Searched refs:vecMac3 (Results 1 – 4 of 4) sorted by relevance
/cmsis-dsp-latest/dsppp/Include/dsppp/Helium/ |
D | matrix_multiply.hpp | 126 ACC vecMac0, vecMac1, vecMac2, vecMac3; 148 vecMac3 = vector_traits<T>::temp_acc_zero(); 162 vecMac3 = inner::vmacc(vecMac3, vecInB, *pInA3++); 178 inner::vstore1<1>(pOut3, vecMac3); 201 vecMac3 = vector_traits<T>::temp_acc_zero(); 215 vecMac3 = inner::vmacc(vecMac3, vecInB, *pInA3++); 228 inner::vstore1_z<1>(pOut3, vecMac3, colBLeft,p0);
|
D | matrix_multiply_f32.hpp | 210 ACC vecMac0, vecMac1, vecMac2, vecMac3; 228 vecMac3 = inner::vmul(vecInB, *pInA3++); 238 vecMac3 = inner::vmacc(vecMac3, vecInB, *pInA3++); 248 vecMac3 = inner::vmacc(vecMac3, vecInB, *pInA3++); 258 vecMac3 = inner::vmacc(vecMac3, vecInB, *pInA3++); 266 inner::vstore1<1>(pOut, vecMac3);
|
/cmsis-dsp-latest/Source/MatrixFunctions/ |
D | arm_mat_mult_f32.c | 214 f32x4_t vecMac0, vecMac1, vecMac2, vecMac3; in arm_mat_mult_f32_4x4_mve() local 232 vecMac3 = vmulq(vecInB, *pInA3++); in arm_mat_mult_f32_4x4_mve() 242 vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++); in arm_mat_mult_f32_4x4_mve() 252 vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++); in arm_mat_mult_f32_4x4_mve() 262 vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++); in arm_mat_mult_f32_4x4_mve() 270 vst1q(pOut, vecMac3); in arm_mat_mult_f32_4x4_mve() 335 f32x4_t vecMac0, vecMac1, vecMac2, vecMac3; in arm_mat_mult_f32() local 357 vecMac3 = vdupq_n_f32(0.0f); in arm_mat_mult_f32() 371 vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++); in arm_mat_mult_f32() 387 vst1q(pOut3, vecMac3); in arm_mat_mult_f32() [all …]
|
D | arm_mat_mult_f16.c | 428 f16x8_t vecMac0, vecMac1, vecMac2, vecMac3; in arm_mat_mult_f16() local 450 vecMac3 = vdupq_n_f16(0.0f16); in arm_mat_mult_f16() 464 vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++); in arm_mat_mult_f16() 477 vst1q(pOut3, vecMac3); pOut3 += 8; in arm_mat_mult_f16() 497 vecMac3 = vdupq_n_f16(0.0f16); in arm_mat_mult_f16() 511 vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++); in arm_mat_mult_f16() 524 vstrhq_p_f16(pOut3, vecMac3, p0); in arm_mat_mult_f16()
|