From 95fa383df2919ce16e05672b5b632aa522422d97 Mon Sep 17 00:00:00 2001
From: donproc <donproc@gmail.com>
Date: Wed, 22 Jul 2020 11:02:40 +0800
Subject: [PATCH] optimize embedding cuda kernel lookup_table_v2,test=develop
 (#25587)

---
 paddle/fluid/operators/lookup_table_v2_op.cu | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/paddle/fluid/operators/lookup_table_v2_op.cu b/paddle/fluid/operators/lookup_table_v2_op.cu
index 5c98eab403..b3b0f8f196 100644
--- a/paddle/fluid/operators/lookup_table_v2_op.cu
+++ b/paddle/fluid/operators/lookup_table_v2_op.cu
@@ -105,17 +105,17 @@ class LookupTableV2CUDAKernel : public framework::OpKernel<T> {
     auto *table = table_t->data<T>();
     auto *output = output_t->mutable_data<T>(context.GetPlace());
 
-    dim3 threads(128, 8);
-    dim3 grids(8, 1);
+    dim3 threads(256, 4);
+    dim3 grids(80, 1);
 
     if (padding_idx == -1)
       LookupTableV2<
-          T, 128, 8, 8,
+          T, 256, 4, 80,
           false><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
           output, table, ids, N, K, D, padding_idx);
     else
       LookupTableV2<
-          T, 128, 8, 8,
+          T, 256, 4, 80,
           true><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
           output, table, ids, N, K, D, padding_idx);
   }
-- 
GitLab