fix FindChannelAbsMaxFunctor cuda kernel bug (#58921)

yghstill · web-flow · commit dc8f0b462810 · 2023-11-14T10:24:26.000+08:00
diff --git a/paddle/fluid/operators/fake_quantize_op.cu.h b/paddle/fluid/operators/fake_quantize_op.cu.h
@@ -201,7 +201,7 @@ struct FindChannelAbsMaxFunctor<phi::GPUContext, T> {
         FindChannelAbsMaxKernelQuantAxis1<T>
             <<<grid, block, block * sizeof(T), ctx.stream()>>>(
                 in_data, num, cin, cout, out_abs_max);
-        in_data += num / cin;
+        in_data += cout * max_threads;
       }
 
       int block = cin % max_threads;

Original file line number	Diff line number	Diff line change
`@@ -201,7 +201,7 @@ struct FindChannelAbsMaxFunctor<phi::GPUContext, T> {`
`201`	`201`	`FindChannelAbsMaxKernelQuantAxis1<T>`
`202`	`202`	`<<<grid, block, block * sizeof(T), ctx.stream()>>>(`
`203`	`203`	`in_data, num, cin, cout, out_abs_max);`
`204`		`- in_data += num / cin;`
	`204`	`+ in_data += cout * max_threads;`
`205`	`205`	`}`
`206`	`206`
`207`	`207`	`int block = cin % max_threads;`