diff --git a/bareMetalC/conv.c b/bareMetalC/conv.c
index 4ace514a..82bb32cd 100644
--- a/bareMetalC/conv.c
+++ b/bareMetalC/conv.c
@@ -228,7 +228,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_dw.c b/bareMetalC/conv_dw.c
index a0b00217..dcfa646a 100644
--- a/bareMetalC/conv_dw.c
+++ b/bareMetalC/conv_dw.c
@@ -120,7 +120,7 @@ int main() {
 #ifndef FAST
     tiled_conv_dw_auto(BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM,
             CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-            STRIDE, PADDING, KERNEL_DIM,
+            STRIDE, PADDING, 0, KERNEL_DIM,
 
             (elem_t*)input,
             (elem_t*)weights,
@@ -140,7 +140,7 @@ int main() {
     uint64_t start_gemmini = read_cycles();
     tiled_conv_dw_auto(BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM,
             CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-            STRIDE, PADDING, KERNEL_DIM,
+            STRIDE, PADDING, 0, KERNEL_DIM,
 
             (elem_t*)input,
             (elem_t*)weights,
diff --git a/bareMetalC/conv_dw_perf.c b/bareMetalC/conv_dw_perf.c
index d516cce5..1373d070 100644
--- a/bareMetalC/conv_dw_perf.c
+++ b/bareMetalC/conv_dw_perf.c
@@ -85,7 +85,7 @@ int main (int argc, char * argv[]) {
     uint64_t start_gemmini = read_cycles();
 
     tiled_conv_dw_auto(BATCH_SIZE, IN_DIM, IN_DIM, CHANNELS, OUT_DIM, OUT_DIM,
-        STRIDE, PADDING, KERNEL_DIM,
+        STRIDE, PADDING, 0, KERNEL_DIM,
 
         (elem_t*)input,
         (elem_t*)weights,
diff --git a/bareMetalC/conv_first_layer.c b/bareMetalC/conv_first_layer.c
index 47e3c2af..7303d17e 100644
--- a/bareMetalC/conv_first_layer.c
+++ b/bareMetalC/conv_first_layer.c
@@ -229,7 +229,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_perf.c b/bareMetalC/conv_perf.c
index bacb09c4..a1278f66 100644
--- a/bareMetalC/conv_perf.c
+++ b/bareMetalC/conv_perf.c
@@ -114,7 +114,7 @@ int main (int argc, char * argv[]) {
       tiled_conv_auto(
           BATCH_SIZE, IN_DIM, IN_DIM, IN_CHANNELS,
           OUT_CHANNELS, OUT_DIM, OUT_DIM,
-          STRIDE, 1, 1, PADDING, KERNEL_DIM,
+          STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
           false, false, false, false, false,
 
           (elem_t*)input,
diff --git a/bareMetalC/conv_trans_input_3120.c b/bareMetalC/conv_trans_input_3120.c
index f909682c..8534816e 100644
--- a/bareMetalC/conv_trans_input_3120.c
+++ b/bareMetalC/conv_trans_input_3120.c
@@ -163,7 +163,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, TRANS_INPUT_3120, TRANS_WEIGHT_1203, TRANS_WEIGHT_0132,
 
         (elem_t*)input,
@@ -183,7 +183,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, TRANS_INPUT_3120, TRANS_WEIGHT_1203, TRANS_WEIGHT_0132,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_trans_input_3120_with_kernel_dilation.c b/bareMetalC/conv_trans_input_3120_with_kernel_dilation.c
index e6d45f55..86e07853 100644
--- a/bareMetalC/conv_trans_input_3120_with_kernel_dilation.c
+++ b/bareMetalC/conv_trans_input_3120_with_kernel_dilation.c
@@ -165,7 +165,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, KERNEL_DILATION, PADDING, KERNEL_DIM,
+        STRIDE, 1, KERNEL_DILATION, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, TRANS_INPUT_3120, TRANS_WEIGHT_1203, TRANS_WEIGHT_0132,
 
         (elem_t*)input,
@@ -185,7 +185,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, KERNEL_DILATION, PADDING, KERNEL_DIM,
+        STRIDE, 1, KERNEL_DILATION, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, TRANS_INPUT_3120, TRANS_WEIGHT_1203, TRANS_WEIGHT_0132,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_trans_output_1203.c b/bareMetalC/conv_trans_output_1203.c
index 652da5e2..99b2ec6e 100644
--- a/bareMetalC/conv_trans_output_1203.c
+++ b/bareMetalC/conv_trans_output_1203.c
@@ -160,7 +160,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, false, false, false,
 
         (elem_t*)input,
@@ -180,7 +180,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_trans_weight_0132.c b/bareMetalC/conv_trans_weight_0132.c
index b5bb31f6..7ff1b72c 100644
--- a/bareMetalC/conv_trans_weight_0132.c
+++ b/bareMetalC/conv_trans_weight_0132.c
@@ -162,7 +162,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, false, TRANS_WEIGHT_1203, TRANS_WEIGHT_0132,
 
         (elem_t*)input,
@@ -182,7 +182,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, false, TRANS_WEIGHT_1203, TRANS_WEIGHT_0132,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_trans_weight_1203.c b/bareMetalC/conv_trans_weight_1203.c
index 6d53f944..4564d64a 100644
--- a/bareMetalC/conv_trans_weight_1203.c
+++ b/bareMetalC/conv_trans_weight_1203.c
@@ -161,7 +161,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, false, TRANS_WEIGHT_1203, false,
 
         (elem_t*)input,
@@ -181,7 +181,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, TRANS_OUTPUT_1203, false, TRANS_WEIGHT_1203, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_with_input_dilation.c b/bareMetalC/conv_with_input_dilation.c
index f667fdff..d144827c 100644
--- a/bareMetalC/conv_with_input_dilation.c
+++ b/bareMetalC/conv_with_input_dilation.c
@@ -263,7 +263,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, INPUT_DILATION, 1, PADDING, KERNEL_DIM,
+        STRIDE, INPUT_DILATION, 1, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_with_input_dilation_and_neg_padding.c b/bareMetalC/conv_with_input_dilation_and_neg_padding.c
index 2258618d..dafadfa4 100644
--- a/bareMetalC/conv_with_input_dilation_and_neg_padding.c
+++ b/bareMetalC/conv_with_input_dilation_and_neg_padding.c
@@ -255,7 +255,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, INPUT_DILATION, 1, PADDING, KERNEL_DIM,
+        STRIDE, INPUT_DILATION, 1, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_with_input_dilation_and_rot180.c b/bareMetalC/conv_with_input_dilation_and_rot180.c
index c1fbdf3e..b3ff1ce7 100644
--- a/bareMetalC/conv_with_input_dilation_and_rot180.c
+++ b/bareMetalC/conv_with_input_dilation_and_rot180.c
@@ -269,7 +269,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, INPUT_DILATION, 1, PADDING, KERNEL_DIM,
+        STRIDE, INPUT_DILATION, 1, PADDING, 0, KERNEL_DIM,
         WROT180, false, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_with_kernel_dilation.c b/bareMetalC/conv_with_kernel_dilation.c
index 8ffeab75..d8217eaf 100644
--- a/bareMetalC/conv_with_kernel_dilation.c
+++ b/bareMetalC/conv_with_kernel_dilation.c
@@ -276,7 +276,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, INPUT_DILATION, KERNEL_DILATION, PADDING, KERNEL_DIM,
+        STRIDE, INPUT_DILATION, KERNEL_DILATION, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         (elem_t*)input,
diff --git a/bareMetalC/conv_with_pool.c b/bareMetalC/conv_with_pool.c
index f896c4a6..3efbb226 100644
--- a/bareMetalC/conv_with_pool.c
+++ b/bareMetalC/conv_with_pool.c
@@ -288,7 +288,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         // 1,
diff --git a/bareMetalC/conv_with_rot180.c b/bareMetalC/conv_with_rot180.c
index ccbec995..13bb19b1 100644
--- a/bareMetalC/conv_with_rot180.c
+++ b/bareMetalC/conv_with_rot180.c
@@ -269,7 +269,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
-        STRIDE, INPUT_DILATION, 1, PADDING, KERNEL_DIM,
+        STRIDE, INPUT_DILATION, 1, PADDING, 0, KERNEL_DIM,
         WROT180, false, false, false, false,
 
         (elem_t*)input,
diff --git a/gemmini-data-collection/templates/conv_template.c b/gemmini-data-collection/templates/conv_template.c
index 49df0aa1..de65da02 100644
--- a/gemmini-data-collection/templates/conv_template.c
+++ b/gemmini-data-collection/templates/conv_template.c
@@ -201,7 +201,7 @@ int main() {
     tiled_conv_auto(
         BATCH_SIZE, IN_DIM, IN_CHANNELS,
         OUT_CHANNELS, OUT_DIM,
-        STRIDE, 1, 1, PADDING, KERNEL_DIM,
+        STRIDE, 1, 1, PADDING, 0, KERNEL_DIM,
         false, false, false, false, false,
 
         (elem_t*)input,
diff --git a/imagenet/alexnet.c b/imagenet/alexnet.c
index 18268397..b2d6ed7b 100644
--- a/imagenet/alexnet.c
+++ b/imagenet/alexnet.c
@@ -104,7 +104,7 @@ int main (int argc, char * argv[]) {
         tiled_conv_auto(
             conv_1_params.batch_size, conv_1_params.in_dim, conv_1_params.in_channels,
             conv_1_params.out_channels, conv_1_params.out_dim,
-            conv_1_params.stride, conv_1_params.padding, conv_1_params.kernel_size,
+            conv_1_params.stride, conv_1_params.padding, 0, conv_1_params.kernel_size,
 
             (elem_t*)images, (elem_t*)conv_1_w, (acc_t*)conv_1_b, (elem_t*)conv_1_out_pooled,
 
@@ -156,7 +156,7 @@ int main (int argc, char * argv[]) {
         tiled_conv_auto(
             conv_2_params.batch_size, conv_2_params.in_dim, conv_2_params.in_channels,
             conv_2_params.out_channels, conv_2_params.out_dim,
-            conv_2_params.stride, conv_2_params.padding, conv_2_params.kernel_size,
+            conv_2_params.stride, conv_2_params.padding, 0, conv_2_params.kernel_size,
 
             (elem_t*)conv_1_out_pooled, (elem_t*)conv_2_w, (acc_t*)conv_2_b, (elem_t*)conv_2_out_pooled,
 
diff --git a/imagenet/mobilenet.c b/imagenet/mobilenet.c
index a88292be..1f16e1f1 100644
--- a/imagenet/mobilenet.c
+++ b/imagenet/mobilenet.c
@@ -97,7 +97,7 @@ int main (int argc, char * argv[]) {
             conv_1_params.batch_size, conv_1_params.in_row_dim, conv_1_params.in_col_dim,
             conv_1_params.in_channels,
             conv_1_params.out_channels, conv_1_params.out_row_dim, conv_1_params.out_col_dim,
-            conv_1_params.stride, 1, 1, conv_1_params.padding, conv_1_params.kernel_size,
+            conv_1_params.stride, 1, 1, conv_1_params.padding, 0, conv_1_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)images, (elem_t*)conv_1_w, (acc_t*)conv_1_b, (elem_t*)conv_1_out,
@@ -126,7 +126,7 @@ int main (int argc, char * argv[]) {
             conv_dw_2_params.batch_size, conv_dw_2_params.in_row_dim, conv_dw_2_params.in_col_dim,
             conv_dw_2_params.in_channels,
             conv_dw_2_params.out_row_dim, conv_dw_2_params.out_col_dim,
-            conv_dw_2_params.stride, conv_dw_2_params.padding, conv_dw_2_params.kernel_size,
+            conv_dw_2_params.stride, conv_dw_2_params.padding, 0, conv_dw_2_params.kernel_size,
 
             (elem_t*)conv_1_out, (elem_t*)conv_dw_2_w, (acc_t*)conv_dw_2_b, (elem_t*)conv_dw_2_out,
 
@@ -206,7 +206,7 @@ int main (int argc, char * argv[]) {
             conv_dw_5_params.batch_size, conv_dw_5_params.in_row_dim, conv_dw_5_params.in_col_dim,
             conv_dw_5_params.in_channels,
             conv_dw_5_params.out_row_dim, conv_dw_5_params.out_col_dim,
-            conv_dw_5_params.stride, conv_dw_5_params.padding, conv_dw_5_params.kernel_size,
+            conv_dw_5_params.stride, conv_dw_5_params.padding, 0, conv_dw_5_params.kernel_size,
 
             (elem_t*)conv_4_out, (elem_t*)conv_dw_5_w, (acc_t*)conv_dw_5_b, (elem_t*)conv_dw_5_out,
 
@@ -287,7 +287,7 @@ int main (int argc, char * argv[]) {
             conv_dw_8_params.batch_size, conv_dw_8_params.in_row_dim, conv_dw_8_params.in_col_dim,
             conv_dw_8_params.in_channels,
             conv_dw_8_params.out_row_dim, conv_dw_8_params.out_col_dim,
-            conv_dw_8_params.stride, conv_dw_8_params.padding, conv_dw_8_params.kernel_size,
+            conv_dw_8_params.stride, conv_dw_8_params.padding, 0, conv_dw_8_params.kernel_size,
 
             (elem_t*)conv_7_out, (elem_t*)conv_dw_8_w, (acc_t*)conv_dw_8_b, (elem_t*)conv_dw_8_out,
 
@@ -383,7 +383,7 @@ int main (int argc, char * argv[]) {
             conv_dw_11_params.batch_size, conv_dw_11_params.in_row_dim, conv_dw_11_params.in_col_dim,
             conv_dw_11_params.in_channels,
             conv_dw_11_params.out_row_dim, conv_dw_11_params.out_col_dim,
-            conv_dw_11_params.stride, conv_dw_11_params.padding, conv_dw_11_params.kernel_size,
+            conv_dw_11_params.stride, conv_dw_11_params.padding, 0, conv_dw_11_params.kernel_size,
 
             (elem_t*)conv_10_out, (elem_t*)conv_dw_11_w, (acc_t*)conv_dw_11_b, (elem_t*)conv_dw_11_out,
 
@@ -463,7 +463,7 @@ int main (int argc, char * argv[]) {
             conv_dw_14_params.batch_size, conv_dw_14_params.in_row_dim, conv_dw_14_params.in_col_dim,
             conv_dw_14_params.in_channels,
             conv_dw_14_params.out_row_dim, conv_dw_14_params.out_col_dim,
-            conv_dw_14_params.stride, conv_dw_14_params.padding, conv_dw_14_params.kernel_size,
+            conv_dw_14_params.stride, conv_dw_14_params.padding, 0, conv_dw_14_params.kernel_size,
 
             (elem_t*)conv_13_out, (elem_t*)conv_dw_14_w, (acc_t*)conv_dw_14_b, (elem_t*)conv_dw_14_out,
 
@@ -559,7 +559,7 @@ int main (int argc, char * argv[]) {
             conv_dw_17_params.batch_size, conv_dw_17_params.in_row_dim, conv_dw_17_params.in_col_dim,
             conv_dw_17_params.in_channels,
             conv_dw_17_params.out_row_dim, conv_dw_17_params.out_col_dim,
-            conv_dw_17_params.stride, conv_dw_17_params.padding, conv_dw_17_params.kernel_size,
+            conv_dw_17_params.stride, conv_dw_17_params.padding, 0, conv_dw_17_params.kernel_size,
 
             (elem_t*)conv_16_out, (elem_t*)conv_dw_17_w, (acc_t*)conv_dw_17_b, (elem_t*)conv_dw_17_out,
 
@@ -655,7 +655,7 @@ int main (int argc, char * argv[]) {
             conv_dw_20_params.batch_size, conv_dw_20_params.in_row_dim, conv_dw_20_params.in_col_dim,
             conv_dw_20_params.in_channels,
             conv_dw_20_params.out_row_dim, conv_dw_20_params.out_col_dim,
-            conv_dw_20_params.stride, conv_dw_20_params.padding, conv_dw_20_params.kernel_size,
+            conv_dw_20_params.stride, conv_dw_20_params.padding, 0, conv_dw_20_params.kernel_size,
 
             (elem_t*)conv_19_out, (elem_t*)conv_dw_20_w, (acc_t*)conv_dw_20_b, (elem_t*)conv_dw_20_out,
 
@@ -734,7 +734,7 @@ int main (int argc, char * argv[]) {
             conv_dw_23_params.batch_size, conv_dw_23_params.in_row_dim, conv_dw_23_params.in_col_dim,
             conv_dw_23_params.in_channels,
             conv_dw_23_params.out_row_dim, conv_dw_23_params.out_col_dim,
-            conv_dw_23_params.stride, conv_dw_23_params.padding, conv_dw_23_params.kernel_size,
+            conv_dw_23_params.stride, conv_dw_23_params.padding, 0, conv_dw_23_params.kernel_size,
 
             (elem_t*)conv_22_out, (elem_t*)conv_dw_23_w, (acc_t*)conv_dw_23_b, (elem_t*)conv_dw_23_out,
 
@@ -830,7 +830,7 @@ int main (int argc, char * argv[]) {
             conv_dw_26_params.batch_size, conv_dw_26_params.in_row_dim, conv_dw_26_params.in_col_dim,
             conv_dw_26_params.in_channels,
             conv_dw_26_params.out_row_dim, conv_dw_26_params.out_col_dim,
-            conv_dw_26_params.stride, conv_dw_26_params.padding, conv_dw_26_params.kernel_size,
+            conv_dw_26_params.stride, conv_dw_26_params.padding, 0, conv_dw_26_params.kernel_size,
 
             (elem_t*)conv_25_out, (elem_t*)conv_dw_26_w, (acc_t*)conv_dw_26_b, (elem_t*)conv_dw_26_out,
 
@@ -926,7 +926,7 @@ int main (int argc, char * argv[]) {
             conv_dw_29_params.batch_size, conv_dw_29_params.in_row_dim, conv_dw_29_params.in_col_dim,
             conv_dw_29_params.in_channels,
             conv_dw_29_params.out_row_dim, conv_dw_29_params.out_col_dim,
-            conv_dw_29_params.stride, conv_dw_29_params.padding, conv_dw_29_params.kernel_size,
+            conv_dw_29_params.stride, conv_dw_29_params.padding, 0, conv_dw_29_params.kernel_size,
 
             (elem_t*)conv_28_out, (elem_t*)conv_dw_29_w, (acc_t*)conv_dw_29_b, (elem_t*)conv_dw_29_out,
 
@@ -1022,7 +1022,7 @@ int main (int argc, char * argv[]) {
             conv_dw_32_params.batch_size, conv_dw_32_params.in_row_dim, conv_dw_32_params.in_col_dim,
             conv_dw_32_params.in_channels,
             conv_dw_32_params.out_row_dim, conv_dw_32_params.out_col_dim,
-            conv_dw_32_params.stride, conv_dw_32_params.padding, conv_dw_32_params.kernel_size,
+            conv_dw_32_params.stride, conv_dw_32_params.padding, 0, conv_dw_32_params.kernel_size,
 
             (elem_t*)conv_31_out, (elem_t*)conv_dw_32_w, (acc_t*)conv_dw_32_b, (elem_t*)conv_dw_32_out,
 
@@ -1102,7 +1102,7 @@ int main (int argc, char * argv[]) {
             conv_dw_35_params.batch_size, conv_dw_35_params.in_row_dim, conv_dw_35_params.in_col_dim,
             conv_dw_35_params.in_channels,
             conv_dw_35_params.out_row_dim, conv_dw_35_params.out_col_dim,
-            conv_dw_35_params.stride, conv_dw_35_params.padding, conv_dw_35_params.kernel_size,
+            conv_dw_35_params.stride, conv_dw_35_params.padding, 0, conv_dw_35_params.kernel_size,
 
             (elem_t*)conv_34_out, (elem_t*)conv_dw_35_w, (acc_t*)conv_dw_35_b, (elem_t*)conv_dw_35_out,
 
@@ -1198,7 +1198,7 @@ int main (int argc, char * argv[]) {
             conv_dw_38_params.batch_size, conv_dw_38_params.in_row_dim, conv_dw_38_params.in_col_dim,
             conv_dw_38_params.in_channels,
             conv_dw_38_params.out_row_dim, conv_dw_38_params.out_col_dim,
-            conv_dw_38_params.stride, conv_dw_38_params.padding, conv_dw_38_params.kernel_size,
+            conv_dw_38_params.stride, conv_dw_38_params.padding, 0, conv_dw_38_params.kernel_size,
 
             (elem_t*)conv_37_out, (elem_t*)conv_dw_38_w, (acc_t*)conv_dw_38_b, (elem_t*)conv_dw_38_out,
 
@@ -1294,7 +1294,7 @@ int main (int argc, char * argv[]) {
             conv_dw_41_params.batch_size, conv_dw_41_params.in_row_dim, conv_dw_41_params.in_col_dim,
             conv_dw_41_params.in_channels,
             conv_dw_41_params.out_row_dim, conv_dw_41_params.out_col_dim,
-            conv_dw_41_params.stride, conv_dw_41_params.padding, conv_dw_41_params.kernel_size,
+            conv_dw_41_params.stride, conv_dw_41_params.padding, 0, conv_dw_41_params.kernel_size,
 
             (elem_t*)conv_40_out, (elem_t*)conv_dw_41_w, (acc_t*)conv_dw_41_b, (elem_t*)conv_dw_41_out,
 
@@ -1374,7 +1374,7 @@ int main (int argc, char * argv[]) {
             conv_dw_44_params.batch_size, conv_dw_44_params.in_row_dim, conv_dw_44_params.in_col_dim,
             conv_dw_44_params.in_channels,
             conv_dw_44_params.out_row_dim, conv_dw_44_params.out_col_dim,
-            conv_dw_44_params.stride, conv_dw_44_params.padding, conv_dw_44_params.kernel_size,
+            conv_dw_44_params.stride, conv_dw_44_params.padding, 0, conv_dw_44_params.kernel_size,
 
             (elem_t*)conv_43_out, (elem_t*)conv_dw_44_w, (acc_t*)conv_dw_44_b, (elem_t*)conv_dw_44_out,
 
@@ -1470,7 +1470,7 @@ int main (int argc, char * argv[]) {
             conv_dw_47_params.batch_size, conv_dw_47_params.in_row_dim, conv_dw_47_params.in_col_dim,
             conv_dw_47_params.in_channels,
             conv_dw_47_params.out_row_dim, conv_dw_47_params.out_col_dim,
-            conv_dw_47_params.stride, conv_dw_47_params.padding, conv_dw_47_params.kernel_size,
+            conv_dw_47_params.stride, conv_dw_47_params.padding, 0, conv_dw_47_params.kernel_size,
 
             (elem_t*)conv_46_out, (elem_t*)conv_dw_47_w, (acc_t*)conv_dw_47_b, (elem_t*)conv_dw_47_out,
 
@@ -1566,7 +1566,7 @@ int main (int argc, char * argv[]) {
             conv_dw_50_params.batch_size, conv_dw_50_params.in_row_dim, conv_dw_50_params.in_col_dim,
             conv_dw_50_params.in_channels,
             conv_dw_50_params.out_row_dim, conv_dw_50_params.out_col_dim,
-            conv_dw_50_params.stride, conv_dw_50_params.padding, conv_dw_50_params.kernel_size,
+            conv_dw_50_params.stride, conv_dw_50_params.padding, 0, conv_dw_50_params.kernel_size,
 
             (elem_t*)conv_49_out, (elem_t*)conv_dw_50_w, (acc_t*)conv_dw_50_b, (elem_t*)conv_dw_50_out,
 
diff --git a/imagenet/resnet50.c b/imagenet/resnet50.c
index 4bbd7849..f92e6400 100644
--- a/imagenet/resnet50.c
+++ b/imagenet/resnet50.c
@@ -109,7 +109,7 @@ int main (int argc, char * argv[]) {
             conv_1_params.batch_size, conv_1_params.in_row_dim, conv_1_params.in_col_dim,
             conv_1_params.in_channels,
             conv_1_params.out_channels, conv_1_params.out_row_dim, conv_1_params.out_col_dim,
-            conv_1_params.stride, 1, 1, conv_1_params.padding, conv_1_params.kernel_size,
+            conv_1_params.stride, 1, 1, conv_1_params.padding, 0, conv_1_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)images, (elem_t*)conv_1_w, (acc_t*)conv_1_b, (elem_t*)conv_1_out_pooled,
@@ -187,7 +187,7 @@ int main (int argc, char * argv[]) {
             conv_3_params.batch_size, conv_3_params.in_row_dim, conv_3_params.in_col_dim,
             conv_3_params.in_channels,
             conv_3_params.out_channels, conv_3_params.out_row_dim, conv_3_params.out_col_dim,
-            conv_3_params.stride, 1, 1, conv_3_params.padding, conv_3_params.kernel_size,
+            conv_3_params.stride, 1, 1, conv_3_params.padding, 0, conv_3_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_2_out, (elem_t*)conv_3_w, (acc_t*)conv_3_b, (elem_t*)conv_3_out,
@@ -332,7 +332,7 @@ int main (int argc, char * argv[]) {
             conv_7_params.batch_size, conv_7_params.in_row_dim, conv_7_params.in_col_dim,
             conv_7_params.in_channels,
             conv_7_params.out_channels, conv_7_params.out_row_dim, conv_7_params.out_col_dim,
-            conv_7_params.stride, 1, 1, conv_7_params.padding, conv_7_params.kernel_size,
+            conv_7_params.stride, 1, 1, conv_7_params.padding, 0, conv_7_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_6_out, (elem_t*)conv_7_w, (acc_t*)conv_7_b, (elem_t*)conv_7_out,
@@ -441,7 +441,7 @@ int main (int argc, char * argv[]) {
             conv_10_params.batch_size, conv_10_params.in_row_dim, conv_10_params.in_col_dim,
             conv_10_params.in_channels,
             conv_10_params.out_channels, conv_10_params.out_row_dim, conv_10_params.out_col_dim,
-            conv_10_params.stride, 1, 1, conv_10_params.padding, conv_10_params.kernel_size,
+            conv_10_params.stride, 1, 1, conv_10_params.padding, 0, conv_10_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_9_out, (elem_t*)conv_10_w, (acc_t*)conv_10_b, (elem_t*)conv_10_out,
@@ -550,7 +550,7 @@ int main (int argc, char * argv[]) {
             conv_13_params.batch_size, conv_13_params.in_row_dim, conv_13_params.in_col_dim,
             conv_13_params.in_channels,
             conv_13_params.out_channels, conv_13_params.out_row_dim, conv_13_params.out_col_dim,
-            conv_13_params.stride, 1, 1, conv_13_params.padding, conv_13_params.kernel_size,
+            conv_13_params.stride, 1, 1, conv_13_params.padding, 0, conv_13_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_12_out, (elem_t*)conv_13_w, (acc_t*)conv_13_b, (elem_t*)conv_13_out,
@@ -620,7 +620,7 @@ int main (int argc, char * argv[]) {
             conv_15_params.batch_size, conv_15_params.in_row_dim, conv_15_params.in_col_dim,
             conv_15_params.in_channels,
             conv_15_params.out_channels, conv_15_params.out_row_dim, conv_15_params.out_col_dim,
-            // conv_15_params.stride, 1, 1, conv_15_params.padding, conv_15_params.kernel_size,
+            // conv_15_params.stride, 1, 1, conv_15_params.padding, 0, conv_15_params.kernel_size,
             // false, false, false, false, false,
 
             (elem_t*)conv_11_out, (elem_t*)conv_15_w, (acc_t*)conv_15_b, (elem_t*)conv_15_out,
@@ -704,7 +704,7 @@ int main (int argc, char * argv[]) {
             conv_17_params.batch_size, conv_17_params.in_row_dim, conv_17_params.in_col_dim,
             conv_17_params.in_channels,
             conv_17_params.out_channels, conv_17_params.out_row_dim, conv_17_params.out_col_dim,
-            conv_17_params.stride, 1, 1, conv_17_params.padding, conv_17_params.kernel_size,
+            conv_17_params.stride, 1, 1, conv_17_params.padding, 0, conv_17_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_16_out, (elem_t*)conv_17_w, (acc_t*)conv_17_b, (elem_t*)conv_17_out,
@@ -813,7 +813,7 @@ int main (int argc, char * argv[]) {
             conv_20_params.batch_size, conv_20_params.in_row_dim, conv_20_params.in_col_dim,
             conv_20_params.in_channels,
             conv_20_params.out_channels, conv_20_params.out_row_dim, conv_20_params.out_col_dim,
-            conv_20_params.stride, 1, 1, conv_20_params.padding, conv_20_params.kernel_size,
+            conv_20_params.stride, 1, 1, conv_20_params.padding, 0, conv_20_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_19_out, (elem_t*)conv_20_w, (acc_t*)conv_20_b, (elem_t*)conv_20_out,
@@ -922,7 +922,7 @@ int main (int argc, char * argv[]) {
             conv_23_params.batch_size, conv_23_params.in_row_dim, conv_23_params.in_col_dim,
             conv_23_params.in_channels,
             conv_23_params.out_channels, conv_23_params.out_row_dim, conv_23_params.out_col_dim,
-            conv_23_params.stride, 1, 1, conv_23_params.padding, conv_23_params.kernel_size,
+            conv_23_params.stride, 1, 1, conv_23_params.padding, 0, conv_23_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_22_out, (elem_t*)conv_23_w, (acc_t*)conv_23_b, (elem_t*)conv_23_out,
@@ -1031,7 +1031,7 @@ int main (int argc, char * argv[]) {
             conv_26_params.batch_size, conv_26_params.in_row_dim, conv_26_params.in_col_dim,
             conv_26_params.in_channels,
             conv_26_params.out_channels, conv_26_params.out_row_dim, conv_26_params.out_col_dim,
-            conv_26_params.stride, 1, 1, conv_26_params.padding, conv_26_params.kernel_size,
+            conv_26_params.stride, 1, 1, conv_26_params.padding, 0, conv_26_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_25_out, (elem_t*)conv_26_w, (acc_t*)conv_26_b, (elem_t*)conv_26_out,
@@ -1101,7 +1101,7 @@ int main (int argc, char * argv[]) {
             conv_28_params.batch_size, conv_28_params.in_row_dim, conv_28_params.in_col_dim,
             conv_28_params.in_channels,
             conv_28_params.out_channels, conv_28_params.out_row_dim, conv_28_params.out_col_dim,
-            // conv_28_params.stride, 1, 1, conv_28_params.padding, conv_28_params.kernel_size,
+            // conv_28_params.stride, 1, 1, conv_28_params.padding, 0, conv_28_params.kernel_size,
             // false, false, false, false, false,
 
             (elem_t*)conv_24_out, (elem_t*)conv_28_w, (acc_t*)conv_28_b, (elem_t*)conv_28_out,
@@ -1185,7 +1185,7 @@ int main (int argc, char * argv[]) {
             conv_30_params.batch_size, conv_30_params.in_row_dim, conv_30_params.in_col_dim,
             conv_30_params.in_channels,
             conv_30_params.out_channels, conv_30_params.out_row_dim, conv_30_params.out_col_dim,
-            conv_30_params.stride, 1, 1, conv_30_params.padding, conv_30_params.kernel_size,
+            conv_30_params.stride, 1, 1, conv_30_params.padding, 0, conv_30_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_29_out, (elem_t*)conv_30_w, (acc_t*)conv_30_b, (elem_t*)conv_30_out,
@@ -1294,7 +1294,7 @@ int main (int argc, char * argv[]) {
             conv_33_params.batch_size, conv_33_params.in_row_dim, conv_33_params.in_col_dim,
             conv_33_params.in_channels,
             conv_33_params.out_channels, conv_33_params.out_row_dim, conv_33_params.out_col_dim,
-            conv_33_params.stride, 1, 1, conv_33_params.padding, conv_33_params.kernel_size,
+            conv_33_params.stride, 1, 1, conv_33_params.padding, 0, conv_33_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_32_out, (elem_t*)conv_33_w, (acc_t*)conv_33_b, (elem_t*)conv_33_out,
@@ -1403,7 +1403,7 @@ int main (int argc, char * argv[]) {
             conv_36_params.batch_size, conv_36_params.in_row_dim, conv_36_params.in_col_dim,
             conv_36_params.in_channels,
             conv_36_params.out_channels, conv_36_params.out_row_dim, conv_36_params.out_col_dim,
-            conv_36_params.stride, 1, 1, conv_36_params.padding, conv_36_params.kernel_size,
+            conv_36_params.stride, 1, 1, conv_36_params.padding, 0, conv_36_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_35_out, (elem_t*)conv_36_w, (acc_t*)conv_36_b, (elem_t*)conv_36_out,
@@ -1512,7 +1512,7 @@ int main (int argc, char * argv[]) {
             conv_39_params.batch_size, conv_39_params.in_row_dim, conv_39_params.in_col_dim,
             conv_39_params.in_channels,
             conv_39_params.out_channels, conv_39_params.out_row_dim, conv_39_params.out_col_dim,
-            conv_39_params.stride, 1, 1, conv_39_params.padding, conv_39_params.kernel_size,
+            conv_39_params.stride, 1, 1, conv_39_params.padding, 0, conv_39_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_38_out, (elem_t*)conv_39_w, (acc_t*)conv_39_b, (elem_t*)conv_39_out,
@@ -1621,7 +1621,7 @@ int main (int argc, char * argv[]) {
             conv_42_params.batch_size, conv_42_params.in_row_dim, conv_42_params.in_col_dim,
             conv_42_params.in_channels,
             conv_42_params.out_channels, conv_42_params.out_row_dim, conv_42_params.out_col_dim,
-            conv_42_params.stride, 1, 1, conv_42_params.padding, conv_42_params.kernel_size,
+            conv_42_params.stride, 1, 1, conv_42_params.padding, 0, conv_42_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_41_out, (elem_t*)conv_42_w, (acc_t*)conv_42_b, (elem_t*)conv_42_out,
@@ -1730,7 +1730,7 @@ int main (int argc, char * argv[]) {
             conv_45_params.batch_size, conv_45_params.in_row_dim, conv_45_params.in_col_dim,
             conv_45_params.in_channels,
             conv_45_params.out_channels, conv_45_params.out_row_dim, conv_45_params.out_col_dim,
-            conv_45_params.stride, 1, 1, conv_45_params.padding, conv_45_params.kernel_size,
+            conv_45_params.stride, 1, 1, conv_45_params.padding, 0, conv_45_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_44_out, (elem_t*)conv_45_w, (acc_t*)conv_45_b, (elem_t*)conv_45_out,
@@ -1799,7 +1799,7 @@ int main (int argc, char * argv[]) {
             conv_47_params.batch_size, conv_47_params.in_row_dim, conv_47_params.in_col_dim,
             conv_47_params.in_channels,
             conv_47_params.out_channels, conv_47_params.out_row_dim, conv_47_params.out_col_dim,
-            conv_47_params.stride, 1, 1, conv_47_params.padding, conv_47_params.kernel_size,
+            conv_47_params.stride, 1, 1, conv_47_params.padding, 0, conv_47_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_43_out, (elem_t*)conv_47_w, (acc_t*)conv_47_b, (elem_t*)conv_47_out,
@@ -1883,7 +1883,7 @@ int main (int argc, char * argv[]) {
             conv_49_params.batch_size, conv_49_params.in_row_dim, conv_49_params.in_col_dim,
             conv_49_params.in_channels,
             conv_49_params.out_channels, conv_49_params.out_row_dim, conv_49_params.out_col_dim,
-            conv_49_params.stride, 1, 1, conv_49_params.padding, conv_49_params.kernel_size,
+            conv_49_params.stride, 1, 1, conv_49_params.padding, 0, conv_49_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_48_out, (elem_t*)conv_49_w, (acc_t*)conv_49_b, (elem_t*)conv_49_out,
@@ -1992,7 +1992,7 @@ int main (int argc, char * argv[]) {
             conv_52_params.batch_size, conv_52_params.in_row_dim, conv_52_params.in_col_dim,
             conv_52_params.in_channels,
             conv_52_params.out_channels, conv_52_params.out_row_dim, conv_52_params.out_col_dim,
-            conv_52_params.stride, 1, 1, conv_52_params.padding, conv_52_params.kernel_size,
+            conv_52_params.stride, 1, 1, conv_52_params.padding, 0, conv_52_params.kernel_size,
             false, false, false, false, false,
 
             (elem_t*)conv_51_out, (elem_t*)conv_52_w, (acc_t*)conv_52_b, (elem_t*)conv_52_out,
diff --git a/include/gemmini.h b/include/gemmini.h
index 76b50a48..c28dafa7 100644
--- a/include/gemmini.h
+++ b/include/gemmini.h
@@ -254,8 +254,11 @@ static acc_scale_t_bits acc_scale_t_to_acc_scale_t_bits(acc_scale_t x) {
 
 // Note: The "pixel_repeats" parameter below is still experimental, andthere is
 // a high chance that it will be removed in future releases.
+#define gemmini_extended6_config_ld(stride, scale, shrunk, block_mvin_stride, pixel_repeats, padding_value, id) \
+  ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, ((uint64_t)(scale_t_to_scale_t_bits(scale)) << 32) | ((uint64_t)(block_mvin_stride) << 16) | ((uint64_t)(pixel_repeats) << 8) | ((id) << 3) | ((shrunk) << 2) | CONFIG_LD, ((uint64_t)(padding_value) << 32) | stride, k_CONFIG)
+
 #define gemmini_extended5_config_ld(stride, scale, shrunk, block_mvin_stride, pixel_repeats, id) \
-  ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, ((uint64_t)(scale_t_to_scale_t_bits(scale)) << 32) | ((uint64_t)(block_mvin_stride) << 16) | ((uint64_t)(pixel_repeats) << 8) | ((id) << 3) | ((shrunk) << 2) | CONFIG_LD, stride, k_CONFIG)
+  gemmini_extended6_config_ld(stride, scale, shrunk, block_mvin_stride, pixel_repeats, 0, id) \
 
 #define gemmini_extended4_config_ld(stride, scale, shrunk, block_mvin_stride, id) \
   gemmini_extended5_config_ld(stride, scale, shrunk, block_mvin_stride, 1, id) \
@@ -351,7 +354,7 @@ static void counter_reset() {
   }
 
 // weight-stationary conv loop
-#define gemmini_loop_conv_ws(batch_size, in_dim, in_channels, out_channels, out_dim, pool_out_dim, stride, padding, kernel_dim, kernel_dilation, pool_size, pool_stride, pool_padding, batches, porows, pocols, pochs, krows, kcols, kchs, lpad, rpad, upad, dpad, plpad, prpad, pupad, pdpad, orows, ocols, weights, output, bias, input, no_bias, no_pool, downsample, wrot180, input_dilated, activation, trans_output_1203, trans_weight_1203, trans_weight_0132, trans_input_3120, max_pixels_per_row, dw) \
+#define gemmini_loop_conv_ws(batch_size, in_dim, in_channels, out_channels, out_dim, pool_out_dim, stride, padding, padding_value, kernel_dim, kernel_dilation, pool_size, pool_stride, pool_padding, batches, porows, pocols, pochs, krows, kcols, kchs, lpad, rpad, upad, dpad, plpad, prpad, pupad, pdpad, orows, ocols, weights, output, bias, input, no_bias, no_pool, downsample, wrot180, input_dilated, activation, trans_output_1203, trans_weight_1203, trans_weight_0132, trans_input_3120, max_pixels_per_row, dw) \
   { \
     ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, ((uint64_t)(out_channels) << 48) | ((uint64_t)(in_channels) << 32) | ((uint64_t)(in_dim) << 16) | (uint64_t)(batch_size), \
       ((uint64_t)(padding) << 48) | ((uint64_t)(stride) << 32) | ((uint64_t)(pool_out_dim) << 16) | (uint64_t)(out_dim), k_LOOP_CONV_WS_CONFIG_1) \
@@ -365,7 +368,7 @@ static void counter_reset() {
       output, k_LOOP_CONV_WS_CONFIG_5) \
     ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, bias, \
       input, k_LOOP_CONV_WS_CONFIG_6) \
-    ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, ((uint64_t)(max_pixels_per_row) << 8) | ((dw) << 6) | ((trans_input_3120) << 5) | ((trans_weight_0132) << 4) | ((trans_weight_1203) << 3) | ((trans_output_1203) << 2) | ((wrot180) << 1) | (no_bias), \
+    ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, ((uint64_t)(padding_value) << 48) | ((uint64_t)(max_pixels_per_row) << 8) | ((dw) << 6) | ((trans_input_3120) << 5) | ((trans_weight_0132) << 4) | ((trans_weight_1203) << 3) | ((trans_output_1203) << 2) | ((wrot180) << 1) | (no_bias), \
       ((activation) << 3)| ((input_dilated) << 2) | ((downsample) << 1) | (no_pool), \
       k_LOOP_CONV_WS) \
   }
@@ -1328,7 +1331,7 @@ static void sp_tiled_conv(
         int out_channels, int out_row_dim, int out_col_dim,
         int pool_out_row_dim, int pool_out_col_dim,
 
-        int stride, int padding, int kernel_dim, int kernel_dilation,
+        int stride, int padding, elem_t padding_value, int kernel_dim, int kernel_dilation,
 
         int pool_size, int pool_stride, int pool_padding,
 
@@ -1414,7 +1417,7 @@ static void sp_tiled_conv(
   }
 
   if (in_row_dim == in_col_dim && out_row_dim == out_col_dim && pool_out_row_dim == pool_out_col_dim) {
-    gemmini_loop_conv_ws(batch_size, in_row_dim, in_channels, out_channels, out_row_dim, pool_out_row_dim, stride, padding, kernel_dim, kernel_dilation, pool_size, pool_stride, pool_padding, batches, porows, pocols, pochs, krows, kcols, kchs, lpad, rpad, upad, dpad, plpad, prpad, pupad, pdpad, orows, ocols, weights, output, bias, input, no_bias, no_pool, downsample, wrot180, input_dilated, act, trans_output_1203, trans_weight_1203, trans_weight_0132, trans_input_3120, max_pixels_per_row, dw);
+    gemmini_loop_conv_ws(batch_size, in_row_dim, in_channels, out_channels, out_row_dim, pool_out_row_dim, stride, padding, padding_value, kernel_dim, kernel_dilation, pool_size, pool_stride, pool_padding, batches, porows, pocols, pochs, krows, kcols, kchs, lpad, rpad, upad, dpad, plpad, prpad, pupad, pdpad, orows, ocols, weights, output, bias, input, no_bias, no_pool, downsample, wrot180, input_dilated, act, trans_output_1203, trans_weight_1203, trans_weight_0132, trans_input_3120, max_pixels_per_row, dw);
     return;
   }
 
@@ -1470,7 +1473,7 @@ static void sp_tiled_conv(
       ichs * (irows >> downsample) * (icols >> downsample) :
       batches * (irows >> downsample) * (icols >> downsample);
 
-    gemmini_extended5_config_ld(dram_stride << downsample, MVIN_SCALE_IDENTITY, false, spad_stride, max_pixels_per_row, 0);
+    gemmini_extended6_config_ld(dram_stride << downsample, MVIN_SCALE_IDENTITY, false, spad_stride, max_pixels_per_row, padding_value, 0);
 
     const int b_it = trans_input_3120 ? max_chs_per_mvin : 1;
     const int ich_it = trans_input_3120 ? 1 : max_chs_per_mvin;
@@ -1810,7 +1813,7 @@ static int tiled_conv_total_spad_rows(bool acc,
 static void conv_cpu_without_pool(
         int batch_size, int in_row_dim, int in_col_dim, int in_channels,
         int out_channels, int out_row_dim, int out_col_dim,
-        int stride, int input_dilation, int kernel_dilation, int padding, int kernel_dim,
+        int stride, int input_dilation, int kernel_dilation, int padding, elem_t padding_value, int kernel_dim,
         bool wrot180, bool trans_output_1203, bool trans_input_3120,
         bool trans_weight_1203, bool trans_weight_0132,
 
@@ -1850,7 +1853,7 @@ static void conv_cpu_without_pool(
                 }
 
                 elem_t ipixel = irow < 0 || irow >= in_row_dim || icol < 0 || icol >= in_col_dim ?
-                    0 : *in;
+                    padding_value : *in;
 
                 const int krow_ = wrot180 ? kernel_dim - krow - 1 : krow;
                 const int kcol_ = wrot180 ? kernel_dim - kcol - 1 : kcol;
@@ -1886,7 +1889,7 @@ static void conv_cpu_without_pool(
 static void conv_dw_cpu_without_pool(
         int batch_size, int in_row_dim, int in_col_dim,
         int channels, int out_row_dim, int out_col_dim,
-        int stride, int padding, int kernel_dim,
+        int stride, int padding, elem_t padding_value, int kernel_dim,
 
         const elem_t * input,
         const elem_t * weights,
@@ -1912,7 +1915,7 @@ static void conv_dw_cpu_without_pool(
               const elem_t * in = input + (b * in_row_dim * in_col_dim + irow * in_col_dim + icol) * channels + ch;
 
               const elem_t ipixel = irow < 0 || irow >= in_row_dim || icol < 0 || icol >= in_col_dim ?
-                  0 : *in;
+                  padding_value : *in;
 
               const elem_t weight = *(weights + (ch * kernel_dim + krow) * kernel_dim  + kcol);
 
@@ -1933,7 +1936,7 @@ static void conv_dw_cpu_without_pool(
 static void conv_cpu(
         int batch_size, int in_row_dim, int in_col_dim, int in_channels,
         int out_channels, int out_row_dim, int out_col_dim,
-        int stride, int input_dilation, int kernel_dilation, int padding, int kernel_dim,
+        int stride, int input_dilation, int kernel_dilation, int padding, elem_t padding_value, int kernel_dim,
         bool wrot180, bool trans_output_1203, bool trans_input_3120,
         bool trans_weight_1203, bool trans_weight_0132,
 
@@ -1950,7 +1953,7 @@ static void conv_cpu(
     conv_cpu_without_pool(
         batch_size, in_row_dim, in_col_dim, in_channels,
         out_channels, out_row_dim, out_col_dim,
-        stride, input_dilation, kernel_dilation, padding, kernel_dim,
+        stride, input_dilation, kernel_dilation, padding, padding_value, kernel_dim,
         wrot180, trans_output_1203, trans_input_3120,
         trans_weight_1203, trans_weight_0132,
         input, weights, bias, output,
@@ -2005,7 +2008,7 @@ static void conv_cpu(
                       }
 
                       elem_t ipixel = irow < 0 || irow >= in_row_dim || icol < 0 || icol >= in_col_dim ?
-                          0 : *in;
+                          padding_value : *in;
 
                       const int krow_ = wrot180 ? kernel_dim - krow - 1 : krow;
                       const int kcol_ = wrot180 ? kernel_dim - kcol - 1 : kcol;
@@ -2052,7 +2055,7 @@ static void conv_cpu(
 static void conv_dw_cpu(
         int batch_size, int in_row_dim, int in_col_dim,
         int channels, int out_row_dim, int out_col_dim,
-        int stride, int padding, int kernel_dim,
+        int stride, int padding, elem_t padding_value, int kernel_dim,
 
         const elem_t * input,
         const elem_t * weights,
@@ -2067,7 +2070,7 @@ static void conv_dw_cpu(
     conv_dw_cpu_without_pool(
         batch_size, in_row_dim, in_col_dim,
         channels, out_row_dim, out_col_dim,
-        stride, padding, kernel_dim,
+        stride, padding, padding_value, kernel_dim,
         input, weights, bias, output,
         act, scale);
     return;
@@ -2109,7 +2112,7 @@ static void conv_dw_cpu(
                     const elem_t * in = input + (b * in_row_dim * in_col_dim + irow * in_col_dim + icol) * channels + ch;
 
                     elem_t ipixel = irow < 0 || irow >= in_row_dim || icol < 0 || icol >= in_col_dim ?
-                        0 : *in;
+                        padding_value : *in;
 
                     const elem_t weight = *(weights + (ch * kernel_dim + krow) * kernel_dim  + kcol);
 
@@ -2142,7 +2145,7 @@ static void tiled_conv(
         int batch_size,
         int in_row_dim, int in_col_dim, int in_channels,
         int out_channels, int out_row_dim, int out_col_dim,
-        int stride, int input_dilation, int kernel_dilation, int padding, int kernel_dim,
+        int stride, int input_dilation, int kernel_dilation, int padding, elem_t padding_value, int kernel_dim,
         bool wrot180, bool trans_output_1203, bool trans_input_3120,
         bool trans_weight_1203, bool trans_weight_0132,
 
@@ -2177,7 +2180,7 @@ static void tiled_conv(
       conv_cpu(
         batch_size, in_row_dim, in_col_dim, in_channels,
         out_channels, out_row_dim, out_col_dim,
-        stride, input_dilation, kernel_dilation, padding, kernel_dim,
+        stride, input_dilation, kernel_dilation, padding, padding_value, kernel_dim,
         wrot180, trans_output_1203, trans_input_3120,
         trans_weight_1203, trans_weight_0132,
         input, weights, bias, output,
@@ -2358,7 +2361,7 @@ static void tiled_conv(
                                     out_channels, out_row_dim, out_col_dim,
                                     pool_out_row_dim, pool_out_col_dim,
 
-                                    stride, padding, kernel_dim, kernel_dilation,
+                                    stride, padding, padding_value, kernel_dim, kernel_dilation,
 
                                     pool_size, pool_stride, pool_padding,
 
@@ -2395,7 +2398,7 @@ static void tiled_conv(
 static void tiled_conv_dw(
     int batch_size, int in_row_dim, int in_col_dim,
     int channels, int out_row_dim, int out_col_dim,
-    int stride, int padding, int kernel_dim,
+    int stride, int padding, elem_t padding_value, int kernel_dim,
 
     int batches,
     int porows, int pocols,
@@ -2419,7 +2422,7 @@ static void tiled_conv_dw(
       conv_dw_cpu(
         batch_size, in_row_dim, in_col_dim,
         channels, out_row_dim, out_col_dim,
-        stride, padding, kernel_dim,
+        stride, padding, padding_value, kernel_dim,
         input, weights, bias, output,
         act, scale,
         pool_size, pool_stride, pool_padding);
@@ -2540,7 +2543,7 @@ static void tiled_conv_dw(
                                 channels, out_row_dim, out_col_dim,
                                 pool_out_row_dim, pool_out_col_dim,
 
-                                stride, padding, kernel_dim, 1,
+                                stride, padding, padding_value, kernel_dim, 1,
 
                                 pool_size, pool_stride, pool_padding,
 
@@ -2576,7 +2579,7 @@ static void tiled_conv_dw(
 static void tiled_conv_auto(
         int batch_size, int in_row_dim, int in_col_dim, int in_channels,
         int out_channels, int out_row_dim, int out_col_dim,
-        int stride, int input_dilation, int kernel_dilation, int padding, int kernel_dim,
+        int stride, int input_dilation, int kernel_dilation, int padding, elem_t padding_value, int kernel_dim,
         bool wrot180, bool trans_output_1203, bool trans_input_3120,
         bool trans_weight_1203, bool trans_weight_0132,
 
@@ -2747,7 +2750,7 @@ static void tiled_conv_auto(
     tiled_conv(
         batch_size, in_row_dim, in_col_dim, in_channels,
         out_channels, out_row_dim, out_col_dim,
-        stride, input_dilation, kernel_dilation, padding, kernel_dim,
+        stride, input_dilation, kernel_dilation, padding, padding_value, kernel_dim,
         wrot180, trans_output_1203, trans_input_3120,
         trans_weight_1203, trans_weight_0132,
 
@@ -2823,7 +2826,7 @@ static void tiled_conv_downsample(
 static void tiled_conv_dw_auto(
     int batch_size, int in_row_dim, int in_col_dim,
     int channels, int out_row_dim, int out_col_dim,
-    int stride, int padding, int kernel_dim,
+    int stride, int padding, elem_t padding_value, int kernel_dim,
 
     elem_t * input,
     elem_t * weights,
@@ -2985,7 +2988,7 @@ static void tiled_conv_dw_auto(
     tiled_conv_dw(
         batch_size, in_row_dim, in_col_dim,
         channels, out_row_dim, out_col_dim,
-        stride, padding, kernel_dim,
+        stride, padding, padding_value, kernel_dim,
 
         batches,
         orows, ocols,