maki/openairinterface5g
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add Intrinsics usage by SIMDE

Browse Source
This commit is contained in:
Bruno Mongazon-Cazavet
2022-09-28 18:21:19 +02:00
committed by Robert Schmidt
parent a1de5e3d11
commit fe1dc24e28
114 changed files with 7900 additions and 11031 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openair1/PHY/CODING/nrLDPC_encoder/ldpc_encoder_optim8segmulti.c
+18 -28
View File
@@ -37,6 +37,7 @@
#include "common/utils/LOG/log.h"
#include "time_meas.h"
#include "openair1/PHY/CODING/nrLDPC_defs.h"
#include "PHY/sse_intrin.h"
#include "ldpc_encode_parity_check.c"
#include "ldpc_generate_coefficient.c"
@@ -61,21 +62,19 @@ int nrLDPC_encod(unsigned char **input,unsigned char **output,int Zc,int Kb,shor
///printf("macro_segment: %d\n", macro_segment);
///printf("macro_segment_end: %d\n", macro_segment_end );
#ifdef __AVX2__
__m256i shufmask = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,0x0101010101010101, 0x0000000000000000);
__m256i andmask = _mm256_set1_epi64x(0x0102040810204080); // every 8 bits -> 8 bytes, pattern repeats.
__m256i zero256 = _mm256_setzero_si256();
__m256i shufmask = simde_mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,0x0101010101010101, 0x0000000000000000);
__m256i andmask = simde_mm256_set1_epi64x(0x0102040810204080); // every 8 bits -> 8 bytes, pattern repeats.
__m256i zero256 = simde_mm256_setzero_si256();
__m256i masks[8];
register __m256i c256;
masks[0] = _mm256_set1_epi8(0x1);
masks[1] = _mm256_set1_epi8(0x2);
masks[2] = _mm256_set1_epi8(0x4);
masks[3] = _mm256_set1_epi8(0x8);
masks[4] = _mm256_set1_epi8(0x10);
masks[5] = _mm256_set1_epi8(0x20);
masks[6] = _mm256_set1_epi8(0x40);
masks[7] = _mm256_set1_epi8(0x80);
#endif
masks[0] = simde_mm256_set1_epi8(0x1);
masks[1] = simde_mm256_set1_epi8(0x2);
masks[2] = simde_mm256_set1_epi8(0x4);
masks[3] = simde_mm256_set1_epi8(0x8);
masks[4] = simde_mm256_set1_epi8(0x10);
masks[5] = simde_mm256_set1_epi8(0x20);
masks[6] = simde_mm256_set1_epi8(0x40);
masks[7] = simde_mm256_set1_epi8(0x80);
@@ -127,12 +126,11 @@ int nrLDPC_encod(unsigned char **input,unsigned char **output,int Zc,int Kb,shor
}
}
#else
#ifdef __AVX2__
for (int i=0; i<block_length>>5; i++) {
c256 = _mm256_and_si256(_mm256_cmpeq_epi8(_mm256_andnot_si256(_mm256_shuffle_epi8(_mm256_set1_epi32(((uint32_t*)input[macro_segment])[i]), shufmask),andmask),zero256),masks[0]);
c256 = simde_mm256_and_si256(simde_mm256_cmpeq_epi8(simde_mm256_andnot_si256(simde_mm256_shuffle_epi8(simde_mm256_set1_epi32(((uint32_t*)input[macro_segment])[i]), shufmask),andmask),zero256),masks[0]);
//for (j=1; j<n_segments; j++) {
for (int j=macro_segment+1; j < macro_segment_end; j++) {
c256 = _mm256_or_si256(_mm256_and_si256(_mm256_cmpeq_epi8(_mm256_andnot_si256(_mm256_shuffle_epi8(_mm256_set1_epi32(((uint32_t*)input[j])[i]), shufmask),andmask),zero256),masks[j-macro_segment]),c256);
c256 = simde_mm256_or_si256(simde_mm256_and_si256(simde_mm256_cmpeq_epi8(simde_mm256_andnot_si256(simde_mm256_shuffle_epi8(simde_mm256_set1_epi32(((uint32_t*)input[j])[i]), shufmask),andmask),zero256),masks[j-macro_segment]),c256);
}
((__m256i *)cc)[i] = c256;
}
@@ -146,9 +144,6 @@ int nrLDPC_encod(unsigned char **input,unsigned char **output,int Zc,int Kb,shor
cc[i] |= (temp << (j-macro_segment));
}
}
#else
AssertFatal(1==0,"Need AVX2 for this\n");
#endif
#endif
if(impp->tinput != NULL) stop_meas(impp->tinput);
@@ -174,7 +169,6 @@ int nrLDPC_encod(unsigned char **input,unsigned char **output,int Zc,int Kb,shor
memcpy(&output[0], &c[2*Zc], (block_length-2*Zc)*sizeof(unsigned char));
memcpy(&output[block_length-2*Zc], &d[0], ((nrows-no_punctured_columns) * Zc-removed_bit)*sizeof(unsigned char));
*/
#ifdef __AVX2__
if ((((2*Zc)&31) == 0) && (((block_length-(2*Zc))&31) == 0)) {
//AssertFatal(((2*Zc)&31) == 0,"2*Zc needs to be a multiple of 32 for now\n");
//AssertFatal(((block_length-(2*Zc))&31) == 0,"block_length-(2*Zc) needs to be a multiple of 32 for now\n");
@@ -185,15 +179,15 @@ int nrLDPC_encod(unsigned char **input,unsigned char **output,int Zc,int Kb,shor
// if (((block_length-(2*Zc))&31)>0) l1++;
for (int i=0;i<l1;i++)
//for (j=0;j<n_segments;j++) ((__m256i *)output[j])[i] = _mm256_and_si256(_mm256_srai_epi16(c256p[i],j),masks[0]);
for (int j=macro_segment; j < macro_segment_end; j++) ((__m256i *)output[j])[i] = _mm256_and_si256(_mm256_srai_epi16(c256p[i],j-macro_segment),masks[0]);
//for (j=0;j<n_segments;j++) ((__m256i *)output[j])[i] = simde_mm256_and_si256(simde_mm256_srai_epi16(c256p[i],j),masks[0]);
for (int j=macro_segment; j < macro_segment_end; j++) ((__m256i *)output[j])[i] = simde_mm256_and_si256(simde_mm256_srai_epi16(c256p[i],j-macro_segment),masks[0]);
// if ((((nrows-no_punctured_columns) * Zc-removed_bit)&31)>0) l2++;
for (int i1=0, i=l1;i1<l2;i1++,i++)
//for (j=0;j<n_segments;j++) ((__m256i *)output[j])[i] = _mm256_and_si256(_mm256_srai_epi16(d256p[i1],j),masks[0]);
for (int j=macro_segment; j < macro_segment_end; j++) ((__m256i *)output[j])[i] = _mm256_and_si256(_mm256_srai_epi16(d256p[i1],j-macro_segment),masks[0]);
//for (j=0;j<n_segments;j++) ((__m256i *)output[j])[i] = simde_mm256_and_si256(simde_mm256_srai_epi16(d256p[i1],j),masks[0]);
for (int j=macro_segment; j < macro_segment_end; j++) ((__m256i *)output[j])[i] = simde_mm256_and_si256(simde_mm256_srai_epi16(d256p[i1],j-macro_segment),masks[0]);
}
else {
#ifdef DEBUG_LDPC
@@ -210,10 +204,6 @@ int nrLDPC_encod(unsigned char **input,unsigned char **output,int Zc,int Kb,shor
output[j][block_length-2*Zc+i] = (dd[i]>>(j-macro_segment))&1;
}
#else
AssertFatal(1==0,"Need AVX2 for now\n");
#endif
if(impp->toutput != NULL) stop_meas(impp->toutput);
return 0;
}