@@ -913,8 +913,8 @@ inline static void ggml_vec_reglu_f32 (const int n, float * y, const float * x,
913913
914914inline static void ggml_vec_reglu_f16 (const int n , ggml_fp16_t * y , const ggml_fp16_t * x , const ggml_fp16_t * g ) {
915915 for (int i = 0 ; i < n ; ++ i ) {
916- float v = GGML_FP16_TO_FP32 (x [i ]);
917- y [i ] = GGML_FP32_TO_FP16 ((v > 0.f ) ? v * GGML_FP16_TO_FP32 (g [i ]) : 0.f );
916+ float v = GGML_CPU_FP16_TO_FP32 (x [i ]);
917+ y [i ] = GGML_CPU_FP32_TO_FP16 ((v > 0.f ) ? v * GGML_CPU_FP16_TO_FP32 (g [i ]) : 0.f );
918918 }
919919}
920920
@@ -927,9 +927,9 @@ inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, c
927927 } else if (x [i ] >= 10.0f ) {
928928 y [i ] = x [i ] * g [i ];
929929 } else {
930- ggml_fp16_t fp16 = GGML_FP32_TO_FP16 (x [i ]);
930+ ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16 (x [i ]);
931931 memcpy (& t , & fp16 , sizeof (uint16_t ));
932- y [i ] = GGML_FP16_TO_FP32 (ggml_table_gelu_f16 [t ]) * g [i ];
932+ y [i ] = GGML_CPU_FP16_TO_FP32 (ggml_table_gelu_f16 [t ]) * g [i ];
933933 }
934934 }
935935}
@@ -944,18 +944,18 @@ inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, c
944944inline static void ggml_vec_geglu_f16 (const int n , ggml_fp16_t * y , const ggml_fp16_t * x , const ggml_fp16_t * g ) {
945945 const uint16_t * i16 = (const uint16_t * ) x ;
946946 for (int i = 0 ; i < n ; ++ i ) {
947- float v = GGML_FP16_TO_FP32 (g [i ]);
948- y [i ] = GGML_FP32_TO_FP16 ( GGML_FP16_TO_FP32 (ggml_table_gelu_f16 [i16 [i ]]) * v );
947+ float v = GGML_CPU_FP16_TO_FP32 (g [i ]);
948+ y [i ] = GGML_CPU_FP32_TO_FP16 ( GGML_CPU_FP16_TO_FP32 (ggml_table_gelu_f16 [i16 [i ]]) * v );
949949 }
950950}
951951
952952void ggml_vec_swiglu_f32 (const int n , float * y , const float * x , const float * g );
953953
954954inline static void ggml_vec_swiglu_f16 (const int n , ggml_fp16_t * y , const ggml_fp16_t * x , const ggml_fp16_t * g ) {
955955 for (int i = 0 ; i < n ; ++ i ) {
956- float v = GGML_FP16_TO_FP32 (x [i ]);
957- float w = GGML_FP16_TO_FP32 (g [i ]);
958- y [i ] = GGML_FP32_TO_FP16 ((v /(1.0f + expf (- v ))) * w );
956+ float v = GGML_CPU_FP16_TO_FP32 (x [i ]);
957+ float w = GGML_CPU_FP16_TO_FP32 (g [i ]);
958+ y [i ] = GGML_CPU_FP32_TO_FP16 ((v /(1.0f + expf (- v ))) * w );
959959 }
960960}
961961
0 commit comments