|
|
@ -116,9 +116,11 @@ inline bool vec_check(size_t len) {
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
namespace internal {
|
|
|
|
namespace internal {
|
|
|
|
|
|
|
|
#ifdef __SSE3__
|
|
|
|
void addToImpl(float* a, const float* b, size_t len);
|
|
|
|
void addToImpl(float* a, const float* b, size_t len);
|
|
|
|
void batchAddToImpl(float* a, const float* b[], int batch, size_t len);
|
|
|
|
void batchAddToImpl(float* a, const float* b[], int batch, size_t len);
|
|
|
|
void colMaxImpl(float* result, const float* data, int dim, int numSamples);
|
|
|
|
void colMaxImpl(float* result, const float* data, int dim, int numSamples);
|
|
|
|
|
|
|
|
#endif
|
|
|
|
#ifdef __AVX__
|
|
|
|
#ifdef __AVX__
|
|
|
|
void decayL1AvxImpl(float* dst, float* src, float lambda, size_t len);
|
|
|
|
void decayL1AvxImpl(float* dst, float* src, float lambda, size_t len);
|
|
|
|
void decayL1AvxImpl(
|
|
|
|
void decayL1AvxImpl(
|
|
|
|