Merge branch 'make-convs-A-stride' of https://github.com/ucb-bar/gemmini-rocc-tests into make-convs-A-stride

Author: hngenc

bareMetalC/conv.c

@@ -8,14 +8,6 @@
 #endif
 #include "include/gemmini_testutils.h"
 
-// #define BATCH_SIZE 4
-// #define IN_DIM 224
-// #define IN_CHANNELS 3
-// #define OUT_CHANNELS 32
-// #define KERNEL_DIM 3
-// #define PADDING 1
-// #define STRIDE 2
-
 #ifndef BAREMETAL
 
 #define BATCH_SIZE 4

bareMetalC/conv_with_pool.c

@@ -8,18 +8,6 @@
 #endif
 #include "include/gemmini_testutils.h"
 
-// #define BATCH_SIZE 4
-// #define IN_DIM 224
-// #define IN_CHANNELS 3
-// #define OUT_CHANNELS 32
-// #define KERNEL_DIM 3
-// #define PADDING 1
-// #define STRIDE 2
-// 
-// #define POOL_SIZE 3
-// #define POOL_STRIDE 2
-// #define POOL_PADDING 1
-
 #ifndef BAREMETAL
 
 #define BATCH_SIZE 4
@@ -396,7 +384,6 @@ int main() {
     uint64_t start_gemmini = read_cycles();
 
     tiled_conv_A_stride_auto(
-    // tiled_conv(
         BATCH_SIZE, IN_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_DIM,
         STRIDE, PADDING, KERNEL_DIM,

include/gemmini.h

@@ -1079,9 +1079,7 @@ void sp_tiled_conv_A_stride(
                 for (int ocol = 0; ocol < ocols; ocol += DIM) {
                     const int I = ocols - ocol > DIM ? DIM : ocols - ocol;
 
-                    // for (int och = 0; och < ochs; och += DIM) {
                     for (int och = 0; och < ochs; och += max_ochs_per_mvin) {
-                        // const int J = ochs - och > DIM ? DIM : ochs - och;
                         const int J = ochs - och > max_ochs_per_mvin ? max_ochs_per_mvin : ochs - och;
 
                         const uint32_t D_sp_addr = D_sp_addr_start + (och / DIM) * batches * orows * ocols + b * orows * ocols + orow * ocols + ocol;
@@ -1145,9 +1143,6 @@ void sp_tiled_conv_A_stride(
 
         gemmini_extended4_config_ld(out_channels * sizeof(elem_t), MVIN_SCALE_IDENTITY, false, krows * kcols * kchs, 1);
 
-        // for (int och = 0; och < ochs; och += DIM) {
-        //     const int J = ochs - och > DIM ? DIM : ochs - och;
-
         for (int och = 0; och < ochs; och += max_ochs_per_mvin) {
             const int J = ochs - och > max_ochs_per_mvin ?
                 max_ochs_per_mvin : ochs - och;
@@ -1157,9 +1152,6 @@ void sp_tiled_conv_A_stride(
                     for (int kch = 0; kch < kchs; kch += DIM) {
                         const int K = kchs - kch > DIM ? DIM : kchs - kch;
 
-                    // for (int kch = 0; kch < kchs; kch += max_kchs_per_mvin) {
-                    //     const int K = kchs - kch > max_kchs_per_mvin ? max_kchs_per_mvin : kchs - kch;
-
                         const uint32_t B_sp_addr = B_sp_addr_start + (och / DIM) * krows * kcols * kchs + krow * kcols * kchs + kcol * kchs + kch;
 
                         gemmini_extended_mvin2(weights + (krow*kernel_dim*in_channels + kcol*in_channels + kch) * out_channels + och,
@@ -1183,7 +1175,6 @@ void sp_tiled_conv_A_stride(
                     for (int krow = 0; krow < krows; krow++) {
                         const int irow = orow * stride + krow;
 
-                        // for (int kcol = 0; kcol < kcols; kcol++) {
                         for (int kcol = 0; kcol < kcols; kcol++) {
                             const int icol = ocol * stride + kcol;