor for vectorize situation

[ZhangX-21]

PR Category

CINN

PR Types

Improvements

Description

Pcard-88155

[paddle-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.

[zyfncg]

考虑后端的处理效率，这里返回类型建议用const &

[ZhangX-21]

DONE

[zyfncg]

pre_load的name是在哪个阶段引入的？

[ZhangX-21]

向量化处理scalar Tensor含有vectorize axis申请了local buffer

origin：
vectorized_var_33_0.x = max((float16)((float)(vectorized_var_1.x)), var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4)) % 200704) / 196)])
vectorized_var_33_0.y = max((float16)((float)(vectorized_var_1.y)), var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4) + 1) % 200704) / 196)])
vectorized_var_33_0.z = max((float16)((float)(vectorized_var_1.z)), var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4) + 2) % 200704) / 196)])
vectorized_var_33_0.w = max((float16)((float)(vectorized_var_1.w)), var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4) + 3) % 200704) / 196)])

preload:
float16  pre_load_var_2_buffer[4]；
float16* pre_load_var_2_local_0 = pre_load_var_2_buffer;
pre_load_var_2_local_0[0] = var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4)) % 200704) / 196)];
pre_load_var_2_local_0[1] = var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4) + 1) % 200704) / 196)];
pre_load_var_2_local_0[2] = var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4) + 2) % 200704) / 196)];
pre_load_var_2_local_0[3] = var_2[(((((int)blockIdx.x * 1024) + ((int)threadIdx.x * 4) + 3) % 200704) / 196)];

vectorized_var_33_0.x = max((float16)((float)(vectorized_var_1.x)), pre_load_var_2_local_0[0])
vectorized_var_33_0.y = max((float16)((float)(vectorized_var_1.y)), pre_load_var_2_local_0[1])
vectorized_var_33_0.z = max((float16)((float)(vectorized_var_1.z)), pre_load_var_2_local_0[2])
vectorized_var_33_0.w = max((float16)((float)(vectorized_var_1.w)), pre_load_var_2_local_0[3])

在PreLoadScalarTensorWithVectorizeAxis申请Tensor和buffer引入pre_load name。