45#ifndef INCLUDED_volk_32fc_x2_dot_prod_32fc_u_H
46#define INCLUDED_volk_32fc_x2_dot_prod_32fc_u_H
55extern void volk_32fc_x2_dot_prod_32fc_sifive_u74(
lv_32fc_t* result,
58 unsigned int num_points);
67 unsigned int num_points)
70 float* res = (
float*)result;
71 float* in = (
float*)input;
72 float* tp = (
float*)taps;
73 unsigned int n_2_ccomplex_blocks = num_points / 2;
75 float sum0[2] = { 0, 0 };
76 float sum1[2] = { 0, 0 };
79 for (
i = 0;
i < n_2_ccomplex_blocks; ++
i) {
80 sum0[0] += in[0] * tp[0] - in[1] * tp[1];
81 sum0[1] += in[0] * tp[1] + in[1] * tp[0];
82 sum1[0] += in[2] * tp[2] - in[3] * tp[3];
83 sum1[1] += in[2] * tp[3] + in[3] * tp[2];
89 res[0] = sum0[0] + sum1[0];
90 res[1] = sum0[1] + sum1[1];
94 *result += input[num_points - 1] * taps[num_points - 1];
101#if LV_HAVE_SSE && LV_HAVE_64
103static inline void volk_32fc_x2_dot_prod_32fc_u_sse_64(
lv_32fc_t* result,
106 unsigned int num_points)
109 const unsigned int num_bytes = num_points * 8;
110 unsigned int isodd = num_points & 1;
113 "# ccomplex_dotprod_generic (float* result, const float *input,\n\t"
114 "# const float *taps, unsigned num_bytes)\n\t"
115 "# float sum0 = 0;\n\t"
116 "# float sum1 = 0;\n\t"
117 "# float sum2 = 0;\n\t"
118 "# float sum3 = 0;\n\t"
120 "# sum0 += input[0] * taps[0] - input[1] * taps[1];\n\t"
121 "# sum1 += input[0] * taps[1] + input[1] * taps[0];\n\t"
122 "# sum2 += input[2] * taps[2] - input[3] * taps[3];\n\t"
123 "# sum3 += input[2] * taps[3] + input[3] * taps[2];\n\t"
126 "# } while (--n_2_ccomplex_blocks != 0);\n\t"
127 "# result[0] = sum0 + sum2;\n\t"
128 "# result[1] = sum1 + sum3;\n\t"
129 "# TODO: prefetch and better scheduling\n\t"
130 " xor %%r9, %%r9\n\t"
131 " xor %%r10, %%r10\n\t"
132 " movq %%rcx, %%rax\n\t"
133 " movq %%rcx, %%r8\n\t"
134 " movq %[rsi], %%r9\n\t"
135 " movq %[rdx], %%r10\n\t"
136 " xorps %%xmm6, %%xmm6 # zero accumulators\n\t"
137 " xorps %%xmm7, %%xmm7 # zero accumulators\n\t"
138 " shr $5, %%rax # rax = n_2_ccomplex_blocks / 2\n\t"
140 " jmp .%=L1_test\n\t"
141 " # 4 taps / loop\n\t"
142 " # something like ?? cycles / loop\n\t"
144 "# complex prod: C += A * B, w/ temp Z & Y (or B), xmmPN=$0x8000000080000000\n\t"
145 "# movups (%%r9), %%xmmA\n\t"
146 "# movups (%%r10), %%xmmB\n\t"
147 "# movups %%xmmA, %%xmmZ\n\t"
148 "# shufps $0xb1, %%xmmZ, %%xmmZ # swap internals\n\t"
149 "# mulps %%xmmB, %%xmmA\n\t"
150 "# mulps %%xmmZ, %%xmmB\n\t"
151 "# # SSE replacement for: pfpnacc %%xmmB, %%xmmA\n\t"
152 "# xorps %%xmmPN, %%xmmA\n\t"
153 "# movups %%xmmA, %%xmmZ\n\t"
154 "# unpcklps %%xmmB, %%xmmA\n\t"
155 "# unpckhps %%xmmB, %%xmmZ\n\t"
156 "# movups %%xmmZ, %%xmmY\n\t"
157 "# shufps $0x44, %%xmmA, %%xmmZ # b01000100\n\t"
158 "# shufps $0xee, %%xmmY, %%xmmA # b11101110\n\t"
159 "# addps %%xmmZ, %%xmmA\n\t"
160 "# addps %%xmmA, %%xmmC\n\t"
161 "# A=xmm0, B=xmm2, Z=xmm4\n\t"
162 "# A'=xmm1, B'=xmm3, Z'=xmm5\n\t"
163 " movups 0(%%r9), %%xmm0\n\t"
164 " movups 16(%%r9), %%xmm1\n\t"
165 " movups %%xmm0, %%xmm4\n\t"
166 " movups 0(%%r10), %%xmm2\n\t"
167 " mulps %%xmm2, %%xmm0\n\t"
168 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
169 " movups 16(%%r10), %%xmm3\n\t"
170 " movups %%xmm1, %%xmm5\n\t"
171 " addps %%xmm0, %%xmm6\n\t"
172 " mulps %%xmm3, %%xmm1\n\t"
173 " shufps $0xb1, %%xmm5, %%xmm5 # swap internals\n\t"
174 " addps %%xmm1, %%xmm6\n\t"
175 " mulps %%xmm4, %%xmm2\n\t"
176 " addps %%xmm2, %%xmm7\n\t"
177 " mulps %%xmm5, %%xmm3\n\t"
179 " addps %%xmm3, %%xmm7\n\t"
180 " add $32, %%r10\n\t"
184 " # We've handled the bulk of multiplies up to here.\n\t"
185 " # Let's sse if original n_2_ccomplex_blocks was odd.\n\t"
186 " # If so, we've got 2 more taps to do.\n\t"
189 " # The count was odd, do 2 more taps.\n\t"
190 " movups 0(%%r9), %%xmm0\n\t"
191 " movups %%xmm0, %%xmm4\n\t"
192 " movups 0(%%r10), %%xmm2\n\t"
193 " mulps %%xmm2, %%xmm0\n\t"
194 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
195 " addps %%xmm0, %%xmm6\n\t"
196 " mulps %%xmm4, %%xmm2\n\t"
197 " addps %%xmm2, %%xmm7\n\t"
199 " # neg inversor\n\t"
200 " xorps %%xmm1, %%xmm1\n\t"
201 " mov $0x80000000, %%r9\n\t"
202 " movd %%r9, %%xmm1\n\t"
203 " shufps $0x11, %%xmm1, %%xmm1 # b00010001 # 0 -0 0 -0\n\t"
205 " xorps %%xmm1, %%xmm6\n\t"
206 " movups %%xmm6, %%xmm2\n\t"
207 " unpcklps %%xmm7, %%xmm6\n\t"
208 " unpckhps %%xmm7, %%xmm2\n\t"
209 " movups %%xmm2, %%xmm3\n\t"
210 " shufps $0x44, %%xmm6, %%xmm2 # b01000100\n\t"
211 " shufps $0xee, %%xmm3, %%xmm6 # b11101110\n\t"
212 " addps %%xmm2, %%xmm6\n\t"
213 " # xmm6 = r1 i2 r3 i4\n\t"
214 " movhlps %%xmm6, %%xmm4 # xmm4 = r3 i4 ?? ??\n\t"
215 " addps %%xmm4, %%xmm6 # xmm6 = r1+r3 i2+i4 ?? ??\n\t"
216 " movlps %%xmm6, (%[rdi]) # store low 2x32 bits (complex) "
219 : [rsi]
"r"(input), [rdx]
"r"(taps),
"c"(num_bytes), [rdi]
"r"(result)
220 :
"rax",
"r8",
"r9",
"r10");
224 *result += input[num_points - 1] * taps[num_points - 1];
235#include <pmmintrin.h>
240 unsigned int num_points)
244 memset(&dotProduct, 0x0, 2 *
sizeof(
float));
246 unsigned int number = 0;
247 const unsigned int halfPoints = num_points / 2;
248 unsigned int isodd = num_points & 1;
250 __m128 x, y, yl, yh, z, tmp1, tmp2, dotProdVal;
257 for (; number < halfPoints; number++) {
286 dotProduct += (dotProductVector[0] + dotProductVector[1]);
289 dotProduct += input[num_points - 1] * taps[num_points - 1];
292 *result = dotProduct;
383#include <immintrin.h>
388 unsigned int num_points)
391 unsigned int isodd = num_points & 3;
394 memset(&dotProduct, 0x0, 2 *
sizeof(
float));
396 unsigned int number = 0;
397 const unsigned int quarterPoints = num_points / 4;
399 __m256 x, y, yl, yh, z, tmp1, tmp2, dotProdVal;
404 dotProdVal = _mm256_setzero_ps();
406 for (; number < quarterPoints; number++) {
407 x = _mm256_loadu_ps((
float*)a);
408 y = _mm256_loadu_ps((
float*)b);
410 yl = _mm256_moveldup_ps(y);
411 yh = _mm256_movehdup_ps(y);
413 tmp1 = _mm256_mul_ps(x, yl);
415 x = _mm256_shuffle_ps(x, x, 0xB1);
417 tmp2 = _mm256_mul_ps(x, yh);
419 z = _mm256_addsub_ps(tmp1,
422 dotProdVal = _mm256_add_ps(dotProdVal,
431 _mm256_storeu_ps((
float*)dotProductVector,
434 dotProduct += (dotProductVector[0] + dotProductVector[1] + dotProductVector[2] +
435 dotProductVector[3]);
437 for (
i = num_points - isodd;
i < num_points;
i++) {
438 dotProduct += input[
i] * taps[
i];
441 *result = dotProduct;
446#if LV_HAVE_AVX && LV_HAVE_FMA
447#include <immintrin.h>
449static inline void volk_32fc_x2_dot_prod_32fc_u_avx_fma(
lv_32fc_t* result,
452 unsigned int num_points)
455 unsigned int isodd = num_points & 3;
458 memset(&dotProduct, 0x0, 2 *
sizeof(
float));
460 unsigned int number = 0;
461 const unsigned int quarterPoints = num_points / 4;
463 __m256 x, y, yl, yh, z, tmp1, tmp2, dotProdVal;
468 dotProdVal = _mm256_setzero_ps();
470 for (; number < quarterPoints; number++) {
472 x = _mm256_loadu_ps((
float*)a);
473 y = _mm256_loadu_ps((
float*)b);
475 yl = _mm256_moveldup_ps(y);
476 yh = _mm256_movehdup_ps(y);
480 x = _mm256_shuffle_ps(x, x, 0xB1);
482 tmp2 = _mm256_mul_ps(x, yh);
484 z = _mm256_fmaddsub_ps(
487 dotProdVal = _mm256_add_ps(dotProdVal,
496 _mm256_storeu_ps((
float*)dotProductVector,
499 dotProduct += (dotProductVector[0] + dotProductVector[1] + dotProductVector[2] +
500 dotProductVector[3]);
502 for (
i = num_points - isodd;
i < num_points;
i++) {
503 dotProduct += input[
i] * taps[
i];
506 *result = dotProduct;
513#ifndef INCLUDED_volk_32fc_x2_dot_prod_32fc_a_H
514#define INCLUDED_volk_32fc_x2_dot_prod_32fc_a_H
522#if LV_HAVE_SSE && LV_HAVE_64
525static inline void volk_32fc_x2_dot_prod_32fc_a_sse_64(
lv_32fc_t* result,
528 unsigned int num_points)
531 const unsigned int num_bytes = num_points * 8;
532 unsigned int isodd = num_points & 1;
535 "# ccomplex_dotprod_generic (float* result, const float *input,\n\t"
536 "# const float *taps, unsigned num_bytes)\n\t"
537 "# float sum0 = 0;\n\t"
538 "# float sum1 = 0;\n\t"
539 "# float sum2 = 0;\n\t"
540 "# float sum3 = 0;\n\t"
542 "# sum0 += input[0] * taps[0] - input[1] * taps[1];\n\t"
543 "# sum1 += input[0] * taps[1] + input[1] * taps[0];\n\t"
544 "# sum2 += input[2] * taps[2] - input[3] * taps[3];\n\t"
545 "# sum3 += input[2] * taps[3] + input[3] * taps[2];\n\t"
548 "# } while (--n_2_ccomplex_blocks != 0);\n\t"
549 "# result[0] = sum0 + sum2;\n\t"
550 "# result[1] = sum1 + sum3;\n\t"
551 "# TODO: prefetch and better scheduling\n\t"
552 " xor %%r9, %%r9\n\t"
553 " xor %%r10, %%r10\n\t"
554 " movq %%rcx, %%rax\n\t"
555 " movq %%rcx, %%r8\n\t"
556 " movq %[rsi], %%r9\n\t"
557 " movq %[rdx], %%r10\n\t"
558 " xorps %%xmm6, %%xmm6 # zero accumulators\n\t"
559 " xorps %%xmm7, %%xmm7 # zero accumulators\n\t"
560 " shr $5, %%rax # rax = n_2_ccomplex_blocks / 2\n\t"
562 " jmp .%=L1_test\n\t"
563 " # 4 taps / loop\n\t"
564 " # something like ?? cycles / loop\n\t"
566 "# complex prod: C += A * B, w/ temp Z & Y (or B), xmmPN=$0x8000000080000000\n\t"
567 "# movaps (%%r9), %%xmmA\n\t"
568 "# movaps (%%r10), %%xmmB\n\t"
569 "# movaps %%xmmA, %%xmmZ\n\t"
570 "# shufps $0xb1, %%xmmZ, %%xmmZ # swap internals\n\t"
571 "# mulps %%xmmB, %%xmmA\n\t"
572 "# mulps %%xmmZ, %%xmmB\n\t"
573 "# # SSE replacement for: pfpnacc %%xmmB, %%xmmA\n\t"
574 "# xorps %%xmmPN, %%xmmA\n\t"
575 "# movaps %%xmmA, %%xmmZ\n\t"
576 "# unpcklps %%xmmB, %%xmmA\n\t"
577 "# unpckhps %%xmmB, %%xmmZ\n\t"
578 "# movaps %%xmmZ, %%xmmY\n\t"
579 "# shufps $0x44, %%xmmA, %%xmmZ # b01000100\n\t"
580 "# shufps $0xee, %%xmmY, %%xmmA # b11101110\n\t"
581 "# addps %%xmmZ, %%xmmA\n\t"
582 "# addps %%xmmA, %%xmmC\n\t"
583 "# A=xmm0, B=xmm2, Z=xmm4\n\t"
584 "# A'=xmm1, B'=xmm3, Z'=xmm5\n\t"
585 " movaps 0(%%r9), %%xmm0\n\t"
586 " movaps 16(%%r9), %%xmm1\n\t"
587 " movaps %%xmm0, %%xmm4\n\t"
588 " movaps 0(%%r10), %%xmm2\n\t"
589 " mulps %%xmm2, %%xmm0\n\t"
590 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
591 " movaps 16(%%r10), %%xmm3\n\t"
592 " movaps %%xmm1, %%xmm5\n\t"
593 " addps %%xmm0, %%xmm6\n\t"
594 " mulps %%xmm3, %%xmm1\n\t"
595 " shufps $0xb1, %%xmm5, %%xmm5 # swap internals\n\t"
596 " addps %%xmm1, %%xmm6\n\t"
597 " mulps %%xmm4, %%xmm2\n\t"
598 " addps %%xmm2, %%xmm7\n\t"
599 " mulps %%xmm5, %%xmm3\n\t"
601 " addps %%xmm3, %%xmm7\n\t"
602 " add $32, %%r10\n\t"
606 " # We've handled the bulk of multiplies up to here.\n\t"
607 " # Let's sse if original n_2_ccomplex_blocks was odd.\n\t"
608 " # If so, we've got 2 more taps to do.\n\t"
611 " # The count was odd, do 2 more taps.\n\t"
612 " movaps 0(%%r9), %%xmm0\n\t"
613 " movaps %%xmm0, %%xmm4\n\t"
614 " movaps 0(%%r10), %%xmm2\n\t"
615 " mulps %%xmm2, %%xmm0\n\t"
616 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
617 " addps %%xmm0, %%xmm6\n\t"
618 " mulps %%xmm4, %%xmm2\n\t"
619 " addps %%xmm2, %%xmm7\n\t"
621 " # neg inversor\n\t"
622 " xorps %%xmm1, %%xmm1\n\t"
623 " mov $0x80000000, %%r9\n\t"
624 " movd %%r9, %%xmm1\n\t"
625 " shufps $0x11, %%xmm1, %%xmm1 # b00010001 # 0 -0 0 -0\n\t"
627 " xorps %%xmm1, %%xmm6\n\t"
628 " movaps %%xmm6, %%xmm2\n\t"
629 " unpcklps %%xmm7, %%xmm6\n\t"
630 " unpckhps %%xmm7, %%xmm2\n\t"
631 " movaps %%xmm2, %%xmm3\n\t"
632 " shufps $0x44, %%xmm6, %%xmm2 # b01000100\n\t"
633 " shufps $0xee, %%xmm3, %%xmm6 # b11101110\n\t"
634 " addps %%xmm2, %%xmm6\n\t"
635 " # xmm6 = r1 i2 r3 i4\n\t"
636 " movhlps %%xmm6, %%xmm4 # xmm4 = r3 i4 ?? ??\n\t"
637 " addps %%xmm4, %%xmm6 # xmm6 = r1+r3 i2+i4 ?? ??\n\t"
638 " movlps %%xmm6, (%[rdi]) # store low 2x32 bits (complex) "
641 : [rsi]
"r"(input), [rdx]
"r"(taps),
"c"(num_bytes), [rdi]
"r"(result)
642 :
"rax",
"r8",
"r9",
"r10");
646 *result += input[num_points - 1] * taps[num_points - 1];
654#if LV_HAVE_SSE && LV_HAVE_32
656static inline void volk_32fc_x2_dot_prod_32fc_a_sse_32(
lv_32fc_t* result,
659 unsigned int num_points)
665 const unsigned int num_bytes = num_points*8;
666 unsigned int isodd = num_points & 1;
671 " #movl %%esp, %%ebp\n\t"
672 " movl 12(%%ebp), %%eax # input\n\t"
673 " movl 16(%%ebp), %%edx # taps\n\t"
674 " movl 20(%%ebp), %%ecx # n_bytes\n\t"
675 " xorps %%xmm6, %%xmm6 # zero accumulators\n\t"
676 " movaps 0(%%eax), %%xmm0\n\t"
677 " xorps %%xmm7, %%xmm7 # zero accumulators\n\t"
678 " movaps 0(%%edx), %%xmm2\n\t"
679 " shrl $5, %%ecx # ecx = n_2_ccomplex_blocks / 2\n\t"
680 " jmp .%=L1_test\n\t"
681 " # 4 taps / loop\n\t"
682 " # something like ?? cycles / loop\n\t"
684 "# complex prod: C += A * B, w/ temp Z & Y (or B), xmmPN=$0x8000000080000000\n\t"
685 "# movaps (%%eax), %%xmmA\n\t"
686 "# movaps (%%edx), %%xmmB\n\t"
687 "# movaps %%xmmA, %%xmmZ\n\t"
688 "# shufps $0xb1, %%xmmZ, %%xmmZ # swap internals\n\t"
689 "# mulps %%xmmB, %%xmmA\n\t"
690 "# mulps %%xmmZ, %%xmmB\n\t"
691 "# # SSE replacement for: pfpnacc %%xmmB, %%xmmA\n\t"
692 "# xorps %%xmmPN, %%xmmA\n\t"
693 "# movaps %%xmmA, %%xmmZ\n\t"
694 "# unpcklps %%xmmB, %%xmmA\n\t"
695 "# unpckhps %%xmmB, %%xmmZ\n\t"
696 "# movaps %%xmmZ, %%xmmY\n\t"
697 "# shufps $0x44, %%xmmA, %%xmmZ # b01000100\n\t"
698 "# shufps $0xee, %%xmmY, %%xmmA # b11101110\n\t"
699 "# addps %%xmmZ, %%xmmA\n\t"
700 "# addps %%xmmA, %%xmmC\n\t"
701 "# A=xmm0, B=xmm2, Z=xmm4\n\t"
702 "# A'=xmm1, B'=xmm3, Z'=xmm5\n\t"
703 " movaps 16(%%eax), %%xmm1\n\t"
704 " movaps %%xmm0, %%xmm4\n\t"
705 " mulps %%xmm2, %%xmm0\n\t"
706 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
707 " movaps 16(%%edx), %%xmm3\n\t"
708 " movaps %%xmm1, %%xmm5\n\t"
709 " addps %%xmm0, %%xmm6\n\t"
710 " mulps %%xmm3, %%xmm1\n\t"
711 " shufps $0xb1, %%xmm5, %%xmm5 # swap internals\n\t"
712 " addps %%xmm1, %%xmm6\n\t"
713 " mulps %%xmm4, %%xmm2\n\t"
714 " movaps 32(%%eax), %%xmm0\n\t"
715 " addps %%xmm2, %%xmm7\n\t"
716 " mulps %%xmm5, %%xmm3\n\t"
717 " addl $32, %%eax\n\t"
718 " movaps 32(%%edx), %%xmm2\n\t"
719 " addps %%xmm3, %%xmm7\n\t"
720 " addl $32, %%edx\n\t"
724 " # We've handled the bulk of multiplies up to here.\n\t"
725 " # Let's sse if original n_2_ccomplex_blocks was odd.\n\t"
726 " # If so, we've got 2 more taps to do.\n\t"
727 " movl 20(%%ebp), %%ecx # n_2_ccomplex_blocks\n\t"
728 " shrl $4, %%ecx\n\t"
729 " andl $1, %%ecx\n\t"
731 " # The count was odd, do 2 more taps.\n\t"
732 " # Note that we've already got mm0/mm2 preloaded\n\t"
733 " # from the main loop.\n\t"
734 " movaps %%xmm0, %%xmm4\n\t"
735 " mulps %%xmm2, %%xmm0\n\t"
736 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
737 " addps %%xmm0, %%xmm6\n\t"
738 " mulps %%xmm4, %%xmm2\n\t"
739 " addps %%xmm2, %%xmm7\n\t"
741 " # neg inversor\n\t"
742 " movl 8(%%ebp), %%eax \n\t"
743 " xorps %%xmm1, %%xmm1\n\t"
744 " movl $0x80000000, (%%eax)\n\t"
745 " movss (%%eax), %%xmm1\n\t"
746 " shufps $0x11, %%xmm1, %%xmm1 # b00010001 # 0 -0 0 -0\n\t"
748 " xorps %%xmm1, %%xmm6\n\t"
749 " movaps %%xmm6, %%xmm2\n\t"
750 " unpcklps %%xmm7, %%xmm6\n\t"
751 " unpckhps %%xmm7, %%xmm2\n\t"
752 " movaps %%xmm2, %%xmm3\n\t"
753 " shufps $0x44, %%xmm6, %%xmm2 # b01000100\n\t"
754 " shufps $0xee, %%xmm3, %%xmm6 # b11101110\n\t"
755 " addps %%xmm2, %%xmm6\n\t"
756 " # xmm6 = r1 i2 r3 i4\n\t"
757 " #movl 8(%%ebp), %%eax # @result\n\t"
758 " movhlps %%xmm6, %%xmm4 # xmm4 = r3 i4 ?? ??\n\t"
759 " addps %%xmm4, %%xmm6 # xmm6 = r1+r3 i2+i4 ?? ??\n\t"
760 " movlps %%xmm6, (%%eax) # store low 2x32 bits (complex) to memory\n\t"
764 :
"eax",
"ecx",
"edx"
768 int getem = num_bytes % 16;
771 *result += (input[num_points - 1] * taps[num_points - 1]);
782#include <pmmintrin.h>
787 unsigned int num_points)
790 const unsigned int num_bytes = num_points * 8;
791 unsigned int isodd = num_points & 1;
794 memset(&dotProduct, 0x0, 2 *
sizeof(
float));
796 unsigned int number = 0;
797 const unsigned int halfPoints = num_bytes >> 4;
799 __m128 x, y, yl, yh, z, tmp1, tmp2, dotProdVal;
806 for (; number < halfPoints; number++) {
835 dotProduct += (dotProductVector[0] + dotProductVector[1]);
838 dotProduct += input[num_points - 1] * taps[num_points - 1];
841 *result = dotProduct;
937 unsigned int num_points)
940 unsigned int quarter_points = num_points / 4;
947 float32x4x2_t a_val, b_val, c_val, accumulator;
948 float32x4x2_t tmp_real, tmp_imag;
949 accumulator.val[0] = vdupq_n_f32(0);
950 accumulator.val[1] = vdupq_n_f32(0);
952 for (number = 0; number < quarter_points; ++number) {
953 a_val = vld2q_f32((
float*)a_ptr);
954 b_val = vld2q_f32((
float*)b_ptr);
960 tmp_real.val[0] = vmulq_f32(a_val.val[0], b_val.val[0]);
962 tmp_real.val[1] = vmulq_f32(a_val.val[1], b_val.val[1]);
966 tmp_imag.val[0] = vmulq_f32(a_val.val[0], b_val.val[1]);
968 tmp_imag.val[1] = vmulq_f32(a_val.val[1], b_val.val[0]);
970 c_val.val[0] = vsubq_f32(tmp_real.val[0], tmp_real.val[1]);
971 c_val.val[1] = vaddq_f32(tmp_imag.val[0], tmp_imag.val[1]);
973 accumulator.val[0] = vaddq_f32(accumulator.val[0], c_val.val[0]);
974 accumulator.val[1] = vaddq_f32(accumulator.val[1], c_val.val[1]);
980 vst2q_f32((
float*)accum_result, accumulator);
981 *result = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
984 for (number = quarter_points * 4; number < num_points; ++number) {
985 *result += (*a_ptr++) * (*b_ptr++);
995 unsigned int num_points)
998 unsigned int quarter_points = num_points / 4;
1005 float32x4x2_t a_val, b_val, accumulator;
1006 float32x4x2_t tmp_imag;
1007 accumulator.val[0] = vdupq_n_f32(0);
1008 accumulator.val[1] = vdupq_n_f32(0);
1010 for (number = 0; number < quarter_points; ++number) {
1011 a_val = vld2q_f32((
float*)a_ptr);
1012 b_val = vld2q_f32((
float*)b_ptr);
1017 tmp_imag.val[1] = vmulq_f32(a_val.val[1], b_val.val[0]);
1018 tmp_imag.val[0] = vmulq_f32(a_val.val[0], b_val.val[0]);
1021 tmp_imag.val[1] = vmlaq_f32(tmp_imag.val[1], a_val.val[0], b_val.val[1]);
1022 tmp_imag.val[0] = vmlsq_f32(tmp_imag.val[0], a_val.val[1], b_val.val[1]);
1024 accumulator.val[0] = vaddq_f32(accumulator.val[0], tmp_imag.val[0]);
1025 accumulator.val[1] = vaddq_f32(accumulator.val[1], tmp_imag.val[1]);
1032 vst2q_f32((
float*)accum_result, accumulator);
1033 *result = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
1036 for (number = quarter_points * 4; number < num_points; ++number) {
1037 *result += (*a_ptr++) * (*b_ptr++);
1046 unsigned int num_points)
1049 unsigned int quarter_points = num_points / 4;
1050 unsigned int number;
1056 float32x4x2_t a_val, b_val, accumulator1, accumulator2;
1057 accumulator1.val[0] = vdupq_n_f32(0);
1058 accumulator1.val[1] = vdupq_n_f32(0);
1059 accumulator2.val[0] = vdupq_n_f32(0);
1060 accumulator2.val[1] = vdupq_n_f32(0);
1062 for (number = 0; number < quarter_points; ++number) {
1063 a_val = vld2q_f32((
float*)a_ptr);
1064 b_val = vld2q_f32((
float*)b_ptr);
1069 accumulator1.val[0] = vmlaq_f32(accumulator1.val[0], a_val.val[0], b_val.val[0]);
1070 accumulator1.val[1] = vmlaq_f32(accumulator1.val[1], a_val.val[0], b_val.val[1]);
1071 accumulator2.val[0] = vmlsq_f32(accumulator2.val[0], a_val.val[1], b_val.val[1]);
1072 accumulator2.val[1] = vmlaq_f32(accumulator2.val[1], a_val.val[1], b_val.val[0]);
1077 accumulator1.val[0] = vaddq_f32(accumulator1.val[0], accumulator2.val[0]);
1078 accumulator1.val[1] = vaddq_f32(accumulator1.val[1], accumulator2.val[1]);
1080 vst2q_f32((
float*)accum_result, accumulator1);
1081 *result = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
1084 for (number = quarter_points * 4; number < num_points; ++number) {
1085 *result += (*a_ptr++) * (*b_ptr++);
1094 unsigned int num_points)
1099 unsigned int quarter_points = num_points / 8;
1100 unsigned int number;
1106 float32x4x4_t a_val, b_val, accumulator1, accumulator2;
1107 float32x4x2_t reduced_accumulator;
1108 accumulator1.val[0] = vdupq_n_f32(0);
1109 accumulator1.val[1] = vdupq_n_f32(0);
1110 accumulator1.val[2] = vdupq_n_f32(0);
1111 accumulator1.val[3] = vdupq_n_f32(0);
1112 accumulator2.val[0] = vdupq_n_f32(0);
1113 accumulator2.val[1] = vdupq_n_f32(0);
1114 accumulator2.val[2] = vdupq_n_f32(0);
1115 accumulator2.val[3] = vdupq_n_f32(0);
1118 for (number = 0; number < quarter_points; ++number) {
1119 a_val = vld4q_f32((
float*)a_ptr);
1120 b_val = vld4q_f32((
float*)b_ptr);
1125 accumulator1.val[0] = vmlaq_f32(accumulator1.val[0], a_val.val[0], b_val.val[0]);
1126 accumulator1.val[1] = vmlaq_f32(accumulator1.val[1], a_val.val[0], b_val.val[1]);
1128 accumulator1.val[2] = vmlaq_f32(accumulator1.val[2], a_val.val[2], b_val.val[2]);
1129 accumulator1.val[3] = vmlaq_f32(accumulator1.val[3], a_val.val[2], b_val.val[3]);
1131 accumulator2.val[0] = vmlsq_f32(accumulator2.val[0], a_val.val[1], b_val.val[1]);
1132 accumulator2.val[1] = vmlaq_f32(accumulator2.val[1], a_val.val[1], b_val.val[0]);
1134 accumulator2.val[2] = vmlsq_f32(accumulator2.val[2], a_val.val[3], b_val.val[3]);
1135 accumulator2.val[3] = vmlaq_f32(accumulator2.val[3], a_val.val[3], b_val.val[2]);
1141 accumulator1.val[0] = vaddq_f32(accumulator1.val[0], accumulator1.val[2]);
1142 accumulator1.val[1] = vaddq_f32(accumulator1.val[1], accumulator1.val[3]);
1143 accumulator2.val[0] = vaddq_f32(accumulator2.val[0], accumulator2.val[2]);
1144 accumulator2.val[1] = vaddq_f32(accumulator2.val[1], accumulator2.val[3]);
1145 reduced_accumulator.val[0] = vaddq_f32(accumulator1.val[0], accumulator2.val[0]);
1146 reduced_accumulator.val[1] = vaddq_f32(accumulator1.val[1], accumulator2.val[1]);
1149 vst2q_f32((
float*)accum_result, reduced_accumulator);
1150 *result = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
1153 for (number = quarter_points * 8; number < num_points; ++number) {
1154 *result += (*a_ptr++) * (*b_ptr++);
1162#include <immintrin.h>
1167 unsigned int num_points)
1170 unsigned int isodd = num_points & 3;
1173 memset(&dotProduct, 0x0, 2 *
sizeof(
float));
1175 unsigned int number = 0;
1176 const unsigned int quarterPoints = num_points / 4;
1178 __m256 x, y, yl, yh, z, tmp1, tmp2, dotProdVal;
1183 dotProdVal = _mm256_setzero_ps();
1185 for (; number < quarterPoints; number++) {
1187 x = _mm256_load_ps((
float*)a);
1188 y = _mm256_load_ps((
float*)b);
1190 yl = _mm256_moveldup_ps(y);
1191 yh = _mm256_movehdup_ps(y);
1193 tmp1 = _mm256_mul_ps(x, yl);
1195 x = _mm256_shuffle_ps(x, x, 0xB1);
1197 tmp2 = _mm256_mul_ps(x, yh);
1199 z = _mm256_addsub_ps(tmp1,
1202 dotProdVal = _mm256_add_ps(dotProdVal,
1211 _mm256_store_ps((
float*)dotProductVector,
1214 dotProduct += (dotProductVector[0] + dotProductVector[1] + dotProductVector[2] +
1215 dotProductVector[3]);
1217 for (
i = num_points - isodd;
i < num_points;
i++) {
1218 dotProduct += input[
i] * taps[
i];
1221 *result = dotProduct;
1226#if LV_HAVE_AVX && LV_HAVE_FMA
1227#include <immintrin.h>
1229static inline void volk_32fc_x2_dot_prod_32fc_a_avx_fma(
lv_32fc_t* result,
1232 unsigned int num_points)
1235 unsigned int isodd = num_points & 3;
1238 memset(&dotProduct, 0x0, 2 *
sizeof(
float));
1240 unsigned int number = 0;
1241 const unsigned int quarterPoints = num_points / 4;
1243 __m256 x, y, yl, yh, z, tmp1, tmp2, dotProdVal;
1248 dotProdVal = _mm256_setzero_ps();
1250 for (; number < quarterPoints; number++) {
1252 x = _mm256_load_ps((
float*)a);
1253 y = _mm256_load_ps((
float*)b);
1255 yl = _mm256_moveldup_ps(y);
1256 yh = _mm256_movehdup_ps(y);
1260 x = _mm256_shuffle_ps(x, x, 0xB1);
1262 tmp2 = _mm256_mul_ps(x, yh);
1264 z = _mm256_fmaddsub_ps(
1267 dotProdVal = _mm256_add_ps(dotProdVal,
1276 _mm256_store_ps((
float*)dotProductVector,
1279 dotProduct += (dotProductVector[0] + dotProductVector[1] + dotProductVector[2] +
1280 dotProductVector[3]);
1282 for (
i = num_points - isodd;
i < num_points;
i++) {
1283 dotProduct += input[
i] * taps[
i];
1286 *result = dotProduct;
FORCE_INLINE __m128 _mm_movehdup_ps(__m128 a)
Definition: sse2neon.h:6611
float32x4_t __m128
Definition: sse2neon.h:235
FORCE_INLINE __m128 _mm_addsub_ps(__m128 a, __m128 b)
Definition: sse2neon.h:6496
#define _mm_shuffle_ps(a, b, imm)
Definition: sse2neon.h:2586
FORCE_INLINE void _mm_storeu_ps(float *p, __m128 a)
Definition: sse2neon.h:2787
FORCE_INLINE __m128 _mm_moveldup_ps(__m128 a)
Definition: sse2neon.h:6627
FORCE_INLINE __m128 _mm_mul_ps(__m128 a, __m128 b)
Definition: sse2neon.h:2205
FORCE_INLINE __m128 _mm_loadu_ps(const float *p)
Definition: sse2neon.h:1941
FORCE_INLINE __m128 _mm_setzero_ps(void)
Definition: sse2neon.h:2531
FORCE_INLINE __m128 _mm_add_ps(__m128 a, __m128 b)
Definition: sse2neon.h:1039
FORCE_INLINE __m128 _mm_load_ps(const float *p)
Definition: sse2neon.h:1858
FORCE_INLINE void _mm_store_ps(float *p, __m128 a)
Definition: sse2neon.h:2704
static void volk_32fc_x2_dot_prod_32fc_neon_optfmaunroll(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:1091
static void volk_32fc_x2_dot_prod_32fc_a_sse3(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:784
static void volk_32fc_x2_dot_prod_32fc_a_avx(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:1164
static void volk_32fc_x2_dot_prod_32fc_u_avx(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:385
static void volk_32fc_x2_dot_prod_32fc_neon(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:934
static void volk_32fc_x2_dot_prod_32fc_generic(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:64
static void volk_32fc_x2_dot_prod_32fc_u_sse3(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:237
static void volk_32fc_x2_dot_prod_32fc_neon_opttests(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:992
static void volk_32fc_x2_dot_prod_32fc_neon_optfma(lv_32fc_t *result, const lv_32fc_t *input, const lv_32fc_t *taps, unsigned int num_points)
Definition: volk_32fc_x2_dot_prod_32fc.h:1043
#define __VOLK_VOLATILE
Definition: volk_common.h:74
#define __VOLK_PREFETCH(addr)
Definition: volk_common.h:72
#define __VOLK_ASM
Definition: volk_common.h:73
#define __VOLK_ATTR_ALIGNED(x)
Definition: volk_common.h:66
float complex lv_32fc_t
Definition: volk_complex.h:74
for i
Definition: volk_config_fixed.tmpl.h:13