diff --git a/sadl/layer_add.h b/sadl/layer_add.h
index 34dffda4db5ee626dc87881147119d07926d3193..762cd6fa4ccc60504f1de89a176568acf4c445c4 100644
--- a/sadl/layer_add.h
+++ b/sadl/layer_add.h
@@ -102,29 +102,6 @@ template<typename T> bool Add<T>::apply(std::vector<Tensor<T> *> &in)
if (in[1]->size() == 1)
{ // ie in[0]->dims().size() == 1? happen if in[1] is a Const
const Tensor<T> &B = *in[1];
-#if __AVX2__
- if constexpr (std::is_same_v<T, int16_t>)
- {
- if (m_out.size() % 16 == 0)
- {
- const __m256i value = _mm256_set1_epi16(B[0]);
- const __m256i max = _mm256_set1_epi16(32767);
- const __m256i min = _mm256_set1_epi16(-32768);
- T *a_ptr = m_out.data();
- for (int64_t i = 0; i < m_out.size(); i += 16, a_ptr+=16)
- {
- __m256i x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(a_ptr));
- __m256i result = _mm256_adds_epi16(x, value);
- result = _mm256_srai_epi16(result, shift);
-#if SATURATE_RESULT
- result = _mm256_min_epi16(_mm256_max_epi16(result, min), max);
-#endif
- _mm256_store_si256((__m256i *) a_ptr, result);
- }
- return true;
- }
- }
-#else
const T value = B[0];
#if DEBUG_MODEL_ANALYZE
std::cout << "\n[ANALYZE] add (in):\t" << m_out.size() << std::endl;
@@ -138,7 +115,6 @@ template<typename T> bool Add<T>::apply(std::vector<Tensor<T> *> &in)
SATURATE(z);
x = static_cast<T>(z);
}
-#endif
}
else if (in[0]->dims().size() == 2)
{
@@ -231,28 +207,6 @@ template<typename T> bool Add<T>::apply(std::vector<Tensor<T> *> &in)
if (in[1]->size() == 1)
{ // for constant
const Tensor<T> &B = *in[1];
-#if __AVX2__
- if constexpr(std::is_same_v<T, int16_t>)
- {
- if (m_out.size() % 16 == 0)
- {
- const __m256i value = _mm256_srai_epi16(_mm256_set1_epi16(B[0]), shift);
- const __m256i max = _mm256_set1_epi16(32767);
- const __m256i min = _mm256_set1_epi16(-32768);
- T *a_ptr = m_out.data();
- for (int64_t i = 0; i < m_out.size(); i += 16, a_ptr+=16)
- {
- __m256i x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(a_ptr));
- __m256i result = _mm256_adds_epi16(x, value);
-#if SATURATE_RESULT
- result = _mm256_min_epi16(_mm256_max_epi16(result, min), max);
-#endif
- _mm256_store_si256((__m256i *) a_ptr, result);
- }
- return true;
- }
- }
-#else
T valt = B[0];
ComputationType<T>::quantize(valt, shift);
const T value = valt;
@@ -267,7 +221,6 @@ template<typename T> bool Add<T>::apply(std::vector<Tensor<T> *> &in)
SATURATE(z);
x = static_cast<T>(z);
}
-#endif
}
else if (in[0]->dims().size() == 2)
{
diff --git a/sadl/layer_biasadd.h b/sadl/layer_biasadd.h
index 76075bbccd890d2bfb8d4f9d3ba369cf206b94f3..5fa733ea2218e8bf38815059747a94cfa64a5021 100644
--- a/sadl/layer_biasadd.h
+++ b/sadl/layer_biasadd.h
@@ -101,39 +101,13 @@ template<typename T> bool BiasAdd<T>::apply(std::vector<Tensor<T> *> &in)
}
else if (in[0]->dims().size() == 2)
{
+ const Tensor<T> &B = *in[1];
+ assert(B.dims().size() == 1 || (B.dims().size() == 2 && B.dims()[0] == 1));
const int N = in[0]->dims()[0];
const int H = in[0]->dims()[1];
#if DEBUG_MODEL_ANALYZE
std::cout << "\n[ANALYZE] add (in):\t" << H << std::endl;
#endif
-#if __AVX2__
- if constexpr(std::is_same_v<T, int16_t>)
- {
- if (H % 16 == 0)
- {
- const __m256i max = _mm256_set1_epi16(32767);
- const __m256i min = _mm256_set1_epi16(-32768);
- for (int n = 0; n < N; ++n) {
- T *a_ptr = &m_out[n];
- T *b_ptr = in[1]->data();
- for (int i = 0; i < H; i += 16, a_ptr+=16, b_ptr+=16)
- {
- __m256i a = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(a_ptr));
- __m256i b = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(b_ptr));
- __m256i result = _mm256_adds_epi16(a, b);
- result = _mm256_slli_epi16(result, shift);
-#if SATURATE_RESULT
- result = _mm256_min_epi16(_mm256_max_epi16(result, min), max);
-#endif
- _mm256_store_si256((__m256i *) a_ptr, result);
- }
- }
- return true;
- }
- }
-#else
- const Tensor<T> &B = *in[1];
- assert(B.dims().size() == 1 || (B.dims().size() == 2 && B.dims()[0] == 1));
for (int n = 0; n < N; ++n)
for (int i = 0; i < H; ++i)
{
@@ -144,7 +118,6 @@ template<typename T> bool BiasAdd<T>::apply(std::vector<Tensor<T> *> &in)
SATURATE(z);
m_out(n, i) = static_cast<T>(z);
}
-#endif
}
else if (in[0]->dims().size() == 3)
{
@@ -232,39 +205,13 @@ template<typename T> bool BiasAdd<T>::apply(std::vector<Tensor<T> *> &in)
}
else if (in[0]->dims().size() == 2)
{
+ const Tensor<T> &B = *in[1];
+ assert(B.dims().size() == 1 || (B.dims().size() == 2 && B.dims()[0] == 1));
const int N = in[0]->dims()[0];
const int H = in[0]->dims()[1];
#if DEBUG_MODEL_ANALYZE
std::cout << "\n[ANALYZE] add (in):\t" << H << std::endl;
#endif
-#if __AVX2__
- if constexpr(std::is_same_v<T, int16_t>)
- {
- if (H % 16 == 0)
- {
- const __m256i max = _mm256_set1_epi16(32767);
- const __m256i min = _mm256_set1_epi16(-32768);
- for (int n = 0; n < N; ++n) {
- T *a_ptr = &m_out[n];
- T *b_ptr = in[1]->data();
- for (int i = 0; i < H; i += 16, a_ptr+=16, b_ptr+=16)
- {
- __m256i a = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(a_ptr));
- __m256i b = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(b_ptr));
- __m256i result = _mm256_adds_epi16(a, b);
- result = _mm256_srai_epi16(result, shift);
-#if SATURATE_RESULT
- result = _mm256_min_epi16(_mm256_max_epi16(result, min), max);
-#endif
- _mm256_store_si256((__m256i *) a_ptr, result);
- }
- }
- return true;
- }
- }
-#else
- const Tensor<T> &B = *in[1];
- assert(B.dims().size() == 1 || (B.dims().size() == 2 && B.dims()[0] == 1));
for (int n = 0; n < N; ++n)
for (int i = 0; i < H; ++i)
{
@@ -275,7 +222,6 @@ template<typename T> bool BiasAdd<T>::apply(std::vector<Tensor<T> *> &in)
SATURATE(z);
m_out(n, i) = static_cast<T>(z);
}
-#endif
}
else if (in[0]->dims().size() == 3)
{