Change channel and bits with w_in_stride for set_ptrs

CHANGELOG.md

@@ -12,7 +12,7 @@
 ### Changed
 
 - `ne16_task_init` got split into smaller parts: `ne16_task_init`, `ne16_task_set_op_to_conv`, `ne16_task_set_weight_offset`, `ne16_task_set_bits`, `ne16_task_set_norm_quant`
-- strides in `ne16_task_set_strides`, and `ne16_task_set_dims` are now strides between consecutive elements in that dimension
+- strides in `ne16_task_set_strides`, `ne16_task_set_dims`, and `ne16_task_set_ptrs` are now strides between consecutive elements in that dimension
 - `ne16_task_queue_size` is now `NE16_TASK_QUEUE_SIZE`
 
 ### Removed

ne16/hal/ne16_task.c

@@ -108,19 +108,18 @@ void ne16_task_set_weight_offset(ne16_task_t *task,
  * it was the start to the padded data.
  * Necessary for input pointer when it's padded.
  */
-uint32_t ne16_pad_ptr(uint32_t ptr, const uint32_t width,
-                      const uint32_t channel, const uint8_t bits,
+uint32_t ne16_pad_ptr(uint32_t ptr, const uint32_t width, uint32_t width_stride,
                       const uint8_t padding_top, const uint8_t padding_left) {
-  return ptr - (padding_top * width + padding_left) * channel * bits / 8;
+  return ptr - (padding_top * width + padding_left) * width_stride;
 }
 
 void ne16_task_set_ptrs(ne16_task_t *task, uint32_t input_ptr, uint32_t w_in,
-                        uint32_t k_in, uint8_t bits_in, uint8_t padding_top,
+                        uint32_t w_in_stride, uint8_t padding_top,
                         uint8_t padding_left, uint32_t output_ptr,
                         uint32_t weights_ptr, uint32_t scale_ptr,
                         uint32_t shift_ptr, uint32_t bias_ptr) {
   task->data.infeat_ptr =
-      ne16_pad_ptr(input_ptr, w_in, k_in, bits_in, padding_top, padding_left);
+      ne16_pad_ptr(input_ptr, w_in, w_in_stride, padding_top, padding_left);
   task->data.outfeat_ptr = output_ptr;
   task->data.weights_ptr = weights_ptr;
   task->data.scale_ptr = scale_ptr;

ne16/hal/ne16_task.h

@@ -131,10 +131,10 @@ uint32_t ne16_get_tile_padding(uint32_t padding, uint32_t i_height,
                                uint32_t i_width, uint32_t n_height,
                                uint32_t n_width);
 uint32_t ne16_pad_ptr(uint32_t ptr, const uint32_t width,
-                      const uint32_t channel, const uint8_t bits,
-                      const uint8_t padding_top, const uint8_t padding_left);
+                      const uint32_t width_stride, const uint8_t padding_top,
+                      const uint8_t padding_left);
 void ne16_task_set_ptrs(ne16_task_t *task, uint32_t input_ptr, uint32_t w_in,
-                        uint32_t k_in, uint8_t bits_in, uint8_t padding_top,
+                        uint32_t w_in_stride, uint8_t padding_top,
                         uint8_t padding_left, uint32_t output_ptr,
                         uint32_t weights_ptr, uint32_t scale_ptr,
                         uint32_t shift_ptr, uint32_t bias_ptr);

neureka/hal/neureka_task.c

@@ -122,20 +122,19 @@ void neureka_task_set_weight_source(neureka_task_t *task,
  * Necessary for input pointer when it's padded.
  */
 uint32_t neureka_pad_ptr(uint32_t ptr, const uint32_t width,
-                         const uint32_t channel, const uint8_t bits,
-                         const uint8_t padding_top,
+                         const uint32_t width_stride, const uint8_t padding_top,
                          const uint8_t padding_left) {
   return ptr - (padding_top * width + padding_left) * channel * bits / 8;
 }
 
 void neureka_task_set_ptrs(neureka_task_t *task, uint32_t input_ptr,
-                           uint32_t w_in, uint32_t k_in, uint8_t bits_in,
+                           uint32_t w_in, uint32_t w_in_stride,
                            uint8_t padding_top, uint8_t padding_left,
                            uint32_t output_ptr, uint32_t weights_ptr,
                            uint32_t scale_ptr, uint32_t shift_ptr,
                            uint32_t bias_ptr) {
-  task->data.infeat_ptr = neureka_pad_ptr(input_ptr, w_in, k_in, bits_in,
-                                          padding_top, padding_left);
+  task->data.infeat_ptr =
+      neureka_pad_ptr(input_ptr, w_in, w_in_stride, padding_top, padding_left);
   task->data.outfeat_ptr = output_ptr;
   task->data.weights_ptr = weights_ptr;
   task->data.scale_ptr = scale_ptr;

neureka/hal/neureka_task.h

@@ -140,10 +140,10 @@ uint32_t neureka_get_tile_padding(uint32_t padding, uint32_t i_height,
                                   uint32_t i_width, uint32_t n_height,
                                   uint32_t n_width);
 uint32_t neureka_pad_ptr(uint32_t ptr, const uint32_t width,
-                         const uint32_t channel, const uint8_t bits,
-                         const uint8_t padding_top, const uint8_t padding_left);
+                         const uint32_t width_stride, const uint8_t padding_top,
+                         const uint8_t padding_left);
 void neureka_task_set_ptrs(neureka_task_t *task, uint32_t input_ptr,
-                           uint32_t w_in, uint32_t k_in, uint8_t bits_in,
+                           uint32_t w_in, uint32_t w_in_stride,
                            uint8_t padding_top, uint8_t padding_left,
                            uint32_t output_ptr, uint32_t weights_ptr,
                            uint32_t scale_ptr, uint32_t shift_ptr,

test/app/src/nnx_layer.c

@@ -167,8 +167,8 @@ static void task_prepare(nnx_task_t *task) {
                     PADDING_LEFT);
 #endif
 
-  nnx_task_set_ptrs(task, (uint32_t)input, INPUT_WIDTH, INPUT_CHANNEL,
-                    INPUT_BITS, PADDING_TOP, PADDING_LEFT, (uint32_t)output,
+  nnx_task_set_ptrs(task, (uint32_t)input, INPUT_WIDTH, w_in_stride,
+                    PADDING_TOP, PADDING_LEFT, (uint32_t)output,
                     (uint32_t)weight,
 #if HAS_NORM_QUANT == 1
                     (uint32_t)scale, NULL,