Home
last modified time | relevance | path

Searched refs:vecInB (Results 1 – 5 of 5) sorted by relevance

/cmsis-dsp-latest/dsppp/Include/dsppp/Helium/
Dmatrix_multiply_f32.hpp43 VEC vecInA, vecInB, vecDst;
63 vecInB = vldrwq_gather_shifted_offset(pSrcB.const_ptr(), vecOffsB);
67 vecInB = inner::vload1_gen_stride<0, 1, 0, 1>::run(pSrcB.const_ptr());
69 vecDst = inner::vmul(vecInA, vecInB);
94 vecInB = vldrwq_gather_shifted_offset(pSrcB.const_ptr(), vecOffsB);
99 vecInB = inner::vload1_gen_stride<s, s+1, s, s+1>::run(pSrcB.const_ptr());
107 vecDst = inner::vmacc(vecDst, vecInA, vecInB);
141 VEC vecInB;
155 vecInB = inner::vload1_z<1>(pSrBVec, MATRIX_DIM3,p0);
158 vecMac0 = inner::vmul(vecInB, *pInA0++);
[all …]
Dmatrix_multiply_f16.hpp52 VEC vecInA, vecInB, vecDst;
67 vecInB = vldrhq_gather_shifted_offset(pSrcB.const_ptr(), vecOffsB);
74 vecDst = vmulq(vecInA, vecInB);
90 vecInB = vldrhq_gather_shifted_offset(pSrcB.const_ptr(), vecOffsB);
97 vecDst = vfmaq(vecDst, vecInA, vecInB);
135 float16x8_t vecInA, vecInB, vecDst;
151 vecInB = vldrhq_gather_shifted_offset(pSrcB.const_ptr(), vecOffsB);
157 vecDst = vmulq(vecInA, vecInB);
174 vecInB = vldrhq_gather_shifted_offset(pSrcB.const_ptr(), vecOffsB);
180 vecDst = vfmaq(vecDst, vecInA, vecInB);
[all …]
Dmatrix_multiply.hpp127 VEC vecInB;
157 vecInB = inner::vload1<1>(pInB0); /* vldrwq_f32(pInB0, 0); */
159 vecMac0 = inner::vmacc(vecMac0, vecInB, *pInA0++);
160 vecMac1 = inner::vmacc(vecMac1, vecInB, *pInA1++);
161 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++);
162 vecMac3 = inner::vmacc(vecMac3, vecInB, *pInA3++);
210 vecInB = inner::vload1_z<1>(pInB0, colBLeft,p0);
212 vecMac0 = inner::vmacc(vecMac0, vecInB, *pInA0++);
213 vecMac1 = inner::vmacc(vecMac1, vecInB, *pInA1++);
214 vecMac2 = inner::vmacc(vecMac2, vecInB, *pInA2++);
[all …]
/cmsis-dsp-latest/Source/MatrixFunctions/
Darm_mat_mult_f16.c64 f16x8_t vecInA, vecInB, vecDst; in arm_mat_mult_f16_2x2_mve() local
79 vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB); in arm_mat_mult_f16_2x2_mve()
86 vecDst = vmulq(vecInA, vecInB); in arm_mat_mult_f16_2x2_mve()
102 vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB); in arm_mat_mult_f16_2x2_mve()
109 vecDst = vfmaq(vecDst, vecInA, vecInB); in arm_mat_mult_f16_2x2_mve()
133 f16x8_t vecInA, vecInB, vecDst; in arm_mat_mult_f16_3x3_mve() local
149 vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB); in arm_mat_mult_f16_3x3_mve()
155 vecDst = vmulq(vecInA, vecInB); in arm_mat_mult_f16_3x3_mve()
172 vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB); in arm_mat_mult_f16_3x3_mve()
178 vecDst = vfmaq(vecDst, vecInA, vecInB); in arm_mat_mult_f16_3x3_mve()
[all …]
Darm_mat_mult_f32.c103 f32x4_t vecInA, vecInB, vecDst; in arm_mat_mult_f32_2x2_mve() local
109 vecInB = vldrwq_gather_shifted_offset((float32_t const *) pSrcB->pData, vecOffsB); in arm_mat_mult_f32_2x2_mve()
111 vecDst = vmulq(vecInA, vecInB); in arm_mat_mult_f32_2x2_mve()
117 vecInB = vldrwq_gather_shifted_offset((float32_t const *) pSrcB->pData, vecOffsB); in arm_mat_mult_f32_2x2_mve()
119 vecDst = vfmaq(vecDst, vecInA, vecInB); in arm_mat_mult_f32_2x2_mve()
150 f32x4_t vecInB; in arm_mat_mult_f32_3x3_mve() local
164 vecInB = vldrwq_z_f32(pSrBVec, p0); in arm_mat_mult_f32_3x3_mve()
167 vecMac0 = vmulq(vecInB, *pInA0++); in arm_mat_mult_f32_3x3_mve()
168 vecMac1 = vmulq(vecInB, *pInA1++); in arm_mat_mult_f32_3x3_mve()
169 vecMac2 = vmulq(vecInB, *pInA2++); in arm_mat_mult_f32_3x3_mve()
[all …]