/cmsis-dsp-latest/dsppp/Include/dsppp/Helium/ |
D | matrix_multiply_f32.hpp | 139 T *pInA0, *pInA1, *pInA2; 148 pInA2 = pInA1 + pSrcA.stride(); 160 vecMac2 = inner::vmul(vecInB, *pInA2++); 169 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++); 178 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++); 209 T *pInA0, *pInA1, *pInA2, *pInA3; 217 pInA2 = pInA1 + pSrcA.stride(); 218 pInA3 = pInA2 + pSrcA.stride(); 227 vecMac2 = inner::vmul(vecInB, *pInA2++); 237 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++); [all …]
|
D | matrix_multiply.hpp | 123 T *pInA0, *pInA1, *pInA2, *pInA3; 142 pInA2 = pInA1 + pSrcA.stride(); 143 pInA3 = pInA2 + pSrcA.stride(); 161 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++); 193 pInA2 = pInA1 + pSrcA.stride(); 194 pInA3 = pInA2 + pSrcA.stride(); 214 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++);
|
D | matrix_multiply_fixed.hpp | 109 const T *pInA2 = pInA1 + pSrcA.stride(); in arm_mat_mult_3x3_mve() local 126 vecA2 = inner::vload1_z<1>(pInA2,MATRIX_DIM3,p0); in arm_mat_mult_3x3_mve() 202 const T *pInA2 = pInA1 + pSrcA.stride(); in arm_mat_mult_4x4_mve() local 203 const T *pInA3 = pInA2 + pSrcA.stride(); in arm_mat_mult_4x4_mve() 221 vecA2 = inner::vload1_z<1>(pInA2,MATRIX_DIM4,p0); in arm_mat_mult_4x4_mve() 343 const T *pInA2; 411 pInA2 = pInA + strideA; 417 pSrcA2Vec = (T const *) pInA2;
|
/cmsis-dsp-latest/Source/MatrixFunctions/ |
D | arm_mat_mult_f32.c | 148 float32_t *pInA0, *pInA1, *pInA2; in arm_mat_mult_f32_3x3_mve() local 157 pInA2 = pInA1 + MATRIX_DIM3; in arm_mat_mult_f32_3x3_mve() 169 vecMac2 = vmulq(vecInB, *pInA2++); in arm_mat_mult_f32_3x3_mve() 178 vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++); in arm_mat_mult_f32_3x3_mve() 187 vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++); in arm_mat_mult_f32_3x3_mve() 213 float32_t *pInA0, *pInA1, *pInA2, *pInA3; in arm_mat_mult_f32_4x4_mve() local 221 pInA2 = pInA1 + MATRIX_DIM4; in arm_mat_mult_f32_4x4_mve() 222 pInA3 = pInA2 + MATRIX_DIM4; in arm_mat_mult_f32_4x4_mve() 231 vecMac2 = vmulq(vecInB, *pInA2++); in arm_mat_mult_f32_4x4_mve() 241 vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++); in arm_mat_mult_f32_4x4_mve() [all …]
|
D | arm_mat_vec_mult_f32.c | 81 const float32_t *pInA2, *pInA3; in arm_mat_vec_mult_f32() local 91 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_f32() 92 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_f32() 107 pSrcA2Vec = pInA2; in arm_mat_vec_mult_f32() 296 const float32_t *pInA2; /* input data matrix pointer A of Q31 type */ in arm_mat_vec_mult_f32() local 329 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_f32() 330 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_f32() 341 matData = *(pInA2)++; in arm_mat_vec_mult_f32()
|
D | arm_mat_vec_mult_f16.c | 77 const float16_t *pInA2, *pInA3; in arm_mat_vec_mult_f16() local 87 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_f16() 88 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_f16() 103 pSrcA2Vec = pInA2; in arm_mat_vec_mult_f16() 291 const float16_t *pInA2; /* input data matrix pointer A of Q31 type */ in arm_mat_vec_mult_f16() local 324 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_f16() 325 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_f16() 336 matData = *(pInA2)++; in arm_mat_vec_mult_f16()
|
D | arm_mat_mult_opt_q31.c | 141 q31_t *pInA2 = pInA1 + MATRIX_DIM3; in arm_mat_mult_opt_q31_3x3_mve() local 158 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_3x3_mve() 179 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_3x3_mve() 200 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_3x3_mve() 227 q31_t *pInA2 = pInA1 + MATRIX_DIM4; in arm_mat_mult_opt_q31_4x4_mve() local 228 q31_t *pInA3 = pInA2 + MATRIX_DIM4; in arm_mat_mult_opt_q31_4x4_mve() 242 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_4x4_mve() 267 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_4x4_mve() 293 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_4x4_mve() 319 vecA = vldrwq_s32(pInA2); in arm_mat_mult_opt_q31_4x4_mve() [all …]
|
D | arm_mat_mult_q31.c | 137 q31_t *pInA2 = pInA1 + MATRIX_DIM3; in arm_mat_mult_q31_3x3_mve() local 154 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_3x3_mve() 175 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_3x3_mve() 196 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_3x3_mve() 223 q31_t *pInA2 = pInA1 + MATRIX_DIM4; in arm_mat_mult_q31_4x4_mve() local 224 q31_t *pInA3 = pInA2 + MATRIX_DIM4; in arm_mat_mult_q31_4x4_mve() 238 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_4x4_mve() 263 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_4x4_mve() 289 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_4x4_mve() 315 vecA = vldrwq_s32(pInA2); in arm_mat_mult_q31_4x4_mve() [all …]
|
D | arm_mat_cmplx_mult_f32.c | 146 float32_t *pInA2 = pInA1 + CMPLX_DIM * MATRIX_DIM3; in arm_mat_cmplx_mult_f32_3x3_mve() local 174 vecA = vldrwq_f32(pInA2); in arm_mat_cmplx_mult_f32_3x3_mve() 189 vecA = vldrwq_f32(&pInA2[4]); in arm_mat_cmplx_mult_f32_3x3_mve() 217 vecA = vldrwq_f32(pInA2); in arm_mat_cmplx_mult_f32_3x3_mve() 231 vecA = vldrwq_f32(&pInA2[4]); in arm_mat_cmplx_mult_f32_3x3_mve() 259 vecA = vldrwq_f32(pInA2); in arm_mat_cmplx_mult_f32_3x3_mve() 273 vecA = vldrwq_f32(&pInA2[4]); in arm_mat_cmplx_mult_f32_3x3_mve() 303 float32_t *pInA2 = pInA1 + CMPLX_DIM * MATRIX_DIM4; in arm_mat_cmplx_mult_f32_4x4_mve() local 304 float32_t *pInA3 = pInA2 + CMPLX_DIM * MATRIX_DIM4; in arm_mat_cmplx_mult_f32_4x4_mve() 330 vecA = vldrwq_f32(pInA2); in arm_mat_cmplx_mult_f32_4x4_mve() [all …]
|
D | arm_mat_cmplx_mult_q31.c | 140 q31_t const *pInA2 = pInA1 + CMPLX_DIM * MATRIX_DIM3; in arm_mat_cmplx_mult_q31_3x3_mve() local 186 vecA = vldrwq_s32(pInA2); in arm_mat_cmplx_mult_q31_3x3_mve() 190 vecA = vldrwq_z_s32(&pInA2[4], p0); in arm_mat_cmplx_mult_q31_3x3_mve() 227 vecA = vldrwq_s32(pInA2); in arm_mat_cmplx_mult_q31_3x3_mve() 231 vecA = vldrwq_z_s32(&pInA2[4], p0); in arm_mat_cmplx_mult_q31_3x3_mve() 268 vecA = vldrwq_s32(pInA2); in arm_mat_cmplx_mult_q31_3x3_mve() 272 vecA = vldrwq_z_s32(&pInA2[4], p0); in arm_mat_cmplx_mult_q31_3x3_mve() 295 q31_t const *pInA2 = pInA1 + CMPLX_DIM * MATRIX_DIM4; in arm_mat_cmplx_mult_q31_4x4_mve() local 296 q31_t const *pInA3 = pInA2 + CMPLX_DIM * MATRIX_DIM4; in arm_mat_cmplx_mult_q31_4x4_mve() 338 vecA = vldrwq_s32(pInA2); in arm_mat_cmplx_mult_q31_4x4_mve() [all …]
|
D | arm_mat_vec_mult_q15.c | 275 const q15_t *pInA2; /* input data matrix pointer A of Q15 type */ in arm_mat_vec_mult_q15() local 307 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_q15() 308 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_q15() 319 matData = read_q15x2_ia (&pInA2); in arm_mat_vec_mult_q15() 335 sum2 += (q63_t)*pInA2++ * vecData; in arm_mat_vec_mult_q15()
|
D | arm_mat_cmplx_mult_f16.c | 151 float16_t *pInA2 = pInA1 + CMPLX_DIM * MATRIX_DIM; in arm_mat_cmplx_mult_f16_3x3_mve() local 170 vecA2 = vldrhq_f16(pInA2); in arm_mat_cmplx_mult_f16_3x3_mve() 248 float16_t *pInA2 = pInA1 + CMPLX_DIM * MATRIX_DIM; in arm_mat_cmplx_mult_f16_4x4_mve() local 249 float16_t *pInA3 = pInA2 + CMPLX_DIM * MATRIX_DIM; in arm_mat_cmplx_mult_f16_4x4_mve() 272 vecA = vldrhq_f16(pInA2); in arm_mat_cmplx_mult_f16_4x4_mve() 301 vecA = vldrhq_f16(pInA2); in arm_mat_cmplx_mult_f16_4x4_mve() 330 vecA = vldrhq_f16(pInA2); in arm_mat_cmplx_mult_f16_4x4_mve() 359 vecA = vldrhq_f16(pInA2); in arm_mat_cmplx_mult_f16_4x4_mve() 481 float16_t const *pInA2 = pInA1 + numColsA * CMPLX_DIM; in arm_mat_cmplx_mult_f16() local 482 float16_t const *pInA3 = pInA2 + numColsA * CMPLX_DIM; in arm_mat_cmplx_mult_f16() [all …]
|
D | arm_mat_mult_f16.c | 425 float16_t *pInA0, *pInA1, *pInA2, *pInA3; in arm_mat_mult_f16() local 444 pInA2 = pInA1 + numColsA; in arm_mat_mult_f16() 445 pInA3 = pInA2 + numColsA; in arm_mat_mult_f16() 463 vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++); in arm_mat_mult_f16() 490 pInA2 = pInA1 + numColsA; in arm_mat_mult_f16() 491 pInA3 = pInA2 + numColsA; in arm_mat_mult_f16() 510 vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++); in arm_mat_mult_f16()
|
D | arm_mat_mult_q7.c | 120 q7_t *pInA2 = pInA1 + MATRIX_DIM; in arm_mat_mult_q7_3x3_mve() local 134 vecA2 = vldrbq_s8(pInA2); in arm_mat_mult_q7_3x3_mve() 190 q7_t *pInA2 = pInA1 + MATRIX_DIM; in arm_mat_mult_q7_4x4_mve() local 191 q7_t *pInA3 = pInA2 + MATRIX_DIM; in arm_mat_mult_q7_4x4_mve() 204 vecA2 = vldrbq_s8(pInA2); in arm_mat_mult_q7_4x4_mve() 278 q7_t *pInA2; in arm_mat_mult_q7() local 368 pInA2 = pInA + numColsA; in arm_mat_mult_q7() 373 pSrcA2Vec = (q7_t const *)pInA2; in arm_mat_mult_q7()
|
D | arm_mat_vec_mult_q7.c | 285 const q7_t *pInA2; /* input data matrix pointer of Q7 type */ in arm_mat_vec_mult_q7() local 319 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_q7() 320 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_q7() 337 matData = read_q7x4_ia (&pInA2); in arm_mat_vec_mult_q7() 364 sum2 += *pInA2++ * vecData; in arm_mat_vec_mult_q7()
|
D | arm_mat_mult_fast_q15.c | 89 q15_t *pInA2, *pInB2, *px2; in arm_mat_mult_fast_q15() local 258 pInA2 = pInA + numColsA; in arm_mat_mult_fast_q15() 277 inA2 = read_q15x2_ia (&pInA2); in arm_mat_mult_fast_q15() 314 inA2 = *pInA2++; in arm_mat_mult_fast_q15()
|
D | arm_mat_vec_mult_q31.c | 272 const q31_t *pInA2; /* input data matrix pointer A of Q31 type */ in arm_mat_vec_mult_q31() local 304 pInA2 = pInA1 + numCols; in arm_mat_vec_mult_q31() 305 pInA3 = pInA2 + numCols; in arm_mat_vec_mult_q31() 317 matData = *(pInA2)++; in arm_mat_vec_mult_q31()
|
D | arm_mat_mult_q15.c | 138 q15_t *pInA2 = pInA1 + MATRIX_DIM3; in arm_mat_mult_q15_3x3_mve() local 152 vecA2 = vldrhq_s16(pInA2); in arm_mat_mult_q15_3x3_mve() 219 q15_t *pInA2 = pInA1 + MATRIX_DIM4; in arm_mat_mult_q15_4x4_mve() local 220 q15_t *pInA3 = pInA2 + MATRIX_DIM4; in arm_mat_mult_q15_4x4_mve() 233 vecA2 = vldrhq_s16(pInA2); in arm_mat_mult_q15_4x4_mve() 330 q15_t *pInA2; in arm_mat_mult_q15() local 424 pInA2 = pInA + numColsA; in arm_mat_mult_q15() 430 pSrcA2Vec = (q15_t const *) pInA2; in arm_mat_mult_q15()
|
D | arm_mat_mult_fast_q31.c | 73 q31_t *pInA2; in arm_mat_mult_fast_q31() local 131 pInA2 = pInA + numColsA; in arm_mat_mult_fast_q31() 142 inA2 = *pInA2++; in arm_mat_mult_fast_q31()
|
/cmsis-dsp-latest/dsppp/Include/dsppp/Scalar/ |
D | matrix_multiply.hpp | 57 const T *pInA2; /* input data matrix pointer A of Q31 type */ in _dot_m_v() local 86 pInA2 = pInA1 + m.stride(); in _dot_m_v() 87 pInA3 = pInA2 + m.stride(); in _dot_m_v() 99 matData = *(pInA2)++; in _dot_m_v()
|
/cmsis-dsp-latest/dsppp/Include/dsppp/DSP/ |
D | matrix_multiply.hpp | 133 const T *pInA2; /* input data matrix pointer A of Q15 type */ in _dot_m_v() local 164 pInA2 = pInA1 + m.stride(); in _dot_m_v() 165 pInA3 = pInA2 + m.stride(); in _dot_m_v() 179 matData = inner::vload1<1> (pInA2); in _dot_m_v() 180 pInA2 += nb_lanes; in _dot_m_v() 200 sum2 = inner::mac(sum2,*pInA2++ , tmpData); in _dot_m_v()
|
/cmsis-dsp-latest/Source/SVMFunctions/ |
D | arm_svm_linear_predict_f16.c | 77 const float16_t *pInA2, *pInA3; in arm_svm_linear_predict_f16() local 87 pInA2 = pInA1 + numCols; in arm_svm_linear_predict_f16() 88 pInA3 = pInA2 + numCols; in arm_svm_linear_predict_f16() 103 pSrcA2Vec = pInA2; in arm_svm_linear_predict_f16()
|
D | arm_svm_polynomial_predict_f16.c | 103 const float16_t *pInA2, *pInA3; in arm_svm_polynomial_predict_f16() local 113 pInA2 = pInA1 + numCols; in arm_svm_polynomial_predict_f16() 114 pInA3 = pInA2 + numCols; in arm_svm_polynomial_predict_f16() 129 pSrcA2Vec = pInA2; in arm_svm_polynomial_predict_f16()
|
D | arm_svm_sigmoid_predict_f16.c | 80 const float16_t *pInA2, *pInA3; in arm_svm_sigmoid_predict_f16() local 90 pInA2 = pInA1 + numCols; in arm_svm_sigmoid_predict_f16() 91 pInA3 = pInA2 + numCols; in arm_svm_sigmoid_predict_f16() 106 pSrcA2Vec = pInA2; in arm_svm_sigmoid_predict_f16()
|
D | arm_svm_rbf_predict_f16.c | 80 const float16_t *pInA2, *pInA3; in arm_svm_rbf_predict_f16() local 90 pInA2 = pInA1 + numCols; in arm_svm_rbf_predict_f16() 91 pInA3 = pInA2 + numCols; in arm_svm_rbf_predict_f16() 106 pSrcA2Vec = pInA2; in arm_svm_rbf_predict_f16()
|