Skip to content

O
OpenXG-RAN
Commit 243a56ad authored by mashish's avatar mashish
Browse files
Options

avx2 support in 16qam and 64qam in nr_dlsch_llr

parent c19ba39e
Show whitespace changes
Inline Side-by-side
Showing with 254 additions and 3 deletions
openair1/PHY/NR_UE_TRANSPORT/nr_dlsch_llr_computation.c
View file @ 243a56ad
......@@ -37,6 +37,7 @@
#include "PHY/sse_intrin.h"
//#define DEBUG_LLR_SIC
// #define DEBUG_AVX_LLR
int16_t nr_zeros[8] __attribute__ ((aligned(16))) = {0,0,0,0,0,0,0,0};
......@@ -682,9 +683,17 @@ void nr_dlsch_16qam_llr(NR_DL_FRAME_PARMS *frame_parms,
{
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
printf("DLSCH 16QAM AVX\n");
__m256i *rxF = (__m256i *)&rxdataF_comp[0][(symbol * nb_rb * 12)];
__m256i *ch_mag;
__m256i llr256[2];
register __m256i ymm0;
#else
__m128i *rxF = (__m128i*)&rxdataF_comp[0][(symbol*nb_rb*12)];
__m128i *ch_mag;
__m128i llr128[2];
#endif
uint32_t *llr32;
#elif defined(__arm__)
int16x8_t *rxF = (int16x8_t*)&rxdataF_comp[0][(symbol*nb_rb*12)];
......@@ -713,7 +722,11 @@ void nr_dlsch_16qam_llr(NR_DL_FRAME_PARMS *frame_parms,
#endif
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
ch_mag = (__m256i *)&dl_ch_mag[0][(symbol * nb_rb * 12)];
#else
ch_mag = (__m128i*)&dl_ch_mag[0][(symbol*nb_rb*12)];
#endif
#elif defined(__arm__)
ch_mag = (int16x8_t*)&dl_ch_mag[0][(symbol*nb_rb*12)];
#endif
......@@ -724,6 +737,12 @@ void nr_dlsch_16qam_llr(NR_DL_FRAME_PARMS *frame_parms,
else
*llr32p += (len<<2);
#ifdef __AVX2__
unsigned char len_mod8 = len&7;
printf("length = %d, len_mod8 = %d\n", len, len_mod4);
len >>= 3; // length in quad words (4 REs)
len += (len_mod8 == 0 ? 0 : 1);
#else
// printf("len=%d\n", len);
len_mod4 = len&3;
// printf("len_mod4=%d\n", len_mod4);
......@@ -731,9 +750,99 @@ void nr_dlsch_16qam_llr(NR_DL_FRAME_PARMS *frame_parms,
// printf("len>>=2=%d\n", len);
len+=(len_mod4==0 ? 0 : 1);
// printf("len+=%d\n", len);
#endif
for (i=0; i<len; i++) {
#if defined(__x86_64__) || defined(__i386)
#ifdef __AVX2__
ymm0 = _mm256_abs_epi16(rxF[i]);
ymm0 = _mm256_subs_epi16(ch_mag[i], ymm0);
llr256[0] = _mm256_unpacklo_epi32(rxF[i], ymm0);
llr256[1] = _mm256_unpackhi_epi32(rxF[i], ymm0);
// Extract LLR of 1st symbol
llr32[0] = _mm256_extract_epi32(llr256[0], 0);
llr32[1] = _mm256_extract_epi32(llr256[0], 1);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx0 = (int16_t *)&llr32[0];
int16_t *llr16avx1 = (int16_t *)&llr32[1];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 0), (short)_mm256_extract_epi16(rxF[i], 1),
llr16avx0[0], llr16avx0[1], llr16avx1[0], llr16avx1[1]);
#endif
// Extract LLR of 2nd symbol
llr32[2] = _mm256_extract_epi32(llr256[0], 2);
llr32[3] = _mm256_extract_epi32(llr256[0], 3);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx2 = (int16_t *)&llr32[2];
int16_t *llr16avx3 = (int16_t *)&llr32[3];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 2), (short)_mm256_extract_epi16(rxF[i], 3),
llr16avx2[0], llr16avx2[1], llr16avx3[0], llr16avx3[1]);
#endif
// Extract LLR of 3rd symbol
llr32[4] = _mm256_extract_epi32(llr256[1], 0);
llr32[5] = _mm256_extract_epi32(llr256[1], 1);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx4 = (int16_t *)&llr32[4];
int16_t *llr16avx5 = (int16_t *)&llr32[5];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 4), (short)_mm256_extract_epi16(rxF[i], 5),
llr16avx4[0], llr16avx4[1], llr16avx5[0], llr16avx5[1]);
#endif
// Extract LLR of 4th symbol
llr32[6] = _mm256_extract_epi32(llr256[1], 2);
llr32[7] = _mm256_extract_epi32(llr256[1], 3);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx6 = (int16_t *)&llr32[6];
int16_t *llr16avx7 = (int16_t *)&llr32[7];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 6), (short)_mm256_extract_epi16(rxF[i], 7),
llr16avx6[0], llr16avx6[1], llr16avx7[0], llr16avx7[1]);
#endif
// Extract LLR of 5th symbol
llr32[8] = _mm256_extract_epi32(llr256[0], 4);
llr32[9] = _mm256_extract_epi32(llr256[0], 5);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx8 = (int16_t *)&llr32[8];
int16_t *llr16avx9 = (int16_t *)&llr32[9];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 8), (short)_mm256_extract_epi16(rxF[i], 9),
llr16avx8[0], llr16avx8[1], llr16avx9[0], llr16avx9[1]);
#endif
// Extract LLR of 6th symbol
llr32[10] = _mm256_extract_epi32(llr256[0], 6);
llr32[11] = _mm256_extract_epi32(llr256[0], 7);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx10 = (int16_t *)&llr32[10];
int16_t *llr16avx11 = (int16_t *)&llr32[11];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 10), (short)_mm256_extract_epi16(rxF[i], 11),
llr16avx10[0], llr16avx10[1], llr16avx11[0], llr16avx11[1]);
#endif
// Extract LLR of 7th symbol
llr32[12] = _mm256_extract_epi32(llr256[1], 4);
llr32[13] = _mm256_extract_epi32(llr256[1], 5);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx12 = (int16_t *)&llr32[12];
int16_t *llr16avx13 = (int16_t *)&llr32[13];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 12), (short)_mm256_extract_epi16(rxF[i], 13),
llr16avx12[0], llr16avx12[1], llr16avx13[0], llr16avx13[1]);
#endif
// Extract LLR of 8th symbol
llr32[14] = _mm256_extract_epi32(llr256[1], 6);
llr32[15] = _mm256_extract_epi32(llr256[1], 7);
#ifdef DEBUG_AVX_LLR
int16_t *llr16avx14 = (int16_t *)&llr32[14];
int16_t *llr16avx15 = (int16_t *)&llr32[15];
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d] \n",
(short)_mm256_extract_epi16(rxF[i], 14), (short)_mm256_extract_epi16(rxF[i], 15),
llr16avx14[0], llr16avx14[1], llr16avx15[0], llr16avx15[1]);
#endif
llr32 += 16;
#else
xmm0 = _mm_abs_epi16(rxF[i]);
xmm0 = _mm_subs_epi16(ch_mag[i],xmm0);
......@@ -749,6 +858,7 @@ void nr_dlsch_16qam_llr(NR_DL_FRAME_PARMS *frame_parms,
llr32[6] = _mm_extract_epi32(llr128[1],2); //((uint32_t *)&llr128[1])[2];
llr32[7] = _mm_extract_epi32(llr128[1],3); //((uint32_t *)&llr128[1])[3];
llr32+=8;
#endif
#elif defined(__arm__)
xmm0 = vabsq_s16(rxF[i]);
xmm0 = vqsubq_s16(ch_mag[i],xmm0);
......@@ -798,14 +908,20 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
uint8_t beamforming_mode)
{
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
printf("DLSCH 64QAM AVX\n");
__m256i *rxF = (__m256i *)&rxdataF_comp[0][(symbol*nb_rb*12)];
__m256i *ch_mag,*ch_magb;
register __m256i ymm1,ymm2;
#else
__m128i *rxF = (__m128i*)&rxdataF_comp[0][(symbol*nb_rb*12)];
__m128i *ch_mag,*ch_magb;
#endif
#elif defined(__arm__)
int16x8_t *rxF = (int16x8_t*)&rxdataF_comp[0][(symbol*nb_rb*12)];
int16x8_t *ch_mag,*ch_magb,xmm1,xmm2;
#endif
int i,len2;
unsigned char len_mod4;
short *llr;
int16_t *llr2;
int8_t *pllr_symbol;
......@@ -822,8 +938,13 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
pllr_symbol += llr_offset;
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
ch_mag = (__m256i*)&dl_ch_mag[0][(symbol*nb_rb*12)];
ch_magb = (__m256i*)&dl_ch_magb[0][(symbol*nb_rb*12)];
#else
ch_mag = (__m128i*)&dl_ch_mag[0][(symbol*nb_rb*12)];
ch_magb = (__m128i*)&dl_ch_magb[0][(symbol*nb_rb*12)];
#endif
#elif defined(__arm__)
ch_mag = (int16x8_t*)&dl_ch_mag[0][(symbol*nb_rb*12)];
ch_magb = (int16x8_t*)&dl_ch_magb[0][(symbol*nb_rb*12)];
......@@ -840,18 +961,30 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
llr2 = llr;
llr += (len*6);
len_mod4 =len&3;
#ifdef __AVX2__
unsigned char len_mod8 = len&7;
len2 = len>>3; // length in quad words (4 REs)
len2 += (len_mod8 == 0 ? 0 : 1);
#else
unsigned char len_mod4 =len&3;
len2=len>>2; // length in quad words (4 REs)
len2+=((len_mod4==0)?0:1);
#endif
for (i=0; i<len2; i++) {
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
ymm1 = _mm256_abs_epi16(rxF[i]);
ymm1 = _mm256_subs_epi16(ch_mag[i], ymm1);
ymm2 = _mm256_abs_epi16(ymm1);
ymm2 = _mm256_subs_epi16(ch_magb[i], ymm2);
#else
xmm1 = _mm_abs_epi16(rxF[i]);
xmm1 = _mm_subs_epi16(ch_mag[i],xmm1);
xmm2 = _mm_abs_epi16(xmm1);
xmm2 = _mm_subs_epi16(ch_magb[i],xmm2);
#endif
#elif defined(__arm__)
xmm1 = vabsq_s16(rxF[i]);
xmm1 = vsubq_s16(ch_mag[i],xmm1);
......@@ -871,13 +1004,28 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
llr2+=6;
}
*/
// Extract LLR of 1st symbol
llr2[0] = ((short *)&rxF[i])[0];
llr2[1] = ((short *)&rxF[i])[1];
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
llr2[2] = _mm256_extract_epi16(ymm1, 0);
llr2[3] = _mm256_extract_epi16(ymm1, 1);
llr2[4] = _mm256_extract_epi16(ymm2, 0);
llr2[5] = _mm256_extract_epi16(ymm2, 1);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 0),
(short)_mm256_extract_epi16(rxF[i], 1),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
#else
llr2[2] = _mm_extract_epi16(xmm1,0);
llr2[3] = _mm_extract_epi16(xmm1,1);//((short *)&xmm1)[j+1];
llr2[4] = _mm_extract_epi16(xmm2,0);//((short *)&xmm2)[j];
llr2[5] = _mm_extract_epi16(xmm2,1);//((short *)&xmm2)[j+1];
#endif
#elif defined(__arm__)
llr2[2] = vgetq_lane_s16(xmm1,0);
llr2[3] = vgetq_lane_s16(xmm1,1);//((short *)&xmm1)[j+1];
......@@ -886,13 +1034,28 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
#endif
llr2+=6;
// Extract LLR of 2nd symbol
llr2[0] = ((short *)&rxF[i])[2];
llr2[1] = ((short *)&rxF[i])[3];
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
llr2[2] = _mm256_extract_epi16(ymm1, 2);
llr2[3] = _mm256_extract_epi16(ymm1, 3);
llr2[4] = _mm256_extract_epi16(ymm2, 2);
llr2[5] = _mm256_extract_epi16(ymm2, 3);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 2),
(short)_mm256_extract_epi16(rxF[i], 3),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
#else
llr2[2] = _mm_extract_epi16(xmm1,2);
llr2[3] = _mm_extract_epi16(xmm1,3);//((short *)&xmm1)[j+1];
llr2[4] = _mm_extract_epi16(xmm2,2);//((short *)&xmm2)[j];
llr2[5] = _mm_extract_epi16(xmm2,3);//((short *)&xmm2)[j+1];
#endif
#elif defined(__arm__)
llr2[2] = vgetq_lane_s16(xmm1,2);
llr2[3] = vgetq_lane_s16(xmm1,3);//((short *)&xmm1)[j+1];
......@@ -901,13 +1064,28 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
#endif
llr2+=6;
// Extract LLR of 3rd symbol
llr2[0] = ((short *)&rxF[i])[4];
llr2[1] = ((short *)&rxF[i])[5];
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
llr2[2] = _mm256_extract_epi16(ymm1, 4);
llr2[3] = _mm256_extract_epi16(ymm1, 5);
llr2[4] = _mm256_extract_epi16(ymm2, 4);
llr2[5] = _mm256_extract_epi16(ymm2, 5);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 4),
(short)_mm256_extract_epi16(rxF[i], 5),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
#else
llr2[2] = _mm_extract_epi16(xmm1,4);
llr2[3] = _mm_extract_epi16(xmm1,5);//((short *)&xmm1)[j+1];
llr2[4] = _mm_extract_epi16(xmm2,4);//((short *)&xmm2)[j];
llr2[5] = _mm_extract_epi16(xmm2,5);//((short *)&xmm2)[j+1];
#endif
#elif defined(__arm__)
llr2[2] = vgetq_lane_s16(xmm1,4);
llr2[3] = vgetq_lane_s16(xmm1,5);//((short *)&xmm1)[j+1];
......@@ -915,13 +1093,28 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
llr2[5] = vgetq_lane_s16(xmm2,5);//((short *)&xmm2)[j+1];
#endif
llr2+=6;
// Extract LLR of 4th symbol
llr2[0] = ((short *)&rxF[i])[6];
llr2[1] = ((short *)&rxF[i])[7];
#if defined(__x86_64__) || defined(__i386__)
#ifdef __AVX2__
llr2[2] = _mm256_extract_epi16(ymm1, 6);
llr2[3] = _mm256_extract_epi16(ymm1, 7);
llr2[4] = _mm256_extract_epi16(ymm2, 6);
llr2[5] = _mm256_extract_epi16(ymm2, 7);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 6),
(short)_mm256_extract_epi16(rxF[i], 7),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
#else
llr2[2] = _mm_extract_epi16(xmm1,6);
llr2[3] = _mm_extract_epi16(xmm1,7);//((short *)&xmm1)[j+1];
llr2[4] = _mm_extract_epi16(xmm2,6);//((short *)&xmm2)[j];
llr2[5] = _mm_extract_epi16(xmm2,7);//((short *)&xmm2)[j+1];
#endif
#elif defined(__arm__)
llr2[2] = vgetq_lane_s16(xmm1,6);
llr2[3] = vgetq_lane_s16(xmm1,7);//((short *)&xmm1)[j+1];
......@@ -930,6 +1123,64 @@ void nr_dlsch_64qam_llr(NR_DL_FRAME_PARMS *frame_parms,
#endif
llr2+=6;
#ifdef __AVX__
// Extract LLR of 5th symbol
llr2[0] = ((short *)&rxF[i])[8];
llr2[1] = ((short *)&rxF[i])[9];
llr2[2] = _mm256_extract_epi16(ymm1, 8);
llr2[3] = _mm256_extract_epi16(ymm1, 9);
llr2[4] = _mm256_extract_epi16(ymm2, 8);
llr2[5] = _mm256_extract_epi16(ymm2, 9);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 8), (short)_mm256_extract_epi16(rxF[i], 9),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
llr2 += 6;
// Extract LLR of 6th symbol
llr2[0] = ((short *)&rxF[i])[10];
llr2[1] = ((short *)&rxF[i])[11];
llr2[2] = _mm256_extract_epi16(ymm1, 10);
llr2[3] = _mm256_extract_epi16(ymm1, 11);
llr2[4] = _mm256_extract_epi16(ymm2, 10);
llr2[5] = _mm256_extract_epi16(ymm2, 11);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 10), (short)_mm256_extract_epi16(rxF[i], 11),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
llr2 += 6;
// Extract LLR of 7th symbol
llr2[0] = ((short *)&rxF[i])[12];
llr2[1] = ((short *)&rxF[i])[13];
llr2[2] = _mm256_extract_epi16(ymm1, 12);
llr2[3] = _mm256_extract_epi16(ymm1, 13);
llr2[4] = _mm256_extract_epi16(ymm2, 12);
llr2[5] = _mm256_extract_epi16(ymm2, 13);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 12), (short)_mm256_extract_epi16(rxF[i], 13),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
llr2 += 6;
// Extract LLR of 8th symbol
llr2[0] = ((short *)&rxF[i])[14];
llr2[1] = ((short *)&rxF[i])[15];
llr2[2] = _mm256_extract_epi16(ymm1, 14);
llr2[3] = _mm256_extract_epi16(ymm1, 15);
llr2[4] = _mm256_extract_epi16(ymm2, 14);
llr2[5] = _mm256_extract_epi16(ymm2, 15);
#ifdef DEBUG_AVX_LLR
printf("llr of symbol (%d, %d) = [%d, %d, %d, %d, %d, %d]\n",
(short)_mm256_extract_epi16(rxF[i], 14), (short)_mm256_extract_epi16(rxF[i], 15),
llr2[0], llr2[1], llr2[2],
llr2[3], llr2[4], llr2[5]);
#endif
llr2 += 6;
#endif
}
#if defined(__x86_64__) || defined(__i386__)
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment