New files (#852)

* Adding new files for gpu setup * Pointing to right files * Minor correction when calculating buffer positions * Updating goldilocks * Transpose tmp polinomials is now optional --------- Co-authored-by: Ricard Borrell <75077385+rickb80@users.noreply.github.com>
0xPolygonHermez · Jul 9, 2024 · f39b263 · f39b263
1 parent 1fc5cf6
commit f39b263
Show file tree

Hide file tree

Showing 33 changed files with 1,285 additions and 1,299 deletions.
diff --git a/.github/workflows/prover_e2e.yml b/.github/workflows/prover_e2e.yml
@@ -28,9 +28,9 @@ jobs:
     - name: remove old proofs
       run: rm -rf runtime/output/*
     - name: list gha config files
-      run: ls /home/gha/v7.0.0-rc.1-fork.10/config/
+      run: ls /home/gha/v7.0.0-rc.14-fork.10/config/
     - name: link files
-      run: ln -s /home/gha/v7.0.0-rc.1-fork.10/config/ config
+      run: ln -s /home/gha/v7.0.0-rc.14-fork.10/config/ config
     - name: list local config files
       run: ls config
     - name: run collection native tests

diff --git a/src/config/definitions.hpp b/src/config/definitions.hpp
@@ -95,6 +95,8 @@
 
 #define NROWS_PACK 4
 
+#define TRANSPOSE_TMP_POLS true
+
 //#define MULTI_ROM_TEST
 
 //#define ENABLE_EXPERIMENTAL_CODE

diff --git a/src/goldilocks b/src/goldilocks
diff --git a/src/starkpil/chelpers_steps.hpp b/src/starkpil/chelpers_steps.hpp
@@ -49,16 +49,16 @@ class CHelpersSteps {
                 }
             }
         }
-        nColsStagesAcc[10] = nColsStagesAcc[9] + nColsStages[9]; // Polinomials f & q
+        nColsStagesAcc[10] = nColsStagesAcc[9] + nColsStages[4]; // Polinomials f & q
         if(stage == 4) {
             offsetsStages[10] = starkInfo.mapOffsets.section[eSection::q_2ns];
             nColsStages[10] = starkInfo.qDim;
         } else if(stage == 5) {
             offsetsStages[10] = starkInfo.mapOffsets.section[eSection::f_2ns];
             nColsStages[10] = 3;
         }
-        nColsStagesAcc[11] = nColsStagesAcc[10] + 3; // xDivXSubXi
-        nCols = nColsStagesAcc[11] + 6;
+        nColsStagesAcc[11] = nColsStagesAcc[10] + nColsStages[10]; // xDivXSubXi
+        nCols = nColsStagesAcc[11] + 6; // 3 for xDivXSubXi and 3 for xDivXSubWxi
     }
 
     inline virtual void storePolinomials(StarkInfo &starkInfo, StepsParams &params, __m256i *bufferT_, uint8_t* storePol, uint64_t row, uint64_t nrowsPack, uint64_t domainExtended) {

diff --git a/src/starkpil/chelpers_steps_avx512.hpp b/src/starkpil/chelpers_steps_avx512.hpp
@@ -49,16 +49,16 @@ class CHelpersStepsAvx512 : public CHelpersSteps {
                 }
             }
         }
-        nColsStagesAcc[10] = nColsStagesAcc[9] + nColsStages[9]; // Polinomials f & q
+        nColsStagesAcc[10] = nColsStagesAcc[9] + nColsStages[4]; // Polinomials f & q
         if(stage == 4) {
             offsetsStages[10] = starkInfo.mapOffsets.section[eSection::q_2ns];
             nColsStages[10] = starkInfo.qDim;
         } else if(stage == 5) {
             offsetsStages[10] = starkInfo.mapOffsets.section[eSection::f_2ns];
             nColsStages[10] = 3;
         }
-        nColsStagesAcc[11] = nColsStagesAcc[10] + 3; // xDivXSubXi
-        nCols = nColsStagesAcc[11] + 6;
+        nColsStagesAcc[11] = nColsStagesAcc[10] + nColsStages[10]; // xDivXSubXi
+        nCols = nColsStagesAcc[11] + 6; // 3 for xDivXSubXi and 3 for xDivXSubWxi
     }
 
     inline virtual void storePolinomials(StarkInfo &starkInfo, StepsParams &params, __m512i *bufferT_, uint8_t* storePol, uint64_t row, uint64_t nrowsPack, uint64_t domainExtended) {

diff --git a/src/starkpil/chelpers_steps_pack.hpp b/src/starkpil/chelpers_steps_pack.hpp
@@ -51,16 +51,16 @@ class CHelpersStepsPack : public CHelpersSteps {
                 }
             }
         }
-        nColsStagesAcc[10] = nColsStagesAcc[9] + nColsStages[9]; // Polinomials f & q
+        nColsStagesAcc[10] = nColsStagesAcc[9] + nColsStages[4]; // Polinomials f & q
         if(stage == 4) {
             offsetsStages[10] = starkInfo.mapOffsets.section[eSection::q_2ns];
             nColsStages[10] = starkInfo.qDim;
         } else if(stage == 5) {
             offsetsStages[10] = starkInfo.mapOffsets.section[eSection::f_2ns];
             nColsStages[10] = 3;
         }
-        nColsStagesAcc[11] = nColsStagesAcc[10] + 3; // xDivXSubXi
-        nCols = nColsStagesAcc[11] + 6;
+        nColsStagesAcc[11] = nColsStagesAcc[10] + nColsStages[10]; // xDivXSubXi
+        nCols = nColsStagesAcc[11] + 6; // 3 for xDivXSubXi and 3 for xDivXSubWxi
     }
 
     inline virtual void storePolinomials(StarkInfo &starkInfo, StepsParams &params, Goldilocks::Element *bufferT_, uint8_t* storePol, uint64_t row, uint64_t nrowsPack, uint64_t domainExtended) {
@@ -77,15 +77,22 @@ class CHelpersStepsPack : public CHelpersSteps {
                 bool isTmpPol = !domainExtended && s == 4;
                 for(uint64_t k = 0; k < nColsStages[s]; ++k) {
                     uint64_t dim = storePol[nColsStagesAcc[s] + k];
-                    if(storePol[nColsStagesAcc[s] + k]) {
-                        Goldilocks::Element *buffT = &bufferT_[(nColsStagesAcc[s] + k)* nrowsPack];
-                        if(isTmpPol) {
-                            for(uint64_t i = 0; i < dim; ++i) {
-                                Goldilocks::copy_pack(nrowsPack, &params.pols[offsetsStages[s] + k * domainSize + row * dim + i], uint64_t(dim), &buffT[i*nrowsPack]);
-                            }
-                        } else {
+                    if(!TRANSPOSE_TMP_POLS) {
+                        for(uint64_t k = 0; k < nColsStages[s]; ++k) {
+                            Goldilocks::Element *buffT = &bufferT_[(nColsStagesAcc[s] + k)* nrowsPack];
                             Goldilocks::copy_pack(nrowsPack, &params.pols[offsetsStages[s] + k + row * nColsStages[s]], nColsStages[s], buffT);
                         }
+                    } else {
+                        if(storePol[nColsStagesAcc[s] + k]) {
+                            Goldilocks::Element *buffT = &bufferT_[(nColsStagesAcc[s] + k)* nrowsPack];
+                            if(isTmpPol) {
+                                for(uint64_t i = 0; i < dim; ++i) {
+                                    Goldilocks::copy_pack(nrowsPack, &params.pols[offsetsStages[s] + k * domainSize + row * dim + i], uint64_t(dim), &buffT[i*nrowsPack]);
+                                }
+                            } else {
+                                Goldilocks::copy_pack(nrowsPack, &params.pols[offsetsStages[s] + k + row * nColsStages[s]], nColsStages[s], buffT);
+                            }
+                        }
                     }
                 }
             }