/hal_nxp-3.6.0/mcux/mcux-sdk/CMSIS/DSP/PrivateInclude/ |
D | arm_vec_filtering.h | 72 mve_pred16_t p0 = vctp32q(k); \ 76 acc1Vec = vfmaq_m_f32(acc1Vec, xVec, yVec, p0); \ 78 acc2Vec = vfmaq_m_f32(acc2Vec, xVec, yVec, p0); \ 80 acc3Vec = vfmaq_m_f32(acc3Vec, xVec, yVec, p0); \ 83 acc0Vec = vfmaq_m_f32(acc0Vec, xVec, yVec, p0); \ 117 mve_pred16_t p0 = vctp32q(k); \ 122 accVec = vfmaq_m_f32(accVec, xVec, yVec, p0);\ 155 mve_pred16_t p0 = vctp32q(k); \ 159 acc1Vec = vfmaq_m_f32(acc1Vec, xVec, yVec, p0); \ 162 p0 = vctp32q(k+1); \ [all …]
|
/hal_nxp-3.6.0/mcux/mcux-sdk/CMSIS/DSP/Source/MatrixFunctions/ |
D | arm_mat_ldlt_f32.c | 46 mve_pred16_t p0 = vctp32q(cnt); \ 48 tmpa=vldrwq_z_f32(&A[i*n + w],p0);\ 49 tmpb=vldrwq_z_f32(&A[j*n + w],p0);\ 51 vstrwq_p(&A[i*n + w], tmpb, p0); \ 52 vstrwq_p(&A[j*n + w], tmpa, p0); \ 130 mve_pred16_t p0; in arm_mat_ldlt_f32() local 131 p0 = vctp16q(cnt); in arm_mat_ldlt_f32() 135 vstrhq_p(&pp[k], vecP, p0); in arm_mat_ldlt_f32() 193 mve_pred16_t p0; in arm_mat_ldlt_f32() local 202 p0 = vctp32q(cnt); in arm_mat_ldlt_f32() [all …]
|
D | arm_mat_cholesky_f16.c | 89 mve_pred16_t p0; in arm_mat_cholesky_f16() local 118 p0 = vctp16q(kCnt); in arm_mat_cholesky_f16() 120 vecGi=vldrhq_z_f16(&pG[i * n + k],p0); in arm_mat_cholesky_f16() 122 vecGj0=vldrhq_z_f16(&pG[(j + 0) * n + k],p0); in arm_mat_cholesky_f16() 123 vecGj1=vldrhq_z_f16(&pG[(j + 1) * n + k],p0); in arm_mat_cholesky_f16() 124 vecGj2=vldrhq_z_f16(&pG[(j + 2) * n + k],p0); in arm_mat_cholesky_f16() 125 vecGj3=vldrhq_z_f16(&pG[(j + 3) * n + k],p0); in arm_mat_cholesky_f16() 127 acc0 = vfmsq_m(acc0, vecGi, vecGj0, p0); in arm_mat_cholesky_f16() 128 acc1 = vfmsq_m(acc1, vecGi, vecGj1, p0); in arm_mat_cholesky_f16() 129 acc2 = vfmsq_m(acc2, vecGi, vecGj2, p0); in arm_mat_cholesky_f16() [all …]
|
D | arm_mat_mult_q7.c | 73 mve_pred16_t p0 = vctp8q(MATRIX_DIM); in arm_mat_mult_q7_2x2_mve() local 79 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_2x2_mve() 94 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_2x2_mve() 123 mve_pred16_t p0 = vctp8q(MATRIX_DIM); in arm_mat_mult_q7_3x3_mve() local 130 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_3x3_mve() 148 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_3x3_mve() 162 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_3x3_mve() 194 mve_pred16_t p0 = vctp8q(MATRIX_DIM); in arm_mat_mult_q7_4x4_mve() local 200 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_4x4_mve() 221 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_4x4_mve() [all …]
|
D | arm_mat_cholesky_f32.c | 97 mve_pred16_t p0; in arm_mat_cholesky_f32() local 124 p0 = vctp32q(kCnt); in arm_mat_cholesky_f32() 126 vecGi=vldrwq_z_f32(&pG[i * n + k],p0); in arm_mat_cholesky_f32() 128 vecGj0=vldrwq_z_f32(&pG[(j + 0) * n + k],p0); in arm_mat_cholesky_f32() 129 vecGj1=vldrwq_z_f32(&pG[(j + 1) * n + k],p0); in arm_mat_cholesky_f32() 130 vecGj2=vldrwq_z_f32(&pG[(j + 2) * n + k],p0); in arm_mat_cholesky_f32() 131 vecGj3=vldrwq_z_f32(&pG[(j + 3) * n + k],p0); in arm_mat_cholesky_f32() 133 acc0 = vfmaq_m(acc0, vecGi, vecGj0, p0); in arm_mat_cholesky_f32() 134 acc1 = vfmaq_m(acc1, vecGi, vecGj1, p0); in arm_mat_cholesky_f32() 135 acc2 = vfmaq_m(acc2, vecGi, vecGj2, p0); in arm_mat_cholesky_f32() [all …]
|
D | arm_mat_inverse_f16.c | 247 mve_pred16_t p0 = vctp16q(blkCnt); in arm_mat_inverse_f16() local 251 vstrhq_p_f16(pTmpB, vecA, p0); in arm_mat_inverse_f16() 252 vstrhq_p_f16(pTmpA, vecB, p0); in arm_mat_inverse_f16() 280 mve_pred16_t p0 = vctp16q(blkCnt); in arm_mat_inverse_f16() local 284 vstrhq_p_f16(pTmpB, vecA, p0); in arm_mat_inverse_f16() 285 vstrhq_p_f16(pTmpA, vecB, p0); in arm_mat_inverse_f16() 350 mve_pred16_t p0 = vctp16q(blkCnt); in arm_mat_inverse_f16() local 355 vstrhq_p_f16(pTmpA, vecA, p0); in arm_mat_inverse_f16() 382 mve_pred16_t p0 = vctp16q(blkCnt); in arm_mat_inverse_f16() local 386 vstrhq_p_f16(pTmpA, vecA, p0); in arm_mat_inverse_f16() [all …]
|
/hal_nxp-3.6.0/mcux/mcux-sdk/CMSIS/DSP/Source/StatisticsFunctions/ |
D | arm_min_f16.c | 73 mve_pred16_t p0; in arm_min_f16() local 87 p0 = vcmpleq(vecSrc, curExtremValVec); in arm_min_f16() 88 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_min_f16() 89 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_min_f16() 105 p0 = vctp16q(blkCnt); in arm_min_f16() 110 p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); in arm_min_f16() 111 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_min_f16() 112 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_min_f16() 121 p0 = vcmpleq(curExtremValVec, minValue); in arm_min_f16() 122 indexVec = vpselq(curExtremIdxVec, vdupq_n_u16(blockSize), p0); in arm_min_f16()
|
D | arm_absmax_f16.c | 73 mve_pred16_t p0; in arm_absmax_f16() local 90 p0 = vcmpgeq(vecSrc, curExtremValVec); in arm_absmax_f16() 91 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmax_f16() 92 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmax_f16() 111 p0 = vctp16q(blkCnt); in arm_absmax_f16() 116 p0 = vcmpgeq_m(vecSrc, curExtremValVec, p0); in arm_absmax_f16() 117 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmax_f16() 118 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmax_f16() 127 p0 = vcmpgeq(curExtremValVec, maxValue); in arm_absmax_f16() 128 indexVec = vpselq(curExtremIdxVec, vdupq_n_u16(blockSize), p0); in arm_absmax_f16()
|
D | arm_absmin_f32.c | 79 mve_pred16_t p0; in arm_absmin_f32() local 96 p0 = vcmpleq(vecSrc, curExtremValVec); in arm_absmin_f32() 97 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_f32() 98 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_f32() 113 p0 = vctp32q(blkCnt); in arm_absmin_f32() 122 p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); in arm_absmin_f32() 123 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_f32() 124 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_f32() 133 p0 = vcmpleq(curExtremValVec, minValue); in arm_absmin_f32() 134 indexVec = vpselq(curExtremIdxVec, vdupq_n_u32(blockSize), p0); in arm_absmin_f32()
|
D | arm_absmin_f16.c | 74 mve_pred16_t p0; in arm_absmin_f16() local 91 p0 = vcmpleq(vecSrc, curExtremValVec); in arm_absmin_f16() 92 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_f16() 93 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_f16() 108 p0 = vctp16q(blkCnt); in arm_absmin_f16() 117 p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); in arm_absmin_f16() 118 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_f16() 119 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_f16() 128 p0 = vcmpleq(curExtremValVec, minValue); in arm_absmin_f16() 129 indexVec = vpselq(curExtremIdxVec, vdupq_n_u16(blockSize), p0); in arm_absmin_f16()
|
D | arm_absmin_q31.c | 69 mve_pred16_t p0; in arm_absmin_q31() local 86 p0 = vcmpleq(vecSrc, curExtremValVec); in arm_absmin_q31() 87 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_q31() 88 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_q31() 107 p0 = vctp32q(blkCnt); in arm_absmin_q31() 112 p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); in arm_absmin_q31() 113 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_q31() 114 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_q31() 123 p0 = vcmpleq(curExtremValVec, minValue); in arm_absmin_q31() 124 indexVec = vpselq(curExtremIdxVec, vdupq_n_u32(blockSize), p0); in arm_absmin_q31()
|
D | arm_absmin_q15.c | 68 mve_pred16_t p0; in arm_absmin_q15() local 85 p0 = vcmpleq(vecSrc, curExtremValVec); in arm_absmin_q15() 86 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_q15() 87 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_q15() 106 p0 = vctp16q(blkCnt); in arm_absmin_q15() 111 p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); in arm_absmin_q15() 112 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_absmin_q15() 113 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_absmin_q15() 122 p0 = vcmpleq(curExtremValVec, minValue); in arm_absmin_q15() 123 indexVec = vpselq(curExtremIdxVec, vdupq_n_u16(blockSize), p0); in arm_absmin_q15()
|
D | arm_absmin_q7.c | 71 mve_pred16_t p0; in arm_small_blk_absmin_q7() local 88 p0 = vcmpleq(vecSrc, curExtremValVec); in arm_small_blk_absmin_q7() 89 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_small_blk_absmin_q7() 90 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_small_blk_absmin_q7() 109 p0 = vctp8q(blkCnt); in arm_small_blk_absmin_q7() 114 p0 = vcmpleq_m(vecSrc, curExtremValVec, p0); in arm_small_blk_absmin_q7() 115 curExtremValVec = vpselq(vecSrc, curExtremValVec, p0); in arm_small_blk_absmin_q7() 116 curExtremIdxVec = vpselq(indexVec, curExtremIdxVec, p0); in arm_small_blk_absmin_q7() 125 p0 = vcmpleq(curExtremValVec, minValue); in arm_small_blk_absmin_q7() 126 idx = vminvq_p_u8(idx, curExtremIdxVec, p0); in arm_small_blk_absmin_q7()
|
/hal_nxp-3.6.0/mcux/mcux-sdk/CMSIS/DSP/Source/TransformFunctions/ |
D | arm_cfft_radix2_f16.c | 125 float16_t p0, p1, p2, p3; in arm_radix2_butterfly_f16() local 155 p0 = xt * cosVal; in arm_radix2_butterfly_f16() 163 pSrc[2 * l] = p0 + p1; in arm_radix2_butterfly_f16() 197 p0 = xt * cosVal; in arm_radix2_butterfly_f16() 205 pSrc[2 * l] = p0 + p1; in arm_radix2_butterfly_f16() 260 p0 = xt * cosVal; in arm_radix2_butterfly_f16() 268 pSrc[2 * l] = p0 + p1; in arm_radix2_butterfly_f16() 294 float16_t p0, p1, p2, p3; in arm_radix2_butterfly_inverse_f16() local 316 p0 = xt * cosVal; in arm_radix2_butterfly_inverse_f16() 324 pSrc[2 * l] = p0 - p1; in arm_radix2_butterfly_inverse_f16() [all …]
|
D | arm_cfft_radix2_f32.c | 123 float32_t p0, p1, p2, p3; in arm_radix2_butterfly_f32() local 153 p0 = xt * cosVal; in arm_radix2_butterfly_f32() 161 pSrc[2 * l] = p0 + p1; in arm_radix2_butterfly_f32() 195 p0 = xt * cosVal; in arm_radix2_butterfly_f32() 203 pSrc[2 * l] = p0 + p1; in arm_radix2_butterfly_f32() 258 p0 = xt * cosVal; in arm_radix2_butterfly_f32() 266 pSrc[2 * l] = p0 + p1; in arm_radix2_butterfly_f32() 292 float32_t p0, p1, p2, p3; in arm_radix2_butterfly_inverse_f32() local 314 p0 = xt * cosVal; in arm_radix2_butterfly_inverse_f32() 322 pSrc[2 * l] = p0 - p1; in arm_radix2_butterfly_inverse_f32() [all …]
|
D | arm_cfft_radix2_q31.c | 99 q31_t p0, p1; in arm_radix2_butterfly_q31() local 123 mult_32x32_keep32_R(p0, xt, cosVal); in arm_radix2_butterfly_q31() 125 multAcc_32x32_keep32_R(p0, yt, sinVal); in arm_radix2_butterfly_q31() 128 pSrc[2U * l] = p0; in arm_radix2_butterfly_q31() 161 mult_32x32_keep32_R(p0, xt, cosVal); in arm_radix2_butterfly_q31() 163 multAcc_32x32_keep32_R(p0, yt, sinVal); in arm_radix2_butterfly_q31() 166 pSrc[2U * l] = p0; in arm_radix2_butterfly_q31() 227 q31_t p0, p1; in arm_radix2_butterfly_inverse_q31() local 251 mult_32x32_keep32_R(p0, xt, cosVal); in arm_radix2_butterfly_inverse_q31() 253 multSub_32x32_keep32_R(p0, yt, sinVal); in arm_radix2_butterfly_inverse_q31() [all …]
|
D | arm_cfft_q31.c | 279 mve_pred16_t p0 = vctp32q(blkCnt); in arm_cfft_radix4by2_q31_mve() local 283 vstrwq_p(pIn0, vecIn0, p0); in arm_cfft_radix4by2_q31_mve() 535 mve_pred16_t p0 = vctp32q(blkCnt); in arm_cfft_radix4by2_inverse_q31_mve() local 539 vstrwq_p(pIn0, vecIn0, p0); in arm_cfft_radix4by2_inverse_q31_mve() 734 q31_t p0, p1; in arm_cfft_radix4by2_q31() local 750 mult_32x32_keep32_R(p0, xt, cosVal); in arm_cfft_radix4by2_q31() 752 multAcc_32x32_keep32_R(p0, yt, sinVal); in arm_cfft_radix4by2_q31() 755 pSrc[2 * l] = p0 << 1; in arm_cfft_radix4by2_q31() 769 p0 = pSrc[4 * i + 0]; in arm_cfft_radix4by2_q31() 774 p0 <<= 1U; in arm_cfft_radix4by2_q31() [all …]
|
/hal_nxp-3.6.0/mcux/mcux-sdk/CMSIS/DSP/Source/SVMFunctions/ |
D | arm_svm_linear_predict_f16.c | 134 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_linear_predict_f16() local 137 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_linear_predict_f16() 138 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_linear_predict_f16() 140 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_linear_predict_f16() 142 vecA = vldrhq_z_f16(pSrcA2Vec, p0); in arm_svm_linear_predict_f16() 144 vecA = vldrhq_z_f16(pSrcA3Vec, p0); in arm_svm_linear_predict_f16() 210 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_linear_predict_f16() local 213 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_linear_predict_f16() 214 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_linear_predict_f16() 216 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_linear_predict_f16() [all …]
|
D | arm_svm_sigmoid_predict_f16.c | 137 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_sigmoid_predict_f16() local 140 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_sigmoid_predict_f16() 141 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_sigmoid_predict_f16() 143 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_sigmoid_predict_f16() 145 vecA = vldrhq_z_f16(pSrcA2Vec, p0); in arm_svm_sigmoid_predict_f16() 147 vecA = vldrhq_z_f16(pSrcA3Vec, p0); in arm_svm_sigmoid_predict_f16() 218 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_sigmoid_predict_f16() local 221 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_sigmoid_predict_f16() 222 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_sigmoid_predict_f16() 224 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_sigmoid_predict_f16() [all …]
|
D | arm_svm_polynomial_predict_f16.c | 137 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_polynomial_predict_f16() local 140 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_polynomial_predict_f16() 141 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_polynomial_predict_f16() 143 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_polynomial_predict_f16() 145 vecA = vldrhq_z_f16(pSrcA2Vec, p0); in arm_svm_polynomial_predict_f16() 147 vecA = vldrhq_z_f16(pSrcA3Vec, p0); in arm_svm_polynomial_predict_f16() 219 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_polynomial_predict_f16() local 222 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_polynomial_predict_f16() 223 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_polynomial_predict_f16() 225 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_polynomial_predict_f16() [all …]
|
D | arm_svm_rbf_predict_f16.c | 142 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_rbf_predict_f16() local 146 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_rbf_predict_f16() 147 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_rbf_predict_f16() 150 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_rbf_predict_f16() 153 vecA = vldrhq_z_f16(pSrcA2Vec, p0);; in arm_svm_rbf_predict_f16() 156 vecA = vldrhq_z_f16(pSrcA3Vec, p0); in arm_svm_rbf_predict_f16() 231 mve_pred16_t p0 = vctp16q(blkCnt); in arm_svm_rbf_predict_f16() local 234 vecIn = vldrhq_z_f16(pInVec, p0); in arm_svm_rbf_predict_f16() 235 vecA = vldrhq_z_f16(pSrcA0Vec, p0); in arm_svm_rbf_predict_f16() 238 vecA = vldrhq_z_f16(pSrcA1Vec, p0); in arm_svm_rbf_predict_f16() [all …]
|
D | arm_svm_linear_predict_f32.c | 131 mve_pred16_t p0 = vctp32q(blkCnt); in arm_svm_linear_predict_f32() local 134 vecIn = vldrwq_z_f32(pInVec, p0); in arm_svm_linear_predict_f32() 135 vecA = vldrwq_z_f32(pSrcA0Vec, p0); in arm_svm_linear_predict_f32() 137 vecA = vldrwq_z_f32(pSrcA1Vec, p0); in arm_svm_linear_predict_f32() 139 vecA = vldrwq_z_f32(pSrcA2Vec, p0); in arm_svm_linear_predict_f32() 141 vecA = vldrwq_z_f32(pSrcA3Vec, p0); in arm_svm_linear_predict_f32() 208 mve_pred16_t p0 = vctp32q(blkCnt); in arm_svm_linear_predict_f32() local 211 vecIn = vldrwq_z_f32(pInVec, p0); in arm_svm_linear_predict_f32() 212 vecA = vldrwq_z_f32(pSrcA0Vec, p0); in arm_svm_linear_predict_f32() 214 vecA = vldrwq_z_f32(pSrcA1Vec, p0); in arm_svm_linear_predict_f32() [all …]
|
/hal_nxp-3.6.0/mcux/mcux-sdk/CMSIS/DSP/Source/FilteringFunctions/ |
D | arm_fir_interpolate_q15.c | 147 mve_pred16_t p0 = vctp16q(tapCnt); in arm_fir_interpolate_q15() local 149 vecState = vldrhq_z_s16(ptr1, p0); in arm_fir_interpolate_q15() 151 vecCoef = vldrhq_gather_shifted_offset_z_s16(ptr2, vec_strides3, p0); in arm_fir_interpolate_q15() 154 vecCoef = vldrhq_gather_shifted_offset_z_s16(ptr2, vec_strides2, p0); in arm_fir_interpolate_q15() 157 vecCoef = vldrhq_gather_shifted_offset_z_s16(ptr2, vec_strides1, p0); in arm_fir_interpolate_q15() 160 vecCoef = vldrhq_gather_shifted_offset_z_s16(ptr2, vec_strides0, p0); in arm_fir_interpolate_q15() 207 mve_pred16_t p0 = vctp16q(tapCnt); in arm_fir_interpolate_q15() local 209 vecState = vldrhq_z_s16(ptr1, p0); in arm_fir_interpolate_q15() 211 vecCoef = vldrhq_gather_shifted_offset_z_s16(ptr2, vec_strides2, p0); in arm_fir_interpolate_q15() 214 vecCoef = vldrhq_gather_shifted_offset_z_s16(ptr2, vec_strides1, p0); in arm_fir_interpolate_q15() [all …]
|
D | arm_fir_interpolate_q31.c | 144 mve_pred16_t p0 = vctp32q(tapCnt); in arm_fir_interpolate_q31() local 146 vecState = vldrwq_z_s32(ptr1, p0); in arm_fir_interpolate_q31() 148 vecCoef = vldrwq_gather_shifted_offset_z_s32(ptr2, vec_strides3, p0); in arm_fir_interpolate_q31() 151 vecCoef = vldrwq_gather_shifted_offset_z_s32(ptr2, vec_strides2, p0); in arm_fir_interpolate_q31() 154 vecCoef = vldrwq_gather_shifted_offset_z_s32(ptr2, vec_strides1, p0); in arm_fir_interpolate_q31() 157 vecCoef = vldrwq_gather_shifted_offset_z_s32(ptr2, vec_strides0, p0); in arm_fir_interpolate_q31() 204 mve_pred16_t p0 = vctp32q(tapCnt); in arm_fir_interpolate_q31() local 206 vecState = vldrwq_z_s32(ptr1, p0); in arm_fir_interpolate_q31() 208 vecCoef = vldrwq_gather_shifted_offset_z_s32(ptr2, vec_strides2, p0); in arm_fir_interpolate_q31() 211 vecCoef = vldrwq_gather_shifted_offset_z_s32(ptr2, vec_strides1, p0); in arm_fir_interpolate_q31() [all …]
|
D | arm_correlate_f16.c | 127 mve_pred16_t p0 = vctp16q(k+1); \ 130 … acc1Vec = vfmaq_m_f16(acc1Vec, xVec, yVec,p0); \ 133 p0 = vctp16q(k); \ 135 … acc0Vec = vfmaq_m_f16(acc0Vec, xVec, yVec,p0); \ 164 mve_pred16_t p0 = vctp16q(k); \ 167 … accVec = vfmaq_m(accVec, xVec, yVec, p0); \ 205 mve_pred16_t p0 = vctp16q(k); \ 208 … acc1Vec = vfmaq_m_f16(acc1Vec, xVec, yVec, p0); \ 210 … acc2Vec = vfmaq_m_f16(acc2Vec, xVec, yVec, p0); \ 212 … acc3Vec = vfmaq_m_f16(acc3Vec, xVec, yVec, p0); \ [all …]
|