optimize perfermance of multiple-dimension reduce (#33761)

2dde0eb0 · Zhang Zheng · GitHub · 4d259b91 · 2dde0eb0 · 2dde0eb0
Showing with 257 addition and 151 deletion

paddle/fluid/operators/reduce_ops/reduce_op.cu.h paddle/fluid/operators/reduce_ops/reduce_op.cu.h +256 -150

paddle/fluid/platform/fast_divmod.h paddle/fluid/platform/fast_divmod.h +1 -1

未找到文件。
--- a/paddle/fluid/operators/reduce_ops/reduce_op.cu.h
+++ b/paddle/fluid/operators/reduce_ops/reduce_op.cu.h
--- a/paddle/fluid/platform/fast_divmod.h
+++ b/paddle/fluid/platform/fast_divmod.h
@@ -54,7 +54,7 @@ struct FastDivMod {
    return (t + n) >> shift_val;
  }

-  __device__ __forceinline__ DivModT Divmod(uint32_t n) {
+  __device__ __forceinline__ DivModT Divmod(uint32_t n) const {
    uint32_t q = Div(n);
    DivModT result = {q, n - q * divisor};
    return result;