Home
last modified time | relevance | path

Searched refs:vecB (Results 1 – 25 of 56) sorted by relevance

123

/cmsis-dsp-latest/Source/MatrixFunctions/
Darm_mat_cmplx_mult_f32.c83 f32x4_t vecB, vecA; in arm_mat_cmplx_mult_f32_2x2_mve() local
93 vecB = vldrwq_gather_shifted_offset(pInB, vecColBOffs0); 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()
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()
114 vecB = vldrwq_gather_shifted_offset(pInB, vecColBOffs0); 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()
121 acc1 = vcmulq(vecA, vecB); in arm_mat_cmplx_mult_f32_2x2_mve()
[all …]
Darm_mat_cmplx_mult_f16.c77 f16x8_t vecB, vecA0, vecA1; in arm_mat_cmplx_mult_f16_2x2_mve() local
102 vecB = vldrhq_gather_shifted_offset(pInB, vecColBOffs0); in arm_mat_cmplx_mult_f16_2x2_mve()
104 acc0 = vcmulq(vecA0, vecB); in arm_mat_cmplx_mult_f16_2x2_mve()
105 acc0 = vcmlaq_rot90(acc0, vecA0, vecB); in arm_mat_cmplx_mult_f16_2x2_mve()
107 acc1 = vcmulq(vecA1, vecB); in arm_mat_cmplx_mult_f16_2x2_mve()
108 acc1 = vcmlaq_rot90(acc1, vecA1, vecB); in arm_mat_cmplx_mult_f16_2x2_mve()
153 f16x8_t vecB, vecA0, vecA1, vecA2; in arm_mat_cmplx_mult_f16_3x3_mve() local
172 vecB = vldrhq_gather_shifted_offset_z(pInB, vecColBOffs0, p0); in arm_mat_cmplx_mult_f16_3x3_mve()
174 acc0 = vcmulq(vecA0, vecB); in arm_mat_cmplx_mult_f16_3x3_mve()
175 acc0 = vcmlaq_rot90(acc0, vecA0, vecB); in arm_mat_cmplx_mult_f16_3x3_mve()
[all …]
Darm_mat_mult_q7.c72 q7x16_t vecB, vecA0, vecA1; in arm_mat_mult_q7_2x2_mve() local
79 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_2x2_mve()
84 acc0 = vmladavq_s8(vecA0, vecB); in arm_mat_mult_q7_2x2_mve()
85 acc1 = vmladavq_s8(vecA1, vecB); in arm_mat_mult_q7_2x2_mve()
94 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_2x2_mve()
96 acc0 = vmladavq_s8(vecA0, vecB); in arm_mat_mult_q7_2x2_mve()
97 acc1 = vmladavq_s8(vecA1, vecB); in arm_mat_mult_q7_2x2_mve()
122 q7x16_t vecB, vecA0, vecA1, vecA2; in arm_mat_mult_q7_3x3_mve() local
130 vecB = vldrbq_gather_offset_z(pInB, vecColBOffs, p0); in arm_mat_mult_q7_3x3_mve()
136 acc0 = vmladavq_s8(vecA0, vecB); in arm_mat_mult_q7_3x3_mve()
[all …]
Darm_mat_mult_q31.c79 q31x4_t vecB, vecA0, vecA1; in arm_mat_mult_q31_2x2_mve() local
89 vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0); in arm_mat_mult_q31_2x2_mve()
95 acc0 = vrmlaldavhq(vecA0, vecB); in arm_mat_mult_q31_2x2_mve()
96 acc1 = vrmlaldavhq(vecA1, vecB); in arm_mat_mult_q31_2x2_mve()
108 vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0); in arm_mat_mult_q31_2x2_mve()
110 acc0 = vrmlaldavhq(vecA0, vecB); in arm_mat_mult_q31_2x2_mve()
111 acc1 = vrmlaldavhq(vecA1, vecB); in arm_mat_mult_q31_2x2_mve()
139 q31x4_t vecB, vecA; in arm_mat_mult_q31_3x3_mve() local
148 vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0); in arm_mat_mult_q31_3x3_mve()
151 acc0 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_q31_3x3_mve()
[all …]
Darm_mat_cmplx_mult_q31.c79 q31x4_t vecB, vecA; in arm_mat_cmplx_mult_q31_2x2_mve() local
90 vecB = vldrwq_gather_shifted_offset(pInB, vecColBOffs0); 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()
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()
108 vecB = vldrwq_gather_shifted_offset(pInB, vecColBOffs0); 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()
114 acc2 = vmlsldavq_s32(vecA, vecB); in arm_mat_cmplx_mult_q31_2x2_mve()
[all …]
Darm_mat_mult_opt_q31.c83 q31x4_t vecB, vecA0, vecA1; in arm_mat_mult_opt_q31_2x2_mve() local
93 vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0); in arm_mat_mult_opt_q31_2x2_mve()
99 acc0 = vrmlaldavhq(vecA0, vecB); in arm_mat_mult_opt_q31_2x2_mve()
100 acc1 = vrmlaldavhq(vecA1, vecB); in arm_mat_mult_opt_q31_2x2_mve()
112 vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0); in arm_mat_mult_opt_q31_2x2_mve()
114 acc0 = vrmlaldavhq(vecA0, vecB); in arm_mat_mult_opt_q31_2x2_mve()
115 acc1 = vrmlaldavhq(vecA1, vecB); in arm_mat_mult_opt_q31_2x2_mve()
143 q31x4_t vecB, vecA; in arm_mat_mult_opt_q31_3x3_mve() local
152 vecB = vldrwq_gather_shifted_offset_z_s32(pInB, vecColBOffs, p0); in arm_mat_mult_opt_q31_3x3_mve()
155 acc0 = vrmlaldavhq(vecA, vecB); in arm_mat_mult_opt_q31_3x3_mve()
[all …]
Darm_mat_mult_q15.c83 q15x8_t vecB, vecA0, vecA1; in arm_mat_mult_q15_2x2_mve() local
90 vecB = vldrhq_gather_shifted_offset_z_s16((q15_t const *)pInB, vecColBOffs, p0); in arm_mat_mult_q15_2x2_mve()
95 acc0 = vmlaldavq(vecA0, vecB); in arm_mat_mult_q15_2x2_mve()
96 acc1 = vmlaldavq(vecA1, vecB); in arm_mat_mult_q15_2x2_mve()
108 vecB = vldrhq_gather_shifted_offset_z_s16(pInB, vecColBOffs, p0); in arm_mat_mult_q15_2x2_mve()
110 acc0 = vmlaldavq(vecA0, vecB); in arm_mat_mult_q15_2x2_mve()
111 acc1 = vmlaldavq(vecA1, vecB); in arm_mat_mult_q15_2x2_mve()
140 q15x8_t vecB, vecA0, vecA1, vecA2; in arm_mat_mult_q15_3x3_mve() local
148 vecB = vldrhq_gather_shifted_offset_z_s16((q15_t const *)pInB, vecColBOffs, p0); in arm_mat_mult_q15_3x3_mve()
154 acc0 = vmlaldavq(vecA0, vecB); in arm_mat_mult_q15_3x3_mve()
[all …]
Darm_mat_cmplx_mult_q15.c136 q15x8_t vecA, vecB, vecB2; in arm_mat_cmplx_mult_q15() local
172 vecB = vldrhq_gather_shifted_offset(pInB, vecOffs); 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()
195 vecB = vldrhq_gather_shifted_offset(pInB, vecOffs); 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()
230 q15x8_t vecA, vecB; in arm_mat_cmplx_mult_q15() local
265 vecB = vldrhq_gather_shifted_offset(pInB, vecOffs); in arm_mat_cmplx_mult_q15()
267 acc0 = vmlsldavaq_s16(acc0, vecA, vecB); in arm_mat_cmplx_mult_q15()
[all …]
/cmsis-dsp-latest/dsppp/Include/dsppp/Helium/
Dmatrix_multiply_fixed.hpp43 VEC vecB, vecA0, vecA1; in arm_mat_mult_2x2_mve() local
50 vecB = inner::vload1_z<StaticStride<MB>::value>(pInB,MATRIX_DIM2,p0); in arm_mat_mult_2x2_mve()
54 vecB = inner::vload1_z(pInB,pSrcB.stride(),MATRIX_DIM2,p0); in arm_mat_mult_2x2_mve()
61 acc0 = inner::vmacc(vecA0, vecB,p0); in arm_mat_mult_2x2_mve()
62 acc1 = inner::vmacc(vecA1, vecB,p0); in arm_mat_mult_2x2_mve()
73 vecB = inner::vload1_z<StaticStride<MB>::value>(pInB,MATRIX_DIM2,p0); in arm_mat_mult_2x2_mve()
77 vecB = inner::vload1_z(pInB,pSrcB.stride(),MATRIX_DIM2,p0); in arm_mat_mult_2x2_mve()
80 acc0 = inner::vmacc(vecA0, vecB,p0); in arm_mat_mult_2x2_mve()
81 acc1 = inner::vmacc(vecA1, vecB,p0); in arm_mat_mult_2x2_mve()
111 VEC vecB, vecA0, vecA1, vecA2; in arm_mat_mult_3x3_mve() local
[all …]
/cmsis-dsp-latest/Source/StatisticsFunctions/
Darm_mse_f32.c60 float32x4_t vecA, vecB; in arm_mse_f32() local
73 vecB = vld1q(pSrcB); in arm_mse_f32()
76 vecA = vsubq(vecA, vecB); in arm_mse_f32()
91 vecB = vld1q(pSrcB); in arm_mse_f32()
93 vecA = vsubq(vecA, vecB); in arm_mse_f32()
114 float32x4_t vecA, vecB; in arm_mse_f32() local
131 vecB = vld1q_f32(pSrcB); in arm_mse_f32()
134 vecA = vsubq_f32(vecA, vecB); in arm_mse_f32()
/cmsis-dsp-latest/Source/BasicMathFunctions/
Darm_add_q31.c64 q31x4_t vecB; in arm_add_q31() local
75 vecB = vld1q(pSrcB); in arm_add_q31()
76 vst1q(pDst, vqaddq(vecA, vecB)); in arm_add_q31()
96 vecB = vld1q(pSrcB); in arm_add_q31()
97 vstrwq_p(pDst, vqaddq(vecA, vecB), p0); in arm_add_q31()
Darm_sub_q31.c64 q31x4_t vecB; in arm_sub_q31() local
75 vecB = vld1q(pSrcB); in arm_sub_q31()
76 vst1q(pDst, vqsubq(vecA, vecB)); in arm_sub_q31()
96 vecB = vld1q(pSrcB); in arm_sub_q31()
97 vstrwq_p(pDst, vqsubq(vecA, vecB), p0); in arm_sub_q31()
Darm_dot_prod_q31.c68 q31x4_t vecB; in arm_dot_prod_q31() local
80 vecB = vld1q(pSrcB); in arm_dot_prod_q31()
81 sum = vrmlaldavhaq(sum, vecA, vecB); in arm_dot_prod_q31()
100 vecB = vld1q(pSrcB); in arm_dot_prod_q31()
101 sum = vrmlaldavhaq_p(sum, vecA, vecB, p0); in arm_dot_prod_q31()
Darm_add_q7.c65 q7x16_t vecB; in arm_add_q7() local
76 vecB = vld1q(pSrcB); in arm_add_q7()
77 vst1q(pDst, vqaddq(vecA, vecB)); in arm_add_q7()
97 vecB = vld1q(pSrcB); in arm_add_q7()
98 vstrbq_p(pDst, vqaddq(vecA, vecB), p0); in arm_add_q7()
Darm_dot_prod_f16.c60 f16x8_t vecA, vecB; in arm_dot_prod_f16() local
78 vecB = vld1q(pSrcB); in arm_dot_prod_f16()
81 vecSum = vfmaq(vecSum, vecA, vecB); in arm_dot_prod_f16()
96 vecB = vld1q(pSrcB); in arm_dot_prod_f16()
97 vecSum = vfmaq_m(vecSum, vecA, vecB, p0); in arm_dot_prod_f16()
Darm_sub_q7.c63 q7x16_t vecB; in arm_sub_q7() local
74 vecB = vld1q(pSrcB); in arm_sub_q7()
75 vst1q(pDst, vqsubq(vecA, vecB)); in arm_sub_q7()
95 vecB = vld1q(pSrcB); in arm_sub_q7()
96 vstrbq_p(pDst, vqsubq(vecA, vecB), p0); in arm_sub_q7()
Darm_dot_prod_q15.c66 q15x8_t vecB; in arm_dot_prod_q15() local
78 vecB = vld1q(pSrcB); in arm_dot_prod_q15()
79 sum = vmlaldavaq(sum, vecA, vecB); in arm_dot_prod_q15()
98 vecB = vld1q(pSrcB); in arm_dot_prod_q15()
99 sum = vmlaldavaq_p(sum, vecA, vecB, p0); in arm_dot_prod_q15()
Darm_mult_q31.c62 q31x4_t vecA, vecB; in arm_mult_q31() local
73 vecB = vld1q(pSrcB); in arm_mult_q31()
74 vst1q(pDst, vqdmulhq(vecA, vecB)); in arm_mult_q31()
94 vecB = vld1q(pSrcB); in arm_mult_q31()
95 vstrwq_p(pDst, vqdmulhq(vecA, vecB), p0); in arm_mult_q31()
Darm_add_q15.c64 q15x8_t vecB; in arm_add_q15() local
75 vecB = vld1q(pSrcB); in arm_add_q15()
76 vst1q(pDst, vqaddq(vecA, vecB)); in arm_add_q15()
96 vecB = vld1q(pSrcB); in arm_add_q15()
97 vstrhq_p(pDst, vqaddq(vecA, vecB), p0); in arm_add_q15()
Darm_mult_q7.c62 q7x16_t vecA, vecB; in arm_mult_q7() local
73 vecB = vld1q(pSrcB); in arm_mult_q7()
74 vst1q(pDst, vqdmulhq(vecA, vecB)); in arm_mult_q7()
94 vecB = vld1q(pSrcB); in arm_mult_q7()
95 vstrbq_p(pDst, vqdmulhq(vecA, vecB), p0); in arm_mult_q7()
Darm_sub_q15.c64 q15x8_t vecB; in arm_sub_q15() local
75 vecB = vld1q(pSrcB); in arm_sub_q15()
76 vst1q(pDst, vqsubq(vecA, vecB)); in arm_sub_q15()
96 vecB = vld1q(pSrcB); in arm_sub_q15()
97 vstrhq_p(pDst, vqsubq(vecA, vecB), p0); in arm_sub_q15()
Darm_dot_prod_q7.c67 q7x16_t vecB; in arm_dot_prod_q7() local
79 vecB = vld1q(pSrcB); in arm_dot_prod_q7()
80 sum = vmladavaq(sum, vecA, vecB); in arm_dot_prod_q7()
99 vecB = vld1q(pSrcB); in arm_dot_prod_q7()
100 sum = vmladavaq_p(sum, vecA, vecB, p0); in arm_dot_prod_q7()
/cmsis-dsp-latest/Source/DistanceFunctions/
Darm_chebyshev_distance_f16.c70 f16x8_t vecA, vecB; in arm_chebyshev_distance_f16() local
79 vecB = vld1q(pB); in arm_chebyshev_distance_f16()
84 vecDiff = vmaxnmaq(vsubq(vecA, vecB), vecDiff); in arm_chebyshev_distance_f16()
99 vecB = vldrhq_z_f16(pB, p0); in arm_chebyshev_distance_f16()
105 vecDiff = vmaxnmaq_m(vecDiff, vsubq(vecA, vecB), p0); in arm_chebyshev_distance_f16()
/cmsis-dsp-latest/Source/ComplexMathFunctions/
Darm_cmplx_mag_squared_f32.c135 float32x4x2_t vecB; in arm_cmplx_mag_squared_f32() local
154 vecB = vld2q_f32(pSrc); in arm_cmplx_mag_squared_f32()
157 vRealB = vmulq_f32(vecB.val[0], vecB.val[0]); in arm_cmplx_mag_squared_f32()
158 vImagB = vmulq_f32(vecB.val[1], vecB.val[1]); in arm_cmplx_mag_squared_f32()
/cmsis-dsp-latest/Source/TransformFunctions/
Darm_cfft_f16.c95 f16x8_t vecA, vecB, vecC, vecD; in _arm_radix4_butterfly_f16_mve() local
140 vecB = vldrhq_f16(inB); in _arm_radix4_butterfly_f16_mve()
146 vecSum1 = vecB + vecD; in _arm_radix4_butterfly_f16_mve()
147 vecDiff1 = vecB - vecD; in _arm_radix4_butterfly_f16_mve()
223 vecB = (f16x8_t)vldrwq_gather_base_f32(vecScGathAddr, 4); in _arm_radix4_butterfly_f16_mve()
226 vecSum1 = vecB + vecD; in _arm_radix4_butterfly_f16_mve()
227 vecDiff1 = vecB - vecD; in _arm_radix4_butterfly_f16_mve()
301 f16x8_t vecA, vecB, vecC, vecD; in _arm_radix4_butterfly_inverse_f16_mve() local
348 vecB = vldrhq_f16(inB); in _arm_radix4_butterfly_inverse_f16_mve()
354 vecSum1 = vecB + vecD; in _arm_radix4_butterfly_inverse_f16_mve()
[all …]

123