/cmsis-dsp-latest/Source/StatisticsFunctions/ |
D | arm_mse_q31.c | 56 q31x4_t vecSrcA,vecSrcB; in arm_mse_q31() local 63 vecSrcA = vld1q(pSrcA); in arm_mse_q31() 66 vecSrcA = vshrq(vecSrcA,1); in arm_mse_q31() 70 vecSrcA = vqsubq(vecSrcA,vecSrcB); in arm_mse_q31() 74 sum = vrmlaldavhaq(sum, vecSrcA, vecSrcA); in arm_mse_q31() 88 vecSrcA = vld1q(pSrcA); in arm_mse_q31() 91 vecSrcA = vshrq(vecSrcA,1); in arm_mse_q31() 94 vecSrcA = vqsubq(vecSrcA,vecSrcB); in arm_mse_q31() 96 sum = vrmlaldavhaq_p(sum, vecSrcA, vecSrcA, p0); in arm_mse_q31()
|
D | arm_mse_q7.c | 62 q7x16_t vecSrcA,vecSrcB; in arm_mse_q7() local 69 vecSrcA = vld1q(pSrcA); in arm_mse_q7() 72 vecSrcA = vshrq(vecSrcA,1); in arm_mse_q7() 75 vecSrcA = vqsubq(vecSrcA,vecSrcB); in arm_mse_q7() 79 sum = vmladavaq(sum, vecSrcA, vecSrcA); in arm_mse_q7() 93 vecSrcA = vld1q(pSrcA); in arm_mse_q7() 96 vecSrcA = vshrq(vecSrcA,1); in arm_mse_q7() 99 vecSrcA = vqsubq(vecSrcA,vecSrcB); in arm_mse_q7() 101 sum = vmladavaq_p(sum, vecSrcA, vecSrcA, p0); in arm_mse_q7()
|
D | arm_mse_q15.c | 56 q15x8_t vecSrcA,vecSrcB; in arm_mse_q15() local 62 vecSrcA = vld1q(pSrcA); in arm_mse_q15() 65 vecSrcA = vshrq(vecSrcA,1); in arm_mse_q15() 68 vecSrcA = vqsubq(vecSrcA,vecSrcB); in arm_mse_q15() 72 sum = vmlaldavaq(sum, vecSrcA, vecSrcA); in arm_mse_q15() 86 vecSrcA = vld1q(pSrcA); in arm_mse_q15() 89 vecSrcA = vshrq(vecSrcA,1); in arm_mse_q15() 92 vecSrcA = vqsubq(vecSrcA,vecSrcB); in arm_mse_q15() 94 sum = vmlaldavaq_p(sum, vecSrcA, vecSrcA, p0); in arm_mse_q15()
|
/cmsis-dsp-latest/Source/ComplexMathFunctions/ |
D | arm_cmplx_dot_prod_q15.c | 67 q15x8_t vecSrcA, vecSrcB; in arm_cmplx_dot_prod_q15() local 74 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 81 accReal = vmlsldavaq(accReal, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q15() 85 accImag = vmlaldavaxq(accImag, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q15() 90 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 103 accReal = vmlsldavaq(accReal, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q15() 106 accImag = vmlaldavaxq(accImag, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q15() 110 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q15() 125 vecSrcA = vldrhq_z_s16(pSrcA, p); in arm_cmplx_dot_prod_q15() 128 accReal = vmlsldavaq_p(accReal, vecSrcA, vecSrcB, p); in arm_cmplx_dot_prod_q15() [all …]
|
D | arm_cmplx_dot_prod_q31.c | 69 q31x4_t vecSrcA, vecSrcB; in arm_cmplx_dot_prod_q31() local 76 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 83 accReal = vrmlsldavhaq(accReal, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q31() 87 accImag = vrmlaldavhaxq(accImag, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q31() 92 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 105 accReal = vrmlsldavhaq(accReal, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q31() 108 accImag = vrmlaldavhaxq(accImag, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_q31() 112 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_q31() 127 vecSrcA = vldrwq_z_s32(pSrcA, p); in arm_cmplx_dot_prod_q31() 130 accReal = vrmlsldavhaq_p(accReal, vecSrcA, vecSrcB, p); in arm_cmplx_dot_prod_q31() [all …]
|
D | arm_cmplx_mult_cmplx_f16.c | 61 f16x8_t vecSrcA, vecSrcB; in arm_cmplx_mult_cmplx_f16() local 69 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f16() 75 vec_acc = vcmulq(vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f16() 79 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f16() 86 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f16() 101 vec_acc = vcmulq(vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f16() 104 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f16() 123 vecSrcA = vldrhq_z_f16(pSrcA, p); in arm_cmplx_mult_cmplx_f16() 125 vec_acc = vcmulq_m(vuninitializedq_f16(),vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_f16() 126 vec_acc = vcmlaq_rot90_m(vec_acc, vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_f16() [all …]
|
D | arm_cmplx_dot_prod_f16.c | 66 f16x8_t vecSrcA, vecSrcB; in arm_cmplx_dot_prod_f16() local 74 vecSrcA = vld1q( pSrcA); in arm_cmplx_dot_prod_f16() 80 vec_acc = vcmlaq(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f16() 84 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f16() 89 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_f16() 102 vec_acc = vcmlaq(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f16() 105 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f16() 120 vecSrcA = vldrhq_z_f16(pSrcA, p); in arm_cmplx_dot_prod_f16() 122 vec_acc = vcmlaq_m(vec_acc, vecSrcA, vecSrcB, p); in arm_cmplx_dot_prod_f16() 123 vec_acc = vcmlaq_rot90_m(vec_acc, vecSrcA, vecSrcB, p); in arm_cmplx_dot_prod_f16() [all …]
|
D | arm_cmplx_mult_cmplx_f32.c | 79 f32x4_t vecSrcA, vecSrcB; in arm_cmplx_mult_cmplx_f32() local 87 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f32() 93 vec_acc = vcmulq(vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f32() 97 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f32() 104 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_f32() 119 vec_acc = vcmulq(vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f32() 122 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_f32() 141 vecSrcA = vldrwq_z_f32(pSrcA, p); in arm_cmplx_mult_cmplx_f32() 143 vec_acc = vcmulq_m(vuninitializedq_f32(),vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_f32() 144 vec_acc = vcmlaq_rot90_m(vec_acc, vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_f32() [all …]
|
D | arm_cmplx_mult_cmplx_q31.c | 60 q31x4_t vecSrcA, vecSrcB; in arm_cmplx_mult_cmplx_q31() local 68 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q31() 76 vecDst = vqdmlsdhq(vuninitializedq_s32(), vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q31() 81 vecDst = vqdmladhxq(vecDst, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q31() 89 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q31() 106 vecDst = vqdmlsdhq(vuninitializedq_s32(), vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q31() 109 vecDst = vqdmladhxq(vecDst, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q31() 131 vecSrcA = vldrwq_z_s32(pSrcA, p); in arm_cmplx_mult_cmplx_q31() 134 vecDst = vqdmlsdhq_m(vuninitializedq_s32(), vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_q31() 135 vecDst = vqdmladhxq_m(vecDst, vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_q31() [all …]
|
D | arm_cmplx_mult_cmplx_q15.c | 60 q15x8_t vecSrcA, vecSrcB; in arm_cmplx_mult_cmplx_q15() local 69 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q15() 78 vecDst = vqdmlsdhq(vuninitializedq_s16(), vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q15() 83 vecDst = vqdmladhxq(vecDst, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q15() 91 vecSrcA = vld1q(pSrcA); in arm_cmplx_mult_cmplx_q15() 108 vecDst = vqdmlsdhq(vuninitializedq_s16(), vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q15() 111 vecDst = vqdmladhxq(vecDst, vecSrcA, vecSrcB); in arm_cmplx_mult_cmplx_q15() 134 vecSrcA = vldrhq_z_s16(pSrcA, p); in arm_cmplx_mult_cmplx_q15() 137 vecDst = vqdmlsdhq_m(vuninitializedq_s16(), vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_q15() 138 vecDst = vqdmladhxq_m(vecDst, vecSrcA, vecSrcB, p); in arm_cmplx_mult_cmplx_q15() [all …]
|
D | arm_cmplx_dot_prod_f32.c | 87 f32x4_t vecSrcA, vecSrcB; in arm_cmplx_dot_prod_f32() local 95 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_f32() 101 vec_acc = vcmlaq(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f32() 105 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f32() 110 vecSrcA = vld1q(pSrcA); in arm_cmplx_dot_prod_f32() 123 vec_acc = vcmlaq(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f32() 126 vec_acc = vcmlaq_rot90(vec_acc, vecSrcA, vecSrcB); in arm_cmplx_dot_prod_f32() 140 vecSrcA = vldrwq_z_f32(pSrcA, p); in arm_cmplx_dot_prod_f32() 142 vec_acc = vcmlaq_m(vec_acc, vecSrcA, vecSrcB, p); in arm_cmplx_dot_prod_f32() 143 vec_acc = vcmlaq_rot90_m(vec_acc, vecSrcA, vecSrcB, p); in arm_cmplx_dot_prod_f32() [all …]
|
/cmsis-dsp-latest/Source/BasicMathFunctions/ |
D | arm_and_u16.c | 65 uint16x8_t vecSrcA, vecSrcB; in arm_and_u16() local 72 vecSrcA = vld1q(pSrcA); in arm_and_u16() 75 vst1q(pDst, vandq_u16(vecSrcA, vecSrcB) ); in arm_and_u16() 91 vecSrcA = vld1q(pSrcA); in arm_and_u16() 93 vstrhq_p(pDst, vandq_u16(vecSrcA, vecSrcB), p0); in arm_and_u16()
|
D | arm_and_u32.c | 57 uint32x4_t vecSrcA, vecSrcB; in arm_and_u32() local 64 vecSrcA = vld1q(pSrcA); in arm_and_u32() 67 vst1q(pDst, vandq_u32(vecSrcA, vecSrcB) ); in arm_and_u32() 83 vecSrcA = vld1q(pSrcA); in arm_and_u32() 85 vstrwq_p(pDst, vandq_u32(vecSrcA, vecSrcB), p0); in arm_and_u32()
|
D | arm_and_u8.c | 58 uint8x16_t vecSrcA, vecSrcB; in arm_and_u8() local 65 vecSrcA = vld1q(pSrcA); in arm_and_u8() 68 vst1q(pDst, vandq_u8(vecSrcA, vecSrcB) ); in arm_and_u8() 84 vecSrcA = vld1q(pSrcA); in arm_and_u8() 86 vstrbq_p(pDst, vandq_u8(vecSrcA, vecSrcB), p0); in arm_and_u8()
|
D | arm_xor_u16.c | 65 uint16x8_t vecSrcA, vecSrcB; in arm_xor_u16() local 72 vecSrcA = vld1q(pSrcA); in arm_xor_u16() 75 vst1q(pDst, veorq_u16(vecSrcA, vecSrcB) ); in arm_xor_u16() 91 vecSrcA = vld1q(pSrcA); in arm_xor_u16() 93 vstrhq_p(pDst, veorq_u16(vecSrcA, vecSrcB), p0); in arm_xor_u16()
|
D | arm_xor_u32.c | 57 uint32x4_t vecSrcA, vecSrcB; in arm_xor_u32() local 64 vecSrcA = vld1q(pSrcA); in arm_xor_u32() 67 vst1q(pDst, veorq_u32(vecSrcA, vecSrcB) ); in arm_xor_u32() 83 vecSrcA = vld1q(pSrcA); in arm_xor_u32() 85 vstrwq_p(pDst, veorq_u32(vecSrcA, vecSrcB), p0); in arm_xor_u32()
|
D | arm_xor_u8.c | 57 uint8x16_t vecSrcA, vecSrcB; in arm_xor_u8() local 64 vecSrcA = vld1q(pSrcA); in arm_xor_u8() 67 vst1q(pDst, veorq_u8(vecSrcA, vecSrcB) ); in arm_xor_u8() 83 vecSrcA = vld1q(pSrcA); in arm_xor_u8() 85 vstrbq_p(pDst, veorq_u8(vecSrcA, vecSrcB), p0); in arm_xor_u8()
|
D | arm_or_u16.c | 65 uint16x8_t vecSrcA, vecSrcB; in arm_or_u16() local 72 vecSrcA = vld1q(pSrcA); in arm_or_u16() 75 vst1q(pDst, vorrq_u16(vecSrcA, vecSrcB) ); in arm_or_u16() 91 vecSrcA = vld1q(pSrcA); in arm_or_u16() 93 vstrhq_p(pDst, vorrq_u16(vecSrcA, vecSrcB), p0); in arm_or_u16()
|
D | arm_or_u32.c | 57 uint32x4_t vecSrcA, vecSrcB; in arm_or_u32() local 64 vecSrcA = vld1q(pSrcA); in arm_or_u32() 67 vst1q(pDst, vorrq_u32(vecSrcA, vecSrcB) ); in arm_or_u32() 83 vecSrcA = vld1q(pSrcA); in arm_or_u32() 85 vstrwq_p(pDst, vorrq_u32(vecSrcA, vecSrcB), p0); in arm_or_u32()
|
D | arm_or_u8.c | 57 uint8x16_t vecSrcA, vecSrcB; in arm_or_u8() local 64 vecSrcA = vld1q(pSrcA); in arm_or_u8() 67 vst1q(pDst, vorrq_u8(vecSrcA, vecSrcB) ); in arm_or_u8() 83 vecSrcA = vld1q(pSrcA); in arm_or_u8() 85 vstrbq_p(pDst, vorrq_u8(vecSrcA, vecSrcB), p0); in arm_or_u8()
|