Commit 9db8db25 authored by Raymond Knopp's avatar Raymond Knopp

LUT memory access fix for code generator

parent e785891f
...@@ -95,7 +95,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -95,7 +95,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
{ {
// Abs and sign of 32 CNs (first BN) // Abs and sign of 32 CNs (first BN)
// ymm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i]; // ymm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",lut_startAddrCnGroups[0]+lut_idxCnProcG3[j][0] + i); fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",(lut_startAddrCnGroups[0]>>5)+lut_idxCnProcG3[j][0] + i);
// sgn = _mm256_sign_epi8(ones, ymm0); // sgn = _mm256_sign_epi8(ones, ymm0);
fprintf(fd," sgn = _mm256_sign_epi8(ones, ymm0);\n"); fprintf(fd," sgn = _mm256_sign_epi8(ones, ymm0);\n");
// min = _mm256_abs_epi8(ymm0); // min = _mm256_abs_epi8(ymm0);
...@@ -103,7 +103,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -103,7 +103,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
// 32 CNs of second BN // 32 CNs of second BN
// ymm0 = p_cnProcBuf[lut_idxCnProcG3[j][1] + i]; // ymm0 = p_cnProcBuf[lut_idxCnProcG3[j][1] + i];
fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",lut_startAddrCnGroups[0]+lut_idxCnProcG3[j][1] + i); fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",(lut_startAddrCnGroups[0]>>5)+lut_idxCnProcG3[j][1] + i);
// min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0)); // min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0));
fprintf(fd," min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0));\n"); fprintf(fd," min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0));\n");
...@@ -116,7 +116,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -116,7 +116,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[0]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",(lut_startAddrCnGroups[0]>>5)+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -153,7 +153,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -153,7 +153,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
{ {
// Abs and sign of 32 CNs (first BN) // Abs and sign of 32 CNs (first BN)
// ymm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i]; // ymm0 = p_cnProcBuf[lut_idxCnProcG3[j][0] + i];
fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",lut_startAddrCnGroups[1]+lut_idxCnProcG4[j][0] + i); fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",(lut_startAddrCnGroups[1]>>5)+lut_idxCnProcG4[j][0] + i);
// sgn = _mm256_sign_epi8(ones, ymm0); // sgn = _mm256_sign_epi8(ones, ymm0);
fprintf(fd," sgn = _mm256_sign_epi8(ones, ymm0);\n"); fprintf(fd," sgn = _mm256_sign_epi8(ones, ymm0);\n");
// min = _mm256_abs_epi8(ymm0); // min = _mm256_abs_epi8(ymm0);
...@@ -163,7 +163,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -163,7 +163,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
// Loop over BNs // Loop over BNs
for (k=1; k<3; k++) for (k=1; k<3; k++)
{ {
fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",lut_startAddrCnGroups[1]+lut_idxCnProcG4[j][k] + i); fprintf(fd," ymm0 = ((__m256i*)&cnProcBuf)[%d];\n",(lut_startAddrCnGroups[1]>>5)+lut_idxCnProcG4[j][k] + i);
// min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0)); // min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0));
fprintf(fd," min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0));\n"); fprintf(fd," min = _mm256_min_epu8(min, _mm256_abs_epi8(ymm0));\n");
...@@ -177,7 +177,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -177,7 +177,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[1]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",(lut_startAddrCnGroups[1]>>5)+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -241,7 +241,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -241,7 +241,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[2]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[2]+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -305,7 +305,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -305,7 +305,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[3]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[3]+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -372,7 +372,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -372,7 +372,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[4]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[4]+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -440,7 +440,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -440,7 +440,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[5]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[5]+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -509,7 +509,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -509,7 +509,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[6]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[6]+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -578,7 +578,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -578,7 +578,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[7]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[7]+(j*bitOffsetInGroup)+i);
} }
} }
} }
...@@ -649,7 +649,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R) ...@@ -649,7 +649,7 @@ void nrLDPC_cnProc_BG1_generator(uint16_t Z,int R)
fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n"); fprintf(fd," min = _mm256_min_epu8(min, maxLLR);\n");
// *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn); // *p_cnProcBufResBit = _mm256_sign_epi8(min, sgn);
// p_cnProcBufResBit++; // p_cnProcBufResBit++;
fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_numCnInCnGroups[8]+(j*bitOffsetInGroup)+i); fprintf(fd," ((__m256i*)cnProcBufRes)[%d] = _mm256_sign_epi8(min, sgn);\n",lut_startAddrCnGroups[8]+(j*bitOffsetInGroup)+i);
} }
} }
} }
......
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