test diff threshold

tenstorrent · Feb 19, 2025 · 23b5886 · 23b5886
1 parent 8c9f534
commit 23b5886
Show file tree

Hide file tree

Showing 4 changed files with 9 additions and 8 deletions.
diff --git a/...nn/operations/data_movement/tilize_with_val_padding/device/tilize_with_val_padding_op.cpp b/...nn/operations/data_movement/tilize_with_val_padding/device/tilize_with_val_padding_op.cpp
@@ -90,7 +90,7 @@ operation::ProgramWithCallbacks TilizeWithValPadding::create_program(
     if (input_tensor_a.memory_config().is_sharded()) {
         return detail::tilize_with_val_padding_multi_core_sharded(input_tensor_a, output_tensor, this->pad_value);
     }
-    if (!this->enough_space_height) {
+    if (!this->enough_space_height || this->enough_space_width) {
         return detail::tilize_with_val_padding_multi_core_block_interleaved(
             input_tensor_a, output_tensor, this->pad_value);
     }

diff --git a/.../data_movement/tilize_with_val_padding/device/tilize_with_val_padding_program_factory.cpp b/.../data_movement/tilize_with_val_padding/device/tilize_with_val_padding_program_factory.cpp
@@ -687,9 +687,9 @@ operation::ProgramWithCallbacks tilize_with_val_padding_multi_core_interleaved(
         if (num_tiles_per_row > num_tiles_per_col) {
             return tilize_with_val_padding_multi_core_col_interleaved(a, output, pad_value);
         }
-        if (num_tiles_per_col > threshold_row_block || num_tiles_per_row > num_tiles_per_col) {
-            return tilize_with_val_padding_multi_core_block_interleaved(a, output, pad_value);
-        }
+        // if (num_tiles_per_col > threshold_row_block || num_tiles_per_row > num_tiles_per_col) {
+        //     return tilize_with_val_padding_multi_core_block_interleaved(a, output, pad_value);
+        // }
     }
 
     auto [ncores, all_cores, core_range, core_range_cliff, nblocks_per_core, nblocks_per_core_cliff] =

diff --git a/...nn/operations/data_movement/untilize_with_unpadding/device/untilize_with_unpadding_op.cpp b/...nn/operations/data_movement/untilize_with_unpadding/device/untilize_with_unpadding_op.cpp
@@ -115,7 +115,7 @@ operation::ProgramWithCallbacks UntilizeWithUnpadding::create_program(
         return detail::untilize_with_unpadding_single_core(
             input_tensor_a, output_tensor, this->use_pack_untilize, this->fp32_dest_acc_en);
     }
-    if (!this->enough_space_height) {
+    if (!this->enough_space_height || !this->enough_space_width) {
         return detail::untilize_with_unpadding_multi_core_block_interleaved(
             input_tensor_a, output_tensor, this->use_pack_untilize, this->fp32_dest_acc_en);
     }

diff --git a/.../data_movement/untilize_with_unpadding/device/untilize_with_unpadding_program_factory.cpp b/.../data_movement/untilize_with_unpadding/device/untilize_with_unpadding_program_factory.cpp
@@ -675,9 +675,10 @@ operation::ProgramWithCallbacks untilize_with_unpadding_multi_core_interleaved(
         if (num_tiles_per_row > num_tiles_per_col) {
             return untilize_with_unpadding_multi_core_col_interleaved(a, output, use_pack_untilize, fp32_dest_acc_en);
         }
-        if (num_tiles_per_col > threshold_row_block || num_tiles_per_row > num_tiles_per_col) {
-            return untilize_with_unpadding_multi_core_block_interleaved(a, output, use_pack_untilize, fp32_dest_acc_en);
-        }
+        // if (num_tiles_per_col > threshold_row_block || num_tiles_per_row > num_tiles_per_col) {
+        //     return untilize_with_unpadding_multi_core_block_interleaved(a, output, use_pack_untilize,
+        //     fp32_dest_acc_en);
+        // }
     }
 
     auto [ncores, all_cores, core_range, core_range_cliff, nblocks_per_core, nblocks_per_core_cliff] =