[ util ] Implement exp_i function

- Add exponential inplace function **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
nnstreamer · Feb 20, 2024 · 8c72f55 · 8c72f55
1 parent cc7da98
commit 8c72f55
Show file tree

Hide file tree

Showing 2 changed files with 23 additions and 12 deletions.
diff --git a/nntrainer/utils/util_simd_neon.cpp b/nntrainer/utils/util_simd_neon.cpp
@@ -100,6 +100,17 @@ void softmax(const unsigned int N, float *X, float *Y) {
   }
 }
 
+void exp_i(const unsigned int N, float *X) {
+  unsigned int i = 0;
+  for (; N - i >= VL_FP32; i += VL_FP32) {
+    vst1q_f32(&X[i], exp_ps(vld1q_f32(&X[i])));
+  }
+  while (i < N) {
+    X[i] = std::exp(X[i]);
+    ++i;
+  }
+}
+
 #ifdef ENABLE_FP16
 void compute_rotary_embedding_value(unsigned int dim, unsigned int half_,
                                     unsigned int w, __fp16 *in, __fp16 *out,

diff --git a/nntrainer/utils/util_simd_neon.h b/nntrainer/utils/util_simd_neon.h
@@ -39,19 +39,16 @@ void calc_trigonometric_vals_dup(unsigned int N_half, float *angle, float *cos_,
  * @param Y float * for Vector Y
  * @param Z float * for Vector Z
  */
-<<<<<<< HEAD
 void swish(const unsigned int N, float *X, float *Y, float *Z);
-=======
-void swish_neon(const unsigned int N, float *X, float *Y, float *Z);
 
 /**
  * @brief returns maximum value of the vector X
- * 
+ *
  * @param N number of elements in X
  * @param X float * for Vector X
  * @return float maximum value of vector X
  */
-float max(const unsigned int N, float* X);
+float max(const unsigned int N, float *X);
 
 /**
  * @brief soft max function with neon y_i = exp(x_i) / sum( exp(x_i) )
@@ -61,7 +58,14 @@ float max(const unsigned int N, float* X);
  * @param Y  float * for Vector Y
  */
 void softmax(const unsigned int N, float *X, float *Y);
->>>>>>> [ util ] Implement softmax function in util_simd
+
+/**
+ * @brief exponential inplace function
+ *
+ * @param N number of elements in X
+ * @param X float * for Vector X
+ */
+void exp_i(const unsigned int N, float *X);
 #ifdef ENABLE_FP16
 /**
  * @brief Accelerating function for rotary embedding layer forwarding
@@ -85,19 +89,16 @@ void compute_rotary_embedding_value(unsigned int dim, unsigned int half_,
  * @param Y __fp16 * for Vector Y
  * @param Z __fp16 * for Vector Z
  */
-<<<<<<< HEAD
 void swish(const unsigned int N, __fp16 *X, __fp16 *Y, __fp16 *Z);
-=======
-void swish_neon(const unsigned int N, __fp16 *X, __fp16 *Y, __fp16 *Z);
 
 /**
  * @brief returns maximum value of the vector X
- * 
+ *
  * @param N number of elements in X
  * @param X __fp16 * for Vector X
  * @return __fp16 maximum value of vector X
  */
-__fp16 max(const unsigned int N, __fp16* X);
+__fp16 max(const unsigned int N, __fp16 *X);
 
 /**
  * @brief soft max function with neon y_i = exp(x_i) / sum( exp(x_i) )
@@ -109,7 +110,6 @@ __fp16 max(const unsigned int N, __fp16* X);
  * @param Y  __fp16 * for Vector Y
  */
 void softmax(const unsigned int N, __fp16 *X, __fp16 *Y);
->>>>>>> [ util ] Implement softmax function in util_simd
 #endif
 
 } // namespace nntrainer::neon