/cmsis-dsp-latest/dsppp/Include/dsppp/Helium/ |
D | matrix_multiply_f16.hpp | 34 const MA &pSrcA, 42 const uint16_t offsetA[8] = { 0, 0, (uint16_t)pSrcA.stride(), (uint16_t)pSrcA.stride(), 43 0, 0, (uint16_t)pSrcA.stride(), (uint16_t)pSrcA.stride() }; 63 vecInA = vldrhq_gather_shifted_offset(pSrcA.const_ptr(), vecOffsA); 82 vecInA = vldrhq_gather_shifted_offset(pSrcA.const_ptr(), vecOffsA); 119 const MA &pSrcA, 124 … (uint16_t)pSrcA.stride(), (uint16_t)pSrcA.stride(), (uint16_t)pSrcA.stride(), 125 … (uint16_t)(2U*pSrcA.stride()), (uint16_t)(2U*pSrcA.stride()) }; 147 vecInA = vldrhq_gather_shifted_offset(pSrcA.const_ptr(), vecOffsA); 166 vecInA = vldrhq_gather_shifted_offset(pSrcA.const_ptr(), vecOffsA); [all …]
|
/cmsis-dsp-latest/Source/MatrixFunctions/ |
D | arm_mat_add_f32.c | 84 const arm_matrix_instance_f32 * pSrcA, in arm_mat_add_f32() argument 96 pDataA = pSrcA->pData; in arm_mat_add_f32() 104 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_f32() 105 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_add_f32() 106 (pSrcA->numRows != pDst->numRows) || (pSrcA->numCols != pDst->numCols)) in arm_mat_add_f32() 117 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_add_f32() 162 const arm_matrix_instance_f32 * pSrcA, in arm_mat_add_f32() argument 166 float32_t *pIn1 = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_add_f32() 177 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_f32() 178 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_add_f32() [all …]
|
D | arm_mat_sub_f32.c | 81 const arm_matrix_instance_f32 * pSrcA, in arm_mat_sub_f32() argument 93 pDataA = pSrcA->pData; in arm_mat_sub_f32() 101 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_f32() 102 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_sub_f32() 103 (pSrcA->numRows != pDst->numRows) || (pSrcA->numCols != pDst->numCols)) in arm_mat_sub_f32() 114 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_sub_f32() 155 const arm_matrix_instance_f32 * pSrcA, in arm_mat_sub_f32() argument 159 float32_t *pIn1 = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_sub_f32() 170 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_f32() 171 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_sub_f32() [all …]
|
D | arm_mat_sub_f16.c | 56 const arm_matrix_instance_f16 * pSrcA, in arm_mat_sub_f16() argument 68 pDataA = pSrcA->pData; in arm_mat_sub_f16() 76 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_f16() 77 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_sub_f16() 78 (pSrcA->numRows != pDst->numRows) || (pSrcA->numCols != pDst->numCols)) in arm_mat_sub_f16() 89 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_sub_f16() 130 const arm_matrix_instance_f16 * pSrcA, in arm_mat_sub_f16() argument 134 float16_t *pInA = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_sub_f16() 145 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_f16() 146 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_sub_f16() [all …]
|
D | arm_mat_add_f16.c | 58 const arm_matrix_instance_f16 * pSrcA, in arm_mat_add_f16() argument 70 pDataA = pSrcA->pData; in arm_mat_add_f16() 78 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_f16() 79 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_add_f16() 80 (pSrcA->numRows != pDst->numRows) || (pSrcA->numCols != pDst->numCols)) in arm_mat_add_f16() 91 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_add_f16() 129 const arm_matrix_instance_f16 * pSrcA, in arm_mat_add_f16() argument 133 float16_t *pInA = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_add_f16() 144 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_f16() 145 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_add_f16() [all …]
|
D | arm_mat_add_q31.c | 55 const arm_matrix_instance_q31 * pSrcA, in arm_mat_add_q31() argument 67 pDataA = pSrcA->pData; in arm_mat_add_q31() 76 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_q31() 77 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_add_q31() 78 (pSrcA->numRows != pDst->numRows) || in arm_mat_add_q31() 79 (pSrcA->numCols != pDst->numCols) ) in arm_mat_add_q31() 90 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_add_q31() 131 const arm_matrix_instance_q31 * pSrcA, in arm_mat_add_q31() argument 135 q31_t *pInA = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_add_q31() 146 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_q31() [all …]
|
D | arm_mat_sub_q31.c | 55 const arm_matrix_instance_q31 * pSrcA, in arm_mat_sub_q31() argument 67 pDataA = pSrcA->pData; in arm_mat_sub_q31() 76 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_q31() 77 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_sub_q31() 78 (pSrcA->numRows != pDst->numRows) || in arm_mat_sub_q31() 79 (pSrcA->numCols != pDst->numCols) ) in arm_mat_sub_q31() 92 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_sub_q31() 133 const arm_matrix_instance_q31 * pSrcA, in arm_mat_sub_q31() argument 137 q31_t *pInA = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_sub_q31() 148 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_q31() [all …]
|
D | arm_mat_add_q15.c | 56 const arm_matrix_instance_q15 * pSrcA, in arm_mat_add_q15() argument 68 pDataA = pSrcA->pData; in arm_mat_add_q15() 77 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_q15() 78 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_add_q15() 79 (pSrcA->numRows != pDst->numRows) || in arm_mat_add_q15() 80 (pSrcA->numCols != pDst->numCols) ) in arm_mat_add_q15() 94 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_add_q15() 132 const arm_matrix_instance_q15 * pSrcA, in arm_mat_add_q15() argument 136 q15_t *pInA = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_add_q15() 147 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_add_q15() [all …]
|
D | arm_mat_sub_q15.c | 56 const arm_matrix_instance_q15 * pSrcA, in arm_mat_sub_q15() argument 69 pDataA = pSrcA->pData; in arm_mat_sub_q15() 78 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_q15() 79 (pSrcA->numCols != pSrcB->numCols) || in arm_mat_sub_q15() 80 (pSrcA->numRows != pDst->numRows) || in arm_mat_sub_q15() 81 (pSrcA->numCols != pDst->numCols) ) in arm_mat_sub_q15() 93 numSamples = (uint32_t) pSrcA->numRows * pSrcA->numCols; in arm_mat_sub_q15() 129 const arm_matrix_instance_q15 * pSrcA, in arm_mat_sub_q15() argument 133 q15_t *pInA = pSrcA->pData; /* input data matrix pointer A */ in arm_mat_sub_q15() 144 if ((pSrcA->numRows != pSrcB->numRows) || in arm_mat_sub_q15() [all …]
|
/cmsis-dsp-latest/Source/ComplexMathFunctions/ |
D | arm_cmplx_mult_cmplx_f16.c | 55 const float16_t * pSrcA, in arm_cmplx_mult_cmplx_f16() argument 69 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f16() 71 pSrcA += 8; in arm_cmplx_mult_cmplx_f16() 76 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f16() 77 pSrcA += 8; in arm_cmplx_mult_cmplx_f16() 86 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f16() 87 pSrcA += 8; in arm_cmplx_mult_cmplx_f16() 102 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f16() 120 pSrcA += 8; in arm_cmplx_mult_cmplx_f16() 123 vecSrcA = vldrhq_z_f16(pSrcA, p); in arm_cmplx_mult_cmplx_f16() [all …]
|
D | arm_cmplx_dot_prod_q15.c | 58 const q15_t * pSrcA, in arm_cmplx_dot_prod_q15() argument 74 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 76 pSrcA += 8; in arm_cmplx_dot_prod_q15() 82 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 83 pSrcA += 8; in arm_cmplx_dot_prod_q15() 90 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 91 pSrcA += 8; in arm_cmplx_dot_prod_q15() 104 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 110 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 122 pSrcA += 8; in arm_cmplx_dot_prod_q15() [all …]
|
D | arm_cmplx_dot_prod_q31.c | 60 const q31_t * pSrcA, in arm_cmplx_dot_prod_q31() argument 76 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 78 pSrcA += 4; in arm_cmplx_dot_prod_q31() 84 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 85 pSrcA += 4; in arm_cmplx_dot_prod_q31() 92 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 93 pSrcA += 4; in arm_cmplx_dot_prod_q31() 106 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 112 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 124 pSrcA += 4; in arm_cmplx_dot_prod_q31() [all …]
|
D | arm_cmplx_mult_cmplx_f32.c | 73 const float32_t * pSrcA, in arm_cmplx_mult_cmplx_f32() argument 87 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f32() 89 pSrcA += 4; in arm_cmplx_mult_cmplx_f32() 94 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f32() 95 pSrcA += 4; in arm_cmplx_mult_cmplx_f32() 104 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f32() 105 pSrcA += 4; in arm_cmplx_mult_cmplx_f32() 120 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f32() 138 pSrcA += 4; in arm_cmplx_mult_cmplx_f32() 141 vecSrcA = vldrwq_z_f32(pSrcA, p); in arm_cmplx_mult_cmplx_f32() [all …]
|
D | arm_cmplx_dot_prod_f16.c | 58 const float16_t * pSrcA, in arm_cmplx_dot_prod_f16() argument 74 vecSrcA = vld1q( pSrcA); in arm_cmplx_dot_prod_f16() 76 pSrcA += 8; in arm_cmplx_dot_prod_f16() 81 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_f16() 82 pSrcA += 8; in arm_cmplx_dot_prod_f16() 89 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_f16() 90 pSrcA += 8; in arm_cmplx_dot_prod_f16() 103 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_f16() 117 pSrcA += 8; in arm_cmplx_dot_prod_f16() 120 vecSrcA = vldrhq_z_f16(pSrcA, p); in arm_cmplx_dot_prod_f16() [all …]
|
D | arm_cmplx_dot_prod_f32.c | 79 const float32_t * pSrcA, in arm_cmplx_dot_prod_f32() argument 95 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_f32() 97 pSrcA += 4; in arm_cmplx_dot_prod_f32() 102 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_f32() 103 pSrcA += 4; in arm_cmplx_dot_prod_f32() 110 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_f32() 111 pSrcA += 4; in arm_cmplx_dot_prod_f32() 124 vecSrcC = vld1q(pSrcA); in arm_cmplx_dot_prod_f32() 138 pSrcA += 4; in arm_cmplx_dot_prod_f32() 140 vecSrcA = vldrwq_z_f32(pSrcA, p); in arm_cmplx_dot_prod_f32() [all …]
|
D | arm_cmplx_mult_cmplx_q31.c | 54 const q31_t * pSrcA, in arm_cmplx_mult_cmplx_q31() argument 68 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q31() 70 pSrcA += 4; in arm_cmplx_mult_cmplx_q31() 77 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q31() 78 pSrcA += 4; in arm_cmplx_mult_cmplx_q31() 89 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q31() 90 pSrcA += 4; in arm_cmplx_mult_cmplx_q31() 107 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q31() 128 pSrcA += 4; in arm_cmplx_mult_cmplx_q31() 131 vecSrcA = vldrwq_z_s32(pSrcA, p); in arm_cmplx_mult_cmplx_q31() [all …]
|
D | arm_cmplx_mult_cmplx_q15.c | 54 const q15_t * pSrcA, in arm_cmplx_mult_cmplx_q15() argument 69 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q15() 71 pSrcA += 8; in arm_cmplx_mult_cmplx_q15() 79 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q15() 80 pSrcA += 8; in arm_cmplx_mult_cmplx_q15() 91 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q15() 92 pSrcA += 8; in arm_cmplx_mult_cmplx_q15() 109 vecSrcC = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q15() 131 pSrcA += 8; in arm_cmplx_mult_cmplx_q15() 134 vecSrcA = vldrhq_z_s16(pSrcA, p); in arm_cmplx_mult_cmplx_q15() [all …]
|
/cmsis-dsp-latest/Include/dsp/ |
D | basic_math_functions.h | 54 const q7_t * pSrcA, 68 const q15_t * pSrcA, 82 const q31_t * pSrcA, 96 const float32_t * pSrcA, 111 const float64_t * pSrcA, 126 const float32_t * pSrcA, 141 const float64_t * pSrcA, 156 const q7_t * pSrcA, 170 const q15_t * pSrcA, 184 const q31_t * pSrcA, [all …]
|
/cmsis-dsp-latest/Source/BasicMathFunctions/ |
D | arm_mult_q7.c | 56 const q7_t * pSrcA, in arm_mult_q7() argument 72 vecA = vld1q(pSrcA); in arm_mult_q7() 82 pSrcA += 16; in arm_mult_q7() 93 vecA = vld1q(pSrcA); in arm_mult_q7() 101 const q7_t * pSrcA, in arm_mult_q7() argument 123 out1 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++)) >> 7), 8); in arm_mult_q7() 124 out2 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++)) >> 7), 8); in arm_mult_q7() 125 out3 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++)) >> 7), 8); in arm_mult_q7() 126 out4 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++)) >> 7), 8); in arm_mult_q7() 131 *pDst++ = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++)) >> 7), 8); in arm_mult_q7() [all …]
|
D | arm_add_q15.c | 57 const q15_t * pSrcA, in arm_add_q15() argument 74 vecA = vld1q(pSrcA); in arm_add_q15() 84 pSrcA += 8; in arm_add_q15() 95 vecA = vld1q(pSrcA); in arm_add_q15() 103 const q15_t * pSrcA, in arm_add_q15() argument 126 inA1 = read_q15x2_ia (&pSrcA); in arm_add_q15() 127 inA2 = read_q15x2_ia (&pSrcA); in arm_add_q15() 136 *pDst++ = (q15_t) __SSAT(((q31_t) *pSrcA++ + *pSrcB++), 16); in arm_add_q15() 137 *pDst++ = (q15_t) __SSAT(((q31_t) *pSrcA++ + *pSrcB++), 16); in arm_add_q15() 138 *pDst++ = (q15_t) __SSAT(((q31_t) *pSrcA++ + *pSrcB++), 16); in arm_add_q15() [all …]
|
D | arm_sub_q15.c | 57 const q15_t * pSrcA, in arm_sub_q15() argument 74 vecA = vld1q(pSrcA); in arm_sub_q15() 84 pSrcA += 8; in arm_sub_q15() 95 vecA = vld1q(pSrcA); in arm_sub_q15() 104 const q15_t * pSrcA, in arm_sub_q15() argument 127 inA1 = read_q15x2_ia (&pSrcA); in arm_sub_q15() 128 inA2 = read_q15x2_ia (&pSrcA); in arm_sub_q15() 137 *pDst++ = (q15_t) __SSAT(((q31_t) *pSrcA++ - *pSrcB++), 16); in arm_sub_q15() 138 *pDst++ = (q15_t) __SSAT(((q31_t) *pSrcA++ - *pSrcB++), 16); in arm_sub_q15() 139 *pDst++ = (q15_t) __SSAT(((q31_t) *pSrcA++ - *pSrcB++), 16); in arm_sub_q15() [all …]
|
D | arm_dot_prod_q15.c | 59 const q15_t * pSrcA, in arm_dot_prod_q15() argument 77 vecA = vld1q(pSrcA); in arm_dot_prod_q15() 87 pSrcA += 8; in arm_dot_prod_q15() 97 vecA = vld1q(pSrcA); in arm_dot_prod_q15() 107 const q15_t * pSrcA, in arm_dot_prod_q15() argument 126 sum = __SMLALD(read_q15x2_ia (&pSrcA), read_q15x2_ia (&pSrcB), sum); in arm_dot_prod_q15() 127 sum = __SMLALD(read_q15x2_ia (&pSrcA), read_q15x2_ia (&pSrcB), sum); in arm_dot_prod_q15() 129 sum += (q63_t)((q31_t) *pSrcA++ * *pSrcB++); in arm_dot_prod_q15() 130 sum += (q63_t)((q31_t) *pSrcA++ * *pSrcB++); in arm_dot_prod_q15() 131 sum += (q63_t)((q31_t) *pSrcA++ * *pSrcB++); in arm_dot_prod_q15() [all …]
|
D | arm_mult_f32.c | 65 const float32_t * pSrcA, in arm_mult_f32() argument 83 vec1 = vld1q(pSrcA); in arm_mult_f32() 89 pSrcA += 4; in arm_mult_f32() 103 vec1 = vld1q(pSrcA); in arm_mult_f32() 112 const float32_t * pSrcA, in arm_mult_f32() argument 132 vec1 = vld1q_f32(pSrcA); in arm_mult_f32() 138 pSrcA += 4; in arm_mult_f32() 160 *pDst++ = (*pSrcA++) * (*pSrcB++); in arm_mult_f32() 162 *pDst++ = (*pSrcA++) * (*pSrcB++); in arm_mult_f32() 164 *pDst++ = (*pSrcA++) * (*pSrcB++); in arm_mult_f32() [all …]
|
D | arm_sub_f32.c | 65 const float32_t * pSrcA, in arm_sub_f32() argument 84 vec1 = vld1q(pSrcA); in arm_sub_f32() 90 pSrcA += 4; in arm_sub_f32() 105 vec1 = vld1q(pSrcA); in arm_sub_f32() 114 const float32_t * pSrcA, in arm_sub_f32() argument 134 vec1 = vld1q_f32(pSrcA); in arm_sub_f32() 140 pSrcA += 4; in arm_sub_f32() 162 *pDst++ = (*pSrcA++) - (*pSrcB++); in arm_sub_f32() 164 *pDst++ = (*pSrcA++) - (*pSrcB++); in arm_sub_f32() 166 *pDst++ = (*pSrcA++) - (*pSrcB++); in arm_sub_f32() [all …]
|
D | arm_add_f32.c | 65 const float32_t * pSrcA, in arm_add_f32() argument 84 vec1 = vld1q(pSrcA); in arm_add_f32() 90 pSrcA += 4; in arm_add_f32() 105 vec1 = vld1q(pSrcA); in arm_add_f32() 114 const float32_t * pSrcA, in arm_add_f32() argument 134 vec1 = vld1q_f32(pSrcA); in arm_add_f32() 140 pSrcA += 4; in arm_add_f32() 162 *pDst++ = (*pSrcA++) + (*pSrcB++); in arm_add_f32() 163 *pDst++ = (*pSrcA++) + (*pSrcB++); in arm_add_f32() 164 *pDst++ = (*pSrcA++) + (*pSrcB++); in arm_add_f32() [all …]
|