webgpu: support EluGrad kernel #7197
Conversation
| const COMPLEX_MULTIPLY_IMAG = 'return areal * bimag + aimag * breal;'; | ||
| const DIV = 'return a / b;'; | ||
| const ELU_DER = ` | ||
| let bGTEZero = f32(b > 0.); |
There was a problem hiding this comment.
| let bGTEZero = f32(b > 0.); | |
| let bGTEZero = f32(b >= 0.); |
| const DIV = 'return a / b;'; | ||
| const ELU_DER = ` | ||
| let bGTEZero = f32(b > 0.); | ||
| return (bGTEZero * a) + (1.0 - bGTEZero) * (a * (b + 1.0)); |
| `; | ||
| const ELU_DER_VEC4 = ` | ||
| let bGTEZero = vec4<f32>(b >= vec4<f32>(0.)); | ||
| return (bGTEZero * a) + (vec4<f32>(1.0) - bGTEZero) * (a * (b + vec4<f32>(1.0))); |
| const COMPLEX_MULTIPLY_REAL = 'return areal * breal - aimag * bimag;'; | ||
| const COMPLEX_MULTIPLY_IMAG = 'return areal * bimag + aimag * breal;'; | ||
| const DIV = 'return a / b;'; | ||
| const ELU_DER = 'return select(a * (b + 1.0), a, b >= 0.);'; |
There was a problem hiding this comment.
I see webgl is using b >= 1.0 https://github.com/tensorflow/tfjs/blob/master/tfjs-backend-webgl/src/kernels/EluGrad.ts#L24 . However, you are using b >= 0.0. Which one is correct?
There was a problem hiding this comment.
On webgl backend, the non PACKED and PACKED versions are different,
I think the PACKED version is correct, but is not 100% sure.
There was a problem hiding this comment.
According to http://bytedeco.org/javacpp-presets/tensorflow/apidocs/org/bytedeco/tensorflow/EluGrad.html, b >= 0. is correct.
Please also fix the WebGL unpacked version, and I think we also need some test cases to cover.
There was a problem hiding this comment.
Thanks! Please hold this a bit until #7251 is reviewed.
qjia7
left a comment
qjia7
left a comment
There was a problem hiding this comment.
LGTM with one nit. Thanks!
| @@ -0,0 +1,38 @@ | |||
| /** | |||
| * @license | |||
| * Copyright 2022 Google LLC. | |||
3 participants
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is