61#ifndef INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_u_H
62#define INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_u_H
73 unsigned int num_points)
76 for (
unsigned int i = 0; i < num_points; ++i) {
77 res += (*input++) *
lv_conj((*taps++));
89 unsigned int num_points)
92 const unsigned int num_bytes = num_points * 8;
94 float* res = (
float*)result;
95 float* in = (
float*)input;
96 float* tp = (
float*)taps;
97 unsigned int n_2_ccomplex_blocks = num_bytes >> 4;
99 float sum0[2] = { 0, 0 };
100 float sum1[2] = { 0, 0 };
103 for (i = 0; i < n_2_ccomplex_blocks; ++i) {
104 sum0[0] += in[0] * tp[0] + in[1] * tp[1];
105 sum0[1] += (-in[0] * tp[1]) + in[1] * tp[0];
106 sum1[0] += in[2] * tp[2] + in[3] * tp[3];
107 sum1[1] += (-in[2] * tp[3]) + in[3] * tp[2];
113 res[0] = sum0[0] + sum1[0];
114 res[1] = sum0[1] + sum1[1];
116 if (num_bytes >> 3 & 1) {
117 *result += input[(num_bytes >> 3) - 1] *
lv_conj(taps[(num_bytes >> 3) - 1]);
125#include <immintrin.h>
130 unsigned int num_points)
133 __m256 sum_a_mult_b_real = _mm256_setzero_ps();
134 __m256 sum_a_mult_b_imag = _mm256_setzero_ps();
136 for (
long unsigned i = 0; i < (num_points & ~3u); i += 4) {
148 __m256 a = _mm256_loadu_ps((
const float*)&input[i]);
149 __m256 b = _mm256_loadu_ps((
const float*)&taps[i]);
150 __m256 b_real = _mm256_moveldup_ps(b);
151 __m256 b_imag = _mm256_movehdup_ps(b);
154 sum_a_mult_b_real = _mm256_add_ps(sum_a_mult_b_real, _mm256_mul_ps(a, b_real));
156 sum_a_mult_b_imag = _mm256_addsub_ps(sum_a_mult_b_imag, _mm256_mul_ps(a, b_imag));
160 sum_a_mult_b_imag = _mm256_permute_ps(sum_a_mult_b_imag,
_MM_SHUFFLE(2, 3, 0, 1));
162 __m256 sum = _mm256_add_ps(sum_a_mult_b_real, sum_a_mult_b_imag);
166 sum = _mm256_add_ps(sum, _mm256_permute2f128_ps(sum, sum, 0x01));
168 sum = _mm256_add_ps(sum, _mm256_permute_ps(sum,
_MM_SHUFFLE(1, 0, 3, 2)));
170 __m128 lower = _mm256_extractf128_ps(sum, 0);
174 for (
long unsigned i = num_points & ~3u; i < num_points; ++i) {
186#include <pmmintrin.h>
187#include <xmmintrin.h>
192 unsigned int num_points)
198 for (
long unsigned i = 0; i < (num_points & ~1u); i += 2) {
232 if (num_points & 1u) {
248 unsigned int num_points)
251 unsigned int quarter_points = num_points / 4;
258 float32x4x2_t a_val, b_val, accumulator;
259 float32x4x2_t tmp_imag;
260 accumulator.val[0] = vdupq_n_f32(0);
261 accumulator.val[1] = vdupq_n_f32(0);
263 for (number = 0; number < quarter_points; ++number) {
264 a_val = vld2q_f32((
float*)a_ptr);
265 b_val = vld2q_f32((
float*)b_ptr);
270 tmp_imag.val[1] = vmulq_f32(a_val.val[1], b_val.val[0]);
271 tmp_imag.val[0] = vmulq_f32(a_val.val[0], b_val.val[0]);
274 tmp_imag.val[1] = vmlsq_f32(tmp_imag.val[1], a_val.val[0], b_val.val[1]);
275 tmp_imag.val[0] = vmlaq_f32(tmp_imag.val[0], a_val.val[1], b_val.val[1]);
277 accumulator.val[0] = vaddq_f32(accumulator.val[0], tmp_imag.val[0]);
278 accumulator.val[1] = vaddq_f32(accumulator.val[1], tmp_imag.val[1]);
285 vst2q_f32((
float*)accum_result, accumulator);
286 *result = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
289 for (number = quarter_points * 4; number < num_points; ++number) {
290 *result += (*a_ptr++) *
lv_conj(*b_ptr++);
298#ifndef INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_a_H
299#define INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_a_H
307#include <immintrin.h>
312 unsigned int num_points)
315 __m256 sum_a_mult_b_real = _mm256_setzero_ps();
316 __m256 sum_a_mult_b_imag = _mm256_setzero_ps();
318 for (
long unsigned i = 0; i < (num_points & ~3u); i += 4) {
330 __m256 a = _mm256_load_ps((
const float*)&input[i]);
331 __m256 b = _mm256_load_ps((
const float*)&taps[i]);
332 __m256 b_real = _mm256_moveldup_ps(b);
333 __m256 b_imag = _mm256_movehdup_ps(b);
336 sum_a_mult_b_real = _mm256_add_ps(sum_a_mult_b_real, _mm256_mul_ps(a, b_real));
338 sum_a_mult_b_imag = _mm256_addsub_ps(sum_a_mult_b_imag, _mm256_mul_ps(a, b_imag));
342 sum_a_mult_b_imag = _mm256_permute_ps(sum_a_mult_b_imag,
_MM_SHUFFLE(2, 3, 0, 1));
344 __m256 sum = _mm256_add_ps(sum_a_mult_b_real, sum_a_mult_b_imag);
348 sum = _mm256_add_ps(sum, _mm256_permute2f128_ps(sum, sum, 0x01));
350 sum = _mm256_add_ps(sum, _mm256_permute_ps(sum,
_MM_SHUFFLE(1, 0, 3, 2)));
352 __m128 lower = _mm256_extractf128_ps(sum, 0);
356 for (
long unsigned i = num_points & ~3u; i < num_points; ++i) {
367#include <pmmintrin.h>
368#include <xmmintrin.h>
373 unsigned int num_points)
379 for (
long unsigned i = 0; i < (num_points & ~1u); i += 2) {
413 if (num_points & 1u) {
425#if LV_HAVE_SSE && LV_HAVE_64
427static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_sse(
lv_32fc_t* result,
430 unsigned int num_points)
433 const unsigned int num_bytes = num_points * 8;
436 static const uint32_t conjugator[4] = {
437 0x00000000, 0x80000000, 0x00000000, 0x80000000
441 "# ccomplex_conjugate_dotprod_generic (float* result, const float *input,\n\t"
442 "# const float *taps, unsigned num_bytes)\n\t"
443 "# float sum0 = 0;\n\t"
444 "# float sum1 = 0;\n\t"
445 "# float sum2 = 0;\n\t"
446 "# float sum3 = 0;\n\t"
448 "# sum0 += input[0] * taps[0] - input[1] * taps[1];\n\t"
449 "# sum1 += input[0] * taps[1] + input[1] * taps[0];\n\t"
450 "# sum2 += input[2] * taps[2] - input[3] * taps[3];\n\t"
451 "# sum3 += input[2] * taps[3] + input[3] * taps[2];\n\t"
454 "# } while (--n_2_ccomplex_blocks != 0);\n\t"
455 "# result[0] = sum0 + sum2;\n\t"
456 "# result[1] = sum1 + sum3;\n\t"
457 "# TODO: prefetch and better scheduling\n\t"
458 " xor %%r9, %%r9\n\t"
459 " xor %%r10, %%r10\n\t"
460 " movq %[conjugator], %%r9\n\t"
461 " movq %%rcx, %%rax\n\t"
462 " movaps 0(%%r9), %%xmm8\n\t"
463 " movq %%rcx, %%r8\n\t"
464 " movq %[rsi], %%r9\n\t"
465 " movq %[rdx], %%r10\n\t"
466 " xorps %%xmm6, %%xmm6 # zero accumulators\n\t"
467 " xorps %%xmm7, %%xmm7 # zero accumulators\n\t"
468 " shr $5, %%rax # rax = n_2_ccomplex_blocks / 2\n\t"
470 " xorps %%xmm8, %%xmm2\n\t"
471 " jmp .%=L1_test\n\t"
472 " # 4 taps / loop\n\t"
473 " # something like ?? cycles / loop\n\t"
475 "# complex prod: C += A * B, w/ temp Z & Y (or B), xmmPN=$0x8000000080000000\n\t"
476 "# movaps (%%r9), %%xmmA\n\t"
477 "# movaps (%%r10), %%xmmB\n\t"
478 "# movaps %%xmmA, %%xmmZ\n\t"
479 "# shufps $0xb1, %%xmmZ, %%xmmZ # swap internals\n\t"
480 "# mulps %%xmmB, %%xmmA\n\t"
481 "# mulps %%xmmZ, %%xmmB\n\t"
482 "# # SSE replacement for: pfpnacc %%xmmB, %%xmmA\n\t"
483 "# xorps %%xmmPN, %%xmmA\n\t"
484 "# movaps %%xmmA, %%xmmZ\n\t"
485 "# unpcklps %%xmmB, %%xmmA\n\t"
486 "# unpckhps %%xmmB, %%xmmZ\n\t"
487 "# movaps %%xmmZ, %%xmmY\n\t"
488 "# shufps $0x44, %%xmmA, %%xmmZ # b01000100\n\t"
489 "# shufps $0xee, %%xmmY, %%xmmA # b11101110\n\t"
490 "# addps %%xmmZ, %%xmmA\n\t"
491 "# addps %%xmmA, %%xmmC\n\t"
492 "# A=xmm0, B=xmm2, Z=xmm4\n\t"
493 "# A'=xmm1, B'=xmm3, Z'=xmm5\n\t"
494 " movaps 0(%%r9), %%xmm0\n\t"
495 " movaps 16(%%r9), %%xmm1\n\t"
496 " movaps %%xmm0, %%xmm4\n\t"
497 " movaps 0(%%r10), %%xmm2\n\t"
498 " xorps %%xmm8, %%xmm2\n\t"
499 " mulps %%xmm2, %%xmm0\n\t"
500 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
501 " movaps 16(%%r10), %%xmm3\n\t"
502 " movaps %%xmm1, %%xmm5\n\t"
503 " xorps %%xmm8, %%xmm3\n\t"
504 " addps %%xmm0, %%xmm6\n\t"
505 " mulps %%xmm3, %%xmm1\n\t"
506 " shufps $0xb1, %%xmm5, %%xmm5 # swap internals\n\t"
507 " addps %%xmm1, %%xmm6\n\t"
508 " mulps %%xmm4, %%xmm2\n\t"
509 " addps %%xmm2, %%xmm7\n\t"
510 " mulps %%xmm5, %%xmm3\n\t"
512 " addps %%xmm3, %%xmm7\n\t"
513 " add $32, %%r10\n\t"
517 " # We've handled the bulk of multiplies up to here.\n\t"
518 " # Let's sse if original n_2_ccomplex_blocks was odd.\n\t"
519 " # If so, we've got 2 more taps to do.\n\t"
522 " # The count was odd, do 2 more taps.\n\t"
523 " # Note that we've already got mm0/mm2 preloaded\n\t"
524 " # from the main loop.\n\t"
525 " movaps 0(%%r9), %%xmm0\n\t"
526 " movaps %%xmm0, %%xmm4\n\t"
527 " movaps 0(%%r10), %%xmm2\n\t"
528 " xorps %%xmm8, %%xmm2\n\t"
529 " mulps %%xmm2, %%xmm0\n\t"
530 " shufps $0xb1, %%xmm4, %%xmm4 # swap internals\n\t"
531 " addps %%xmm0, %%xmm6\n\t"
532 " mulps %%xmm4, %%xmm2\n\t"
533 " addps %%xmm2, %%xmm7\n\t"
535 " # neg inversor\n\t"
536 " xorps %%xmm1, %%xmm1\n\t"
537 " mov $0x80000000, %%r9\n\t"
538 " movd %%r9, %%xmm1\n\t"
539 " shufps $0x11, %%xmm1, %%xmm1 # b00010001 # 0 -0 0 -0\n\t"
541 " xorps %%xmm1, %%xmm6\n\t"
542 " movaps %%xmm6, %%xmm2\n\t"
543 " unpcklps %%xmm7, %%xmm6\n\t"
544 " unpckhps %%xmm7, %%xmm2\n\t"
545 " movaps %%xmm2, %%xmm3\n\t"
546 " shufps $0x44, %%xmm6, %%xmm2 # b01000100\n\t"
547 " shufps $0xee, %%xmm3, %%xmm6 # b11101110\n\t"
548 " addps %%xmm2, %%xmm6\n\t"
549 " # xmm6 = r1 i2 r3 i4\n\t"
550 " movhlps %%xmm6, %%xmm4 # xmm4 = r3 i4 ?? ??\n\t"
551 " addps %%xmm4, %%xmm6 # xmm6 = r1+r3 i2+i4 ?? ??\n\t"
552 " movlps %%xmm6, (%[rdi]) # store low 2x32 bits (complex) "
559 [conjugator]
"r"(conjugator)
560 :
"rax",
"r8",
"r9",
"r10");
562 int getem = num_bytes % 16;
564 for (; getem > 0; getem -= 8) {
565 *result += (input[(num_bytes >> 3) - 1] *
lv_conj(taps[(num_bytes >> 3) - 1]));