[ neon/trivial ] Compare float scaling factors more precisely

- for zero-comparison, use std::fpclassify
- for 1.0-comparison, use std::numeric_limits and epsilon

**Self evaluation:**
1. Build test:     [X]Passed [ ]Failed [ ]Skipped
2. Run test:     [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: skykongkong8 <ss.kong@samsung.com>

diff --git a/nntrainer/tensor/blas_neon.cpp b/nntrainer/tensor/blas_neon.cpp
index fb9b14386f4ebe891b24c143470fad4111ca302a..4a200311954f6258527b86724d2a38f110ceb923 100644 (file)
--- a/nntrainer/tensor/blas_neon.cpp
+++ b/nntrainer/tensor/blas_neon.cpp
@@ -1597,7 +1597,7 @@ void hgemm(const __fp16 *A, const __fp16 *B, __fp16 *C, uint32_t M, uint32_t N,
   unsigned int size = M * N;
   unsigned int size8 = (size >> 3) << 3;
   unsigned int size4 = (size >> 2) << 2;
-  if (beta != 0.F) {
+  if (std::fpclassify(beta) != FP_ZERO) {
     for (; idx < size8; idx += 8) {
       float16x8_t c =
         vmulq_n_f16(vld1q_f16(&C[idx]), static_cast<__fp16>(beta));

diff --git a/nntrainer/tensor/hgemm/hgemm.cpp b/nntrainer/tensor/hgemm/hgemm.cpp
index dd1c173a99e45e1fed4f5a175e50ae4ed3216137..f7788335e91b64718f6ce6a960616208a3eda32b 100644 (file)
--- a/nntrainer/tensor/hgemm/hgemm.cpp
+++ b/nntrainer/tensor/hgemm/hgemm.cpp
@@ -12,6 +12,7 @@
  *
  */
 
+#include <cmath>
 #include <hgemm.h>
 #include <hgemm_kernel_1x4.h>
 #include <hgemm_kernel_1x8.h>
@@ -21,6 +22,7 @@
 #include <hgemm_kernel_8x8.h>
 #include <hgemm_kernel_pack.h>
 #include <hgemm_util.h>
+#include <limits>
 
 #define HGEMM_KERNEL_1x4 hgemm_kernel_1x4
 #define HGEMM_KERNEL_4x4 hgemm_kernel_4x4
@@ -31,7 +33,8 @@
 
 void hgemm_noTrans(const __fp16 *A, const __fp16 *B, float *C32, unsigned int M,
                    unsigned int N, unsigned int K, float alpha, float beta) {
-  if (alpha == 1.F) {
+  const float eps = std::numeric_limits<float>::epsilon();
+  if (std::abs(alpha - 1.F) < eps) {
     // used bitwise operator instead of modulo for performance
     // e.g (M % 8) is same as (M & 0x7) which will extract last 3 bits of M
     if ((M & 0x7) == 0 && (N & 0xF) == 0 && (K & 0x7) == 0) {