webgpu: support LRNGrad kernel #7196
Conversation
| } | ||
| else if (k >= depthBegin && k < depthEnd){ | ||
| var dyi = -2.0 * uniforms.alpha * uniforms.beta | ||
| * getInputImage(b ,r ,c, k) * getOutputImage(b, r, c, d) / norm; |
There was a problem hiding this comment.
Nit
| * getInputImage(b ,r ,c, k) * getOutputImage(b, r, c, d) / norm; | |
| * getInputImage(b, r, c, k) * getOutputImage(b, r, c, d) / norm; |
You may also fix the WebGL code.
| dispatch: [number, number, number]; | ||
| variableNames = ['inputImage', 'outputImage', 'dy']; | ||
| uniforms = | ||
| 'depth : i32, depthRadius : i32, bias : f32, alpha : f32, beta : f32,'; |
There was a problem hiding this comment.
It seems that depth is not necessary as a uniform. You can use let MAX_DEPTH_END = uniforms.outShape[3]; in shader.
| let depthEnd = min(uniforms.depth, d + uniforms.depthRadius + 1); | ||
|
|
||
| let MIN_DEPTH_BEGIN = 0; | ||
| let MAX_DEPTH_END = uniforms.depth; |
There was a problem hiding this comment.
Move L54-L55 out of for (var d = 0; d < uniforms.depth; d++). It seems that you can also use MAX_DEPTH_END for the outermost for.
qjia7
left a comment
qjia7
left a comment
There was a problem hiding this comment.
Two more comments. Sorry missed them in last review.
| continue; | ||
| } | ||
| else if (k >= depthBegin && k < depthEnd) { | ||
| norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k); |
There was a problem hiding this comment.
It seems getInputImage(b, r, c, k) is called twice. Is it better to cache it like below:
let inputValue = getInputImage(b, r, c, k);
norm += inputValue * inputValue ;
There was a problem hiding this comment.
It may cost one more register here, let compiler optimizes the code, Is it OK?
| } | ||
| else if (k >= depthBegin && k < depthEnd){ | ||
| var dyi = -2.0 * uniforms.alpha * uniforms.beta | ||
| * getInputImage(b, r, c, k) * getOutputImage(b, r, c, d) / norm; |
There was a problem hiding this comment.
getOutputImage(b, r, c, d) should be put out of for(var k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; k++) since the arguments are never changed.
There was a problem hiding this comment.
Because getOutputImage(b, r, c, d) is in if-else branch, other branches may does not need to access memory, it is not necessary to put the IO access out of for-loop.
3 participants
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is