fixeria/srsRAN
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

SSE optimization for srslte_bit_interleave_w_offset

Browse source
This commit is contained in:
Xavier Arteaga 2017-10-20 18:17:07 +02:00
parent a570e63c5b
commit c1b296eb2c
1 changed files with 126 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

126
lib/src/phy/utils/bit.c
View file

@ -31,6 +31,12 @@
#include <string.h>
#include <stddef.h>
#ifdef LV_HAVE_SSE
#include <immintrin.h>
#endif /* LV_HAVE_SSE */
#include "srslte/phy/utils/bit.h"
void srslte_bit_interleave(uint8_t *input, uint8_t *output, uint16_t *interleaver, uint32_t nof_bits) {
@ -53,6 +59,125 @@ void srslte_bit_interleave_w_offset(uint8_t *input, uint8_t *output, uint16_t *i
}
w_offset_p=8-w_offset;
}
#ifdef LV_HAVE_SSE
__m64 m64mask = _mm_setr_pi8((uint8_t) 0x80, 0x40, 0x20, 0x10, 0x8, 0x4, 0x2, 0x1);
__m128i m128mask = _mm_set1_epi64(m64mask);
union {
uint8_t v[8];
__m64 m64;
} a, b, c;
union {
__m128i m128;
uint16_t u16[8];
uint8_t u8[16];
struct {
__m64 reg_a;
__m64 reg_b;
} m64;
struct {
uint16_t i0, i1, i2, i3, i4, i5, i6, i7;
} v;
} ipx, epx, ipx2, epx2, b128, a128, c128;
uint32_t i = st;
for (; i < (nof_bits / 8 - 1); i += 2) {
ipx.m128 = _mm_loadu_si128((__m128i *) (interleaver + (i * 8) - w_offset_p));
epx.m128 = _mm_shuffle_epi8(ipx.m128, _mm_set_epi8(0x00, 0x02, 0x04, 0x06, 0x08, 0x0A, 0x0C, 0x0E,
0x00, 0x02, 0x04, 0x06, 0x08, 0x0A, 0x0C, 0x0E));
ipx2.m128 = _mm_loadu_si128((__m128i *) (interleaver + ((i + 1) * 8) - w_offset_p));
epx2.m128 = _mm_shuffle_epi8(ipx2.m128, _mm_set_epi8(0x00, 0x02, 0x04, 0x06, 0x08, 0x0A, 0x0C, 0x0E,
0x00, 0x02, 0x04, 0x06, 0x08, 0x0A, 0x0C, 0x0E));
epx.m64.reg_b = epx2.m64.reg_a;
b128.m128 = _mm_and_si128(epx.m128, _mm_set1_epi8(0x7));
b128.m128 = _mm_shuffle_epi8(m128mask, b128.m128);
ipx.m128 = _mm_srli_epi16(ipx.m128, 3);
ipx2.m128 = _mm_srli_epi16(ipx2.m128, 3);
a128.m128 = _mm_set_epi8(input[ipx2.v.i0],
input[ipx2.v.i1],
input[ipx2.v.i2],
input[ipx2.v.i3],
input[ipx2.v.i4],
input[ipx2.v.i5],
input[ipx2.v.i6],
input[ipx2.v.i7],
input[ipx.v.i0],
input[ipx.v.i1],
input[ipx.v.i2],
input[ipx.v.i3],
input[ipx.v.i4],
input[ipx.v.i5],
input[ipx.v.i6],
input[ipx.v.i7]);
c128.m128 = _mm_cmpeq_epi8(_mm_and_si128(a128.m128, b128.m128), b128.m128);
uint16_t o = (uint16_t) _mm_movemask_epi8(c128.m128);
*((uint16_t *) (output + i)) = o;
}
for (; i < nof_bits / 8; i++) {
ipx.m128 = _mm_loadu_si128((__m128i *) (interleaver + i * 8 - w_offset_p));
epx.m128 = _mm_shuffle_epi8(ipx.m128, _mm_set_epi8(0x00, 0x02, 0x04, 0x06, 0x08, 0x0A, 0x0C, 0x0E,
0x00, 0x02, 0x04, 0x06, 0x08, 0x0A, 0x0C, 0x0E));
b.m64 = _mm_and_si64(epx.m64.reg_a, _mm_set1_pi8(0x7));
b.m64 = _mm_shuffle_pi8(m64mask, b.m64);
ipx.m128 = _mm_srli_epi16(ipx.m128, 3);
a.m64 = _mm_set_pi8(input[ipx.v.i0],
input[ipx.v.i1],
input[ipx.v.i2],
input[ipx.v.i3],
input[ipx.v.i4],
input[ipx.v.i5],
input[ipx.v.i6],
input[ipx.v.i7]);
c.m64 = _mm_cmpeq_pi8(_mm_and_si64(a.m64, b.m64), b.m64);
output[i] = (uint8_t) _mm_movemask_pi8(c.m64);
}
#if 0
/* THIS PIECE OF CODE IS FOR CHECKING SIMD BEHAVIOUR. DO NOT ENABLE. */
uint8_t *output2 = malloc(nof_bits/8);
for (i=st;i<nof_bits/8;i++) {
uint16_t i_p0 = interleaver[i*8+0-w_offset_p];
uint16_t i_p1 = interleaver[i*8+1-w_offset_p];
uint16_t i_p2 = interleaver[i*8+2-w_offset_p];
uint16_t i_p3 = interleaver[i*8+3-w_offset_p];
uint16_t i_p4 = interleaver[i*8+4-w_offset_p];
uint16_t i_p5 = interleaver[i*8+5-w_offset_p];
uint16_t i_p6 = interleaver[i*8+6-w_offset_p];
uint16_t i_p7 = interleaver[i*8+7-w_offset_p];
uint8_t out0 = (input[i_p0/8] & mask[i_p0%8])?mask[0]:(uint8_t)0;
uint8_t out1 = (input[i_p1/8] & mask[i_p1%8])?mask[1]:(uint8_t)0;
uint8_t out2 = (input[i_p2/8] & mask[i_p2%8])?mask[2]:(uint8_t)0;
uint8_t out3 = (input[i_p3/8] & mask[i_p3%8])?mask[3]:(uint8_t)0;
uint8_t out4 = (input[i_p4/8] & mask[i_p4%8])?mask[4]:(uint8_t)0;
uint8_t out5 = (input[i_p5/8] & mask[i_p5%8])?mask[5]:(uint8_t)0;
uint8_t out6 = (input[i_p6/8] & mask[i_p6%8])?mask[6]:(uint8_t)0;
uint8_t out7 = (input[i_p7/8] & mask[i_p7%8])?mask[7]:(uint8_t)0;
output2[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7;
}
for(i = st; i < nof_bits/8; i++) {
if (output[i] != output2[i]) {
printf("%05d/%05d %02X %02X\n", i, nof_bits/8, output[i], output2[i]);
}
//output[i] = output2[i];
}
free(output2);
#endif
#else /* LV_HAVE_SSE */
for (uint32_t i=st;i<nof_bits/8;i++) {
uint16_t i_p0 = interleaver[i*8+0-w_offset_p];
@ -75,6 +200,7 @@ void srslte_bit_interleave_w_offset(uint8_t *input, uint8_t *output, uint16_t *i
output[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7;
}
#endif /* LV_HAVE_SSE */
for (uint32_t j=0;j<nof_bits%8;j++) {
uint16_t i_p = interleaver[(nof_bits/8)*8+j-w_offset];
if (input[i_p/8] & mask[i_p%8]) {