fix test

Tencent · May 8, 2024 · ee36857 · ee36857
1 parent 00b096f
commit ee36857
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 3 deletions.
diff --git a/src/layer/arm/gru_int8.h b/src/layer/arm/gru_int8.h
@@ -447,12 +447,18 @@ static void gru_transform_weight_int8(const Mat& weight_xc, const Mat& weight_xc
             _reciprocal_xc_R0 = vmulq_f32(vrecpsq_f32(_xc_R0, _reciprocal_xc_R0), _reciprocal_xc_R0);
             _reciprocal_xc_U0 = vmulq_f32(vrecpsq_f32(_xc_U0, _reciprocal_xc_U0), _reciprocal_xc_U0);
             _reciprocal_xc_N0 = vmulq_f32(vrecpsq_f32(_xc_N0, _reciprocal_xc_N0), _reciprocal_xc_N0);
+            _reciprocal_xc_R0 = vmulq_f32(vrecpsq_f32(_xc_R0, _reciprocal_xc_R0), _reciprocal_xc_R0);
+            _reciprocal_xc_U0 = vmulq_f32(vrecpsq_f32(_xc_U0, _reciprocal_xc_U0), _reciprocal_xc_U0);
+            _reciprocal_xc_N0 = vmulq_f32(vrecpsq_f32(_xc_N0, _reciprocal_xc_N0), _reciprocal_xc_N0);
             float32x4_t _reciprocal_hc_R0 = vrecpeq_f32(_hc_R0);
             float32x4_t _reciprocal_hc_U0 = vrecpeq_f32(_hc_U0);
             float32x4_t _reciprocal_hc_N0 = vrecpeq_f32(_hc_N0);
             _reciprocal_hc_R0 = vmulq_f32(vrecpsq_f32(_hc_R0, _reciprocal_hc_R0), _reciprocal_hc_R0);
             _reciprocal_hc_U0 = vmulq_f32(vrecpsq_f32(_hc_U0, _reciprocal_hc_U0), _reciprocal_hc_U0);
             _reciprocal_hc_N0 = vmulq_f32(vrecpsq_f32(_hc_N0, _reciprocal_hc_N0), _reciprocal_hc_N0);
+            _reciprocal_hc_R0 = vmulq_f32(vrecpsq_f32(_hc_R0, _reciprocal_hc_R0), _reciprocal_hc_R0);
+            _reciprocal_hc_U0 = vmulq_f32(vrecpsq_f32(_hc_U0, _reciprocal_hc_U0), _reciprocal_hc_U0);
+            _reciprocal_hc_N0 = vmulq_f32(vrecpsq_f32(_hc_N0, _reciprocal_hc_N0), _reciprocal_hc_N0);
 #endif
 
             vst1q_f32(descales_ptr, _reciprocal_xc_R0);

diff --git a/src/layer/arm/neon_mathfun.h b/src/layer/arm/neon_mathfun.h
@@ -276,7 +276,7 @@ static inline float32x4_t div_ps(float32x4_t a, float32x4_t b)
 #else
     float32x4_t reciprocal = vrecpeq_f32(b);
     reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
-    // reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
+    reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
     return vmulq_f32(a, reciprocal);
 #endif
 }

diff --git a/src/layer/arm/rnn_int8.h b/src/layer/arm/rnn_int8.h
@@ -193,8 +193,10 @@ static void rnn_transform_weight_int8(const Mat& weight_xc, const Mat& weight_xc
 #else
             float32x4_t _reciprocal_xc = vrecpeq_f32(_xc);
             _reciprocal_xc = vmulq_f32(vrecpsq_f32(_xc, _reciprocal_xc), _reciprocal_xc);
+            _reciprocal_xc = vmulq_f32(vrecpsq_f32(_xc, _reciprocal_xc), _reciprocal_xc);
             float32x4_t _reciprocal_hc = vrecpeq_f32(_hc);
             _reciprocal_hc = vmulq_f32(vrecpsq_f32(_hc, _reciprocal_hc), _reciprocal_hc);
+            _reciprocal_hc = vmulq_f32(vrecpsq_f32(_hc, _reciprocal_hc), _reciprocal_hc);
 #endif
 
             vst1q_f32(descales_ptr, _reciprocal_xc);

diff --git a/tests/test_rnn.cpp b/tests/test_rnn.cpp
@@ -295,7 +295,7 @@ int test_rnn_int8_with_hidden(const ncnn::Mat& a, int outch, int direction)
     weights[4] = RandomMat(outch * num_directions, 100.f, 200.f);
 
     // initial hidden state
-    ncnn::Mat hidden = RandomMat(outch, num_directions);
+    ncnn::Mat hidden = RandomMat(outch, num_directions, -1.f, 1.f);
 
     std::vector<ncnn::Mat> as(2);
     as[0] = a;
@@ -329,7 +329,7 @@ int test_rnn_int8_with_hidden_input(const ncnn::Mat& a, int outch, int direction
     weights[4] = RandomMat(outch * num_directions, 100.f, 200.f);
 
     // initial hidden state
-    ncnn::Mat hidden = RandomMat(outch, num_directions);
+    ncnn::Mat hidden = RandomMat(outch, num_directions, -1.f, 1.f);
 
     std::vector<ncnn::Mat> as(2);
     as[0] = a;