[CINN Performance] Adjust Spatial Tile Config

[jiahy0825]

PR Category

CINN

PR Types

Performance

Description

pcard-76996

Test Case:

Op: paddle.sin
Shape: [1024, 4096], static shape
Dtype: float16
Improvement: 64.12%, performance is on par with phi

Before

__global__
void __launch_bounds__(256) fn_sin_0__COND__FPA__FPA__FPA_4194304llGE1ll_BPA_AND_FPA_4194304llLE2147483647ll_BPA__BPA_AND_FPA__FPA_1GE1ll_BPA_AND_FPA_1LE1ll_BPA__BPA__BPA___kernel(const float16* __restrict__ var, float16* __restrict__ var_0)
{
  if (((int)blockIdx.x < 16384)) {
    if (((int)threadIdx.x < 256)) {
      var_0[((((int)blockIdx.x / 16) * 4096ll) + ((256ll * ((int)blockIdx.x & 15)) + (int)threadIdx.x))] = cinn_nvgpu_sin_fp16(var[((((int)blockIdx.x / 16ll) * 4096ll) + ((256ll * ((int)blockIdx.x & 15)) + (int)threadIdx.x))]);
    };
  };
}

After

__global__
void __launch_bounds__(256) fn_sin_0__COND__FPA__FPA__FPA_4194304llGE1ll_BPA_AND_FPA_4194304llLE2147483647ll_BPA__BPA_AND_FPA__FPA_1GE1ll_BPA_AND_FPA_1LE1ll_BPA__BPA__BPA___kernel(const float16* __restrict__ var, float16* __restrict__ var_0)
{
  if (((int)blockIdx.x < 4096)) {
    if (((int)threadIdx.x < 256)) {
      var_0[((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x))] = cinn_nvgpu_sin_fp16(var[((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x))]);
      var_0[(1ll + ((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x)))] = cinn_nvgpu_sin_fp16(var[(1ll + ((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x)))]);
      var_0[(2ll + ((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x)))] = cinn_nvgpu_sin_fp16(var[(2ll + ((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x)))]);
      var_0[(3ll + ((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x)))] = cinn_nvgpu_sin_fp16(var[(3ll + ((1024ll * (int)blockIdx.x) + (4ll * (int)threadIdx.x)))]);
    };
  };
}

[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.