/cmsis-dsp-latest/Source/MatrixFunctions/ |
D | arm_mat_cmplx_mult_q31.c | 79 q31x4_t vecB, vecA; in arm_mat_cmplx_mult_q31_2x2_mve() local 91 vecA = vldrwq_s32(pInA0); in arm_mat_cmplx_mult_q31_2x2_mve() 92 acc0 = vmlsldavq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve() 93 acc1 = vmlaldavxq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve() 95 vecA = vldrwq_s32(pInA1); in arm_mat_cmplx_mult_q31_2x2_mve() 96 acc2 = vmlsldavq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve() 97 acc3 = vmlaldavxq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve() 109 vecA = vldrwq_s32(pInA0); in arm_mat_cmplx_mult_q31_2x2_mve() 110 acc0 = vmlsldavq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve() 111 acc1 = vmlaldavxq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve() [all …]
|
D | arm_mat_cmplx_mult_f32.c | 83 f32x4_t vecB, vecA; in arm_mat_cmplx_mult_f32_2x2_mve() local 95 vecA = vldrwq_f32(pInA0); in arm_mat_cmplx_mult_f32_2x2_mve() 96 acc0 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve() 97 acc0 = vcmlaq_rot90(acc0, vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve() 99 vecA = vldrwq_f32(pInA1); in arm_mat_cmplx_mult_f32_2x2_mve() 100 acc1 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve() 101 acc1 = vcmlaq_rot90(acc1, vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve() 116 vecA = vldrwq_f32(pInA0); in arm_mat_cmplx_mult_f32_2x2_mve() 117 acc0 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve() 118 acc0 = vcmlaq_rot90(acc0, vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve() [all …]
|
D | arm_mat_mult_q31.c | 139 q31x4_t vecB, vecA; in arm_mat_mult_q31_3x3_mve() local 150 vecA = vldrwq_s32(pInA0); in arm_mat_mult_q31_3x3_mve() 151 acc0 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_q31_3x3_mve() 152 vecA = vldrwq_s32(pInA1); in arm_mat_mult_q31_3x3_mve() 153 acc1 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_q31_3x3_mve() 154 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_3x3_mve() 155 acc2 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_q31_3x3_mve() 171 vecA = vldrwq_s32(pInA0); in arm_mat_mult_q31_3x3_mve() 172 acc0 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_q31_3x3_mve() 173 vecA = vldrwq_s32(pInA1); in arm_mat_mult_q31_3x3_mve() [all …]
|
D | arm_mat_cmplx_mult_f16.c | 251 f16x8_t vecB, vecA; in arm_mat_cmplx_mult_f16_4x4_mve() local 264 vecA = vldrhq_f16(pInA0); in arm_mat_cmplx_mult_f16_4x4_mve() 265 acc0 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f16_4x4_mve() 266 acc0 = vcmlaq_rot90(acc0, vecA, vecB); in arm_mat_cmplx_mult_f16_4x4_mve() 268 vecA = vldrhq_f16(pInA1); in arm_mat_cmplx_mult_f16_4x4_mve() 269 acc1 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f16_4x4_mve() 270 acc1 = vcmlaq_rot90(acc1, vecA, vecB); in arm_mat_cmplx_mult_f16_4x4_mve() 272 vecA = vldrhq_f16(pInA2); in arm_mat_cmplx_mult_f16_4x4_mve() 273 acc2 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f16_4x4_mve() 274 acc2 = vcmlaq_rot90(acc2, vecA, vecB); in arm_mat_cmplx_mult_f16_4x4_mve() [all …]
|
D | arm_mat_mult_opt_q31.c | 143 q31x4_t vecB, vecA; in arm_mat_mult_opt_q31_3x3_mve() local 154 vecA = vldrwq_s32(pInA0); in arm_mat_mult_opt_q31_3x3_mve() 155 acc0 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_opt_q31_3x3_mve() 156 vecA = vldrwq_s32(pInA1); in arm_mat_mult_opt_q31_3x3_mve() 157 acc1 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_opt_q31_3x3_mve() 158 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_3x3_mve() 159 acc2 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_opt_q31_3x3_mve() 175 vecA = vldrwq_s32(pInA0); in arm_mat_mult_opt_q31_3x3_mve() 176 acc0 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_opt_q31_3x3_mve() 177 vecA = vldrwq_s32(pInA1); in arm_mat_mult_opt_q31_3x3_mve() [all …]
|
D | arm_mat_vec_mult_f32.c | 113 f32x4_t vecA; in arm_mat_vec_mult_f32() local 117 vecA = vld1q(pSrcA0Vec); in arm_mat_vec_mult_f32() 119 acc0 = vfmaq(acc0, vecIn, vecA); in arm_mat_vec_mult_f32() 120 vecA = vld1q(pSrcA1Vec); in arm_mat_vec_mult_f32() 122 acc1 = vfmaq(acc1, vecIn, vecA); in arm_mat_vec_mult_f32() 123 vecA = vld1q(pSrcA2Vec); in arm_mat_vec_mult_f32() 125 acc2 = vfmaq(acc2, vecIn, vecA); in arm_mat_vec_mult_f32() 126 vecA = vld1q(pSrcA3Vec); in arm_mat_vec_mult_f32() 128 acc3 = vfmaq(acc3, vecIn, vecA); in arm_mat_vec_mult_f32() 140 f32x4_t vecA; in arm_mat_vec_mult_f32() local [all …]
|
D | arm_mat_vec_mult_f16.c | 109 f16x8_t vecA; in arm_mat_vec_mult_f16() local 113 vecA = vld1q(pSrcA0Vec); in arm_mat_vec_mult_f16() 115 acc0 = vfmaq(acc0, vecIn, vecA); in arm_mat_vec_mult_f16() 116 vecA = vld1q(pSrcA1Vec); in arm_mat_vec_mult_f16() 118 acc1 = vfmaq(acc1, vecIn, vecA); in arm_mat_vec_mult_f16() 119 vecA = vld1q(pSrcA2Vec); in arm_mat_vec_mult_f16() 121 acc2 = vfmaq(acc2, vecIn, vecA); in arm_mat_vec_mult_f16() 122 vecA = vld1q(pSrcA3Vec); in arm_mat_vec_mult_f16() 124 acc3 = vfmaq(acc3, vecIn, vecA); in arm_mat_vec_mult_f16() 136 f16x8_t vecA; in arm_mat_vec_mult_f16() local [all …]
|
D | arm_mat_ldlt_f32.c | 159 f32x4_t vecA; in arm_mat_ldlt_f32() local 180 vecA = vldrwq_z_f32(&pA[(w + 0)*n+x],p0); in arm_mat_ldlt_f32() 181 vecA = vfmsq_m(vecA, vecW0, vecX, p0); in arm_mat_ldlt_f32() 182 vstrwq_p(&pA[(w + 0)*n+x], vecA, p0); in arm_mat_ldlt_f32() 184 vecA = vldrwq_z_f32(&pA[(w + 1)*n+x],p0); in arm_mat_ldlt_f32() 185 vecA = vfmsq_m(vecA, vecW1, vecX, p0); in arm_mat_ldlt_f32() 186 vstrwq_p(&pA[(w + 1)*n+x], vecA, p0); in arm_mat_ldlt_f32() 188 vecA = vldrwq_z_f32(&pA[(w + 2)*n+x],p0); in arm_mat_ldlt_f32() 189 vecA = vfmsq_m(vecA, vecW2, vecX, p0); in arm_mat_ldlt_f32() 190 vstrwq_p(&pA[(w + 2)*n+x], vecA, p0); in arm_mat_ldlt_f32() [all …]
|
D | arm_mat_solve_upper_triangular_f32.c | 91 f32x4_t vecA; in arm_mat_solve_upper_triangular_f32() local 98 vecA = vld1q_f32(&pA[i * cols + j]); in arm_mat_solve_upper_triangular_f32() 103 vecA = vfmsq(vecA,vdupq_n_f32(pUT[n*i + k]),vecX); in arm_mat_solve_upper_triangular_f32() 112 vecA = vmulq(vecA,vdupq_n_f32(invUT)); in arm_mat_solve_upper_triangular_f32() 115 vst1q(&pX[i*cols+j],vecA); in arm_mat_solve_upper_triangular_f32() 188 f32x4_t vecA; in arm_mat_solve_upper_triangular_f32() local 195 vecA = vld1q_f32(&pA[i * cols + j]); in arm_mat_solve_upper_triangular_f32() 200 vecA = vfmsq_f32(vecA,vdupq_n_f32(pUT[n*i + k]),vecX); in arm_mat_solve_upper_triangular_f32() 209 vecA = vmulq_f32(vecA,vdupq_n_f32(invUT)); in arm_mat_solve_upper_triangular_f32() 212 vst1q_f32(&pX[i*cols+j],vecA); in arm_mat_solve_upper_triangular_f32()
|
D | arm_mat_solve_lower_triangular_f32.c | 98 f32x4_t vecA; in arm_mat_solve_lower_triangular_f32() local 106 vecA = vld1q_f32(&pA[i * cols + j]); in arm_mat_solve_lower_triangular_f32() 111 vecA = vfmsq(vecA,vdupq_n_f32(pLT[n*i + k]),vecX); in arm_mat_solve_lower_triangular_f32() 120 vecA = vmulq(vecA,vdupq_n_f32(invLT)); in arm_mat_solve_lower_triangular_f32() 121 vst1q(&pX[i*cols+j],vecA); in arm_mat_solve_lower_triangular_f32() 199 f32x4_t vecA; in arm_mat_solve_lower_triangular_f32() local 207 vecA = vld1q_f32(&pA[i * cols + j]); in arm_mat_solve_lower_triangular_f32() 212 vecA = vfmsq_f32(vecA,vdupq_n_f32(pLT[n*i + k]),vecX); in arm_mat_solve_lower_triangular_f32() 221 vecA = vmulq_f32(vecA,vdupq_n_f32(invLT)); in arm_mat_solve_lower_triangular_f32() 222 vst1q_f32(&pX[i*cols+j],vecA); in arm_mat_solve_lower_triangular_f32()
|
D | arm_mat_cmplx_mult_q15.c | 136 q15x8_t vecA, vecB, vecB2; in arm_mat_cmplx_mult_q15() local 170 vecA = vld1q(pSrcAVec); in arm_mat_cmplx_mult_q15() 174 acc0 = vmlsldavaq_s16(acc0, vecA, vecB); in arm_mat_cmplx_mult_q15() 175 acc1 = vmlaldavaxq_s16(acc1, vecA, vecB); in arm_mat_cmplx_mult_q15() 182 acc2 = vmlsldavaq_s16(acc2, vecA, vecB2); in arm_mat_cmplx_mult_q15() 183 acc3 = vmlaldavaxq_s16(acc3, vecA, vecB2); in arm_mat_cmplx_mult_q15() 197 vecA = vldrhq_z_s16(pSrcAVec, p0); in arm_mat_cmplx_mult_q15() 199 acc0 = vmlsldavaq_s16(acc0, vecA, vecB); in arm_mat_cmplx_mult_q15() 200 acc1 = vmlaldavaxq_s16(acc1, vecA, vecB); in arm_mat_cmplx_mult_q15() 208 acc2 = vmlsldavaq_s16(acc2, vecA, vecB2); in arm_mat_cmplx_mult_q15() [all …]
|
/cmsis-dsp-latest/Source/SVMFunctions/ |
D | arm_svm_linear_predict_f16.c | 108 f16x8_t vecA; in arm_svm_linear_predict_f16() local 112 vecA = vld1q(pSrcA0Vec); in arm_svm_linear_predict_f16() 114 acc0 = vfmaq(acc0, vecIn, vecA); in arm_svm_linear_predict_f16() 115 vecA = vld1q(pSrcA1Vec); in arm_svm_linear_predict_f16() 117 acc1 = vfmaq(acc1, vecIn, vecA); in arm_svm_linear_predict_f16() 118 vecA = vld1q(pSrcA2Vec); in arm_svm_linear_predict_f16() 120 acc2 = vfmaq(acc2, vecIn, vecA); in arm_svm_linear_predict_f16() 121 vecA = vld1q(pSrcA3Vec); in arm_svm_linear_predict_f16() 123 acc3 = vfmaq(acc3, vecIn, vecA); in arm_svm_linear_predict_f16() 134 f16x8_t vecA; in arm_svm_linear_predict_f16() local [all …]
|
D | arm_svm_polynomial_predict_f16.c | 134 f16x8_t vecA; in arm_svm_polynomial_predict_f16() local 138 vecA = vld1q(pSrcA0Vec); in arm_svm_polynomial_predict_f16() 140 acc0 = vfmaq(acc0, vecIn, vecA); in arm_svm_polynomial_predict_f16() 141 vecA = vld1q(pSrcA1Vec); in arm_svm_polynomial_predict_f16() 143 acc1 = vfmaq(acc1, vecIn, vecA); in arm_svm_polynomial_predict_f16() 144 vecA = vld1q(pSrcA2Vec); in arm_svm_polynomial_predict_f16() 146 acc2 = vfmaq(acc2, vecIn, vecA); in arm_svm_polynomial_predict_f16() 147 vecA = vld1q(pSrcA3Vec); in arm_svm_polynomial_predict_f16() 149 acc3 = vfmaq(acc3, vecIn, vecA); in arm_svm_polynomial_predict_f16() 160 f16x8_t vecA; in arm_svm_polynomial_predict_f16() local [all …]
|
D | arm_svm_sigmoid_predict_f16.c | 111 f16x8_t vecA; in arm_svm_sigmoid_predict_f16() local 115 vecA = vld1q(pSrcA0Vec); in arm_svm_sigmoid_predict_f16() 117 acc0 = vfmaq(acc0, vecIn, vecA); in arm_svm_sigmoid_predict_f16() 118 vecA = vld1q(pSrcA1Vec); in arm_svm_sigmoid_predict_f16() 120 acc1 = vfmaq(acc1, vecIn, vecA); in arm_svm_sigmoid_predict_f16() 121 vecA = vld1q(pSrcA2Vec); in arm_svm_sigmoid_predict_f16() 123 acc2 = vfmaq(acc2, vecIn, vecA); in arm_svm_sigmoid_predict_f16() 124 vecA = vld1q(pSrcA3Vec); in arm_svm_sigmoid_predict_f16() 126 acc3 = vfmaq(acc3, vecIn, vecA); in arm_svm_sigmoid_predict_f16() 137 f16x8_t vecA; in arm_svm_sigmoid_predict_f16() local [all …]
|
D | arm_svm_rbf_predict_f16.c | 111 f16x8_t vecA; in arm_svm_rbf_predict_f16() local 116 vecA = vld1q(pSrcA0Vec); in arm_svm_rbf_predict_f16() 118 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f16() 120 vecA = vld1q(pSrcA1Vec); in arm_svm_rbf_predict_f16() 122 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f16() 124 vecA = vld1q(pSrcA2Vec); in arm_svm_rbf_predict_f16() 126 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f16() 128 vecA = vld1q(pSrcA3Vec); in arm_svm_rbf_predict_f16() 130 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f16() 142 f16x8_t vecA; in arm_svm_rbf_predict_f16() local [all …]
|
D | arm_svm_linear_predict_f32.c | 105 f32x4_t vecA; in arm_svm_linear_predict_f32() local 109 vecA = vld1q(pSrcA0Vec); in arm_svm_linear_predict_f32() 111 acc0 = vfmaq(acc0, vecIn, vecA); in arm_svm_linear_predict_f32() 112 vecA = vld1q(pSrcA1Vec); in arm_svm_linear_predict_f32() 114 acc1 = vfmaq(acc1, vecIn, vecA); in arm_svm_linear_predict_f32() 115 vecA = vld1q(pSrcA2Vec); in arm_svm_linear_predict_f32() 117 acc2 = vfmaq(acc2, vecIn, vecA); in arm_svm_linear_predict_f32() 118 vecA = vld1q(pSrcA3Vec); in arm_svm_linear_predict_f32() 120 acc3 = vfmaq(acc3, vecIn, vecA); in arm_svm_linear_predict_f32() 131 f32x4_t vecA; in arm_svm_linear_predict_f32() local [all …]
|
D | arm_svm_sigmoid_predict_f32.c | 108 f32x4_t vecA; in arm_svm_sigmoid_predict_f32() local 112 vecA = vld1q(pSrcA0Vec); in arm_svm_sigmoid_predict_f32() 114 acc0 = vfmaq(acc0, vecIn, vecA); in arm_svm_sigmoid_predict_f32() 115 vecA = vld1q(pSrcA1Vec); in arm_svm_sigmoid_predict_f32() 117 acc1 = vfmaq(acc1, vecIn, vecA); in arm_svm_sigmoid_predict_f32() 118 vecA = vld1q(pSrcA2Vec); in arm_svm_sigmoid_predict_f32() 120 acc2 = vfmaq(acc2, vecIn, vecA); in arm_svm_sigmoid_predict_f32() 121 vecA = vld1q(pSrcA3Vec); in arm_svm_sigmoid_predict_f32() 123 acc3 = vfmaq(acc3, vecIn, vecA); in arm_svm_sigmoid_predict_f32() 134 f32x4_t vecA; in arm_svm_sigmoid_predict_f32() local [all …]
|
D | arm_svm_polynomial_predict_f32.c | 111 f32x4_t vecA; in arm_svm_polynomial_predict_f32() local 115 vecA = vld1q(pSrcA0Vec); in arm_svm_polynomial_predict_f32() 117 acc0 = vfmaq(acc0, vecIn, vecA); in arm_svm_polynomial_predict_f32() 118 vecA = vld1q(pSrcA1Vec); in arm_svm_polynomial_predict_f32() 120 acc1 = vfmaq(acc1, vecIn, vecA); in arm_svm_polynomial_predict_f32() 121 vecA = vld1q(pSrcA2Vec); in arm_svm_polynomial_predict_f32() 123 acc2 = vfmaq(acc2, vecIn, vecA); in arm_svm_polynomial_predict_f32() 124 vecA = vld1q(pSrcA3Vec); in arm_svm_polynomial_predict_f32() 126 acc3 = vfmaq(acc3, vecIn, vecA); in arm_svm_polynomial_predict_f32() 137 f32x4_t vecA; in arm_svm_polynomial_predict_f32() local [all …]
|
D | arm_svm_rbf_predict_f32.c | 108 f32x4_t vecA; in arm_svm_rbf_predict_f32() local 113 vecA = vld1q(pSrcA0Vec); in arm_svm_rbf_predict_f32() 115 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f32() 117 vecA = vld1q(pSrcA1Vec); in arm_svm_rbf_predict_f32() 119 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f32() 121 vecA = vld1q(pSrcA2Vec); in arm_svm_rbf_predict_f32() 123 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f32() 125 vecA = vld1q(pSrcA3Vec); in arm_svm_rbf_predict_f32() 127 vecDif = vsubq(vecIn, vecA); in arm_svm_rbf_predict_f32() 139 f32x4_t vecA; in arm_svm_rbf_predict_f32() local [all …]
|
/cmsis-dsp-latest/Source/StatisticsFunctions/ |
D | arm_mse_f32.c | 60 float32x4_t vecA, vecB; in arm_mse_f32() local 70 vecA = vld1q(pSrcA); in arm_mse_f32() 76 vecA = vsubq(vecA, vecB); in arm_mse_f32() 78 vecSum = vfmaq(vecSum, vecA, vecA); in arm_mse_f32() 90 vecA = vld1q(pSrcA); in arm_mse_f32() 93 vecA = vsubq(vecA, vecB); in arm_mse_f32() 94 vecSum = vfmaq_m(vecSum, vecA, vecA, p0); in arm_mse_f32() 114 float32x4_t vecA, vecB; in arm_mse_f32() local 128 vecA = vld1q_f32(pSrcA); in arm_mse_f32() 134 vecA = vsubq_f32(vecA, vecB); in arm_mse_f32() [all …]
|
D | arm_mse_f16.c | 60 float16x8_t vecA, vecB; in arm_mse_f16() local 69 vecA = vld1q(pSrcA); in arm_mse_f16() 75 vecA = vsubq(vecA, vecB); in arm_mse_f16() 77 vecSum = vfmaq(vecSum, vecA, vecA); in arm_mse_f16() 89 vecA = vld1q(pSrcA); in arm_mse_f16() 92 vecA = vsubq(vecA, vecB); in arm_mse_f16() 93 vecSum = vfmaq_m(vecSum, vecA, vecA, p0); in arm_mse_f16()
|
/cmsis-dsp-latest/Source/TransformFunctions/ |
D | arm_cfft_f16.c | 95 f16x8_t vecA, vecB, vecC, vecD; in _arm_radix4_butterfly_f16_mve() local 136 vecA = vldrhq_f16(inA); in _arm_radix4_butterfly_f16_mve() 143 vecSum0 = vecA + vecC; /* vecSum0 = vaddq(vecA, vecC) */ in _arm_radix4_butterfly_f16_mve() 144 vecDiff0 = vecA - vecC; /* vecSum0 = vsubq(vecA, vecC) */ in _arm_radix4_butterfly_f16_mve() 194 vecA = vldrhq_f16(inA); in _arm_radix4_butterfly_f16_mve() 214 vecA = (f16x8_t)vldrwq_gather_base_wb_f32(&vecScGathAddr, 64); in _arm_radix4_butterfly_f16_mve() 220 vecSum0 = vecA + vecC; /* vecSum0 = vaddq(vecA, vecC) */ in _arm_radix4_butterfly_f16_mve() 221 vecDiff0 = vecA - vecC; /* vecSum0 = vsubq(vecA, vecC) */ in _arm_radix4_butterfly_f16_mve() 230 vecA = (f16x8_t)vldrwq_gather_base_wb_f32(&vecScGathAddr, 64); in _arm_radix4_butterfly_f16_mve() 301 f16x8_t vecA, vecB, vecC, vecD; in _arm_radix4_butterfly_inverse_f16_mve() local [all …]
|
/cmsis-dsp-latest/Source/BasicMathFunctions/ |
D | arm_add_q31.c | 63 q31x4_t vecA; in arm_add_q31() local 74 vecA = vld1q(pSrcA); in arm_add_q31() 76 vst1q(pDst, vqaddq(vecA, vecB)); in arm_add_q31() 95 vecA = vld1q(pSrcA); in arm_add_q31() 97 vstrwq_p(pDst, vqaddq(vecA, vecB), p0); in arm_add_q31()
|
D | arm_sub_q31.c | 63 q31x4_t vecA; in arm_sub_q31() local 74 vecA = vld1q(pSrcA); in arm_sub_q31() 76 vst1q(pDst, vqsubq(vecA, vecB)); in arm_sub_q31() 95 vecA = vld1q(pSrcA); in arm_sub_q31() 97 vstrwq_p(pDst, vqsubq(vecA, vecB), p0); in arm_sub_q31()
|
D | arm_dot_prod_q31.c | 67 q31x4_t vecA; in arm_dot_prod_q31() local 79 vecA = vld1q(pSrcA); in arm_dot_prod_q31() 81 sum = vrmlaldavhaq(sum, vecA, vecB); in arm_dot_prod_q31() 99 vecA = vld1q(pSrcA); in arm_dot_prod_q31() 101 sum = vrmlaldavhaq_p(sum, vecA, vecB, p0); in arm_dot_prod_q31()
|