don't run (sub) function when height or widht is zero

Author: JaccovG

lib/src/kernels/convolution/mli_krn_conv2d_hwc.h

@@ -386,6 +386,8 @@ static __attribute__ ((always_inline)) void depthwise_convolution2D_hwcn_krnpad(
     perception_area_nopad.clmn_beg = CEIL_DIV(padding_left, stride_width);
     perception_area_nopad.clmn_end = out_width - CEIL_DIV(padding_right, stride_width);
 
+    if ((perception_area_nopad.row_end - perception_area_nopad.row_beg > 0)
+        && (perception_area_nopad.clmn_end - perception_area_nopad.clmn_beg > 0)){
     depthwise_convolution2D_hwcn_nopad<int8_t, int8_t, int32_t, mli_acc32_t>(
                 in_ftrs, weights, biases, out_ftrs, &perception_area_nopad, quant_params,
                     val_min_limit, val_max_limit,
@@ -395,7 +397,7 @@ static __attribute__ ((always_inline)) void depthwise_convolution2D_hwcn_krnpad(
                 stride_height, stride_width,
                 padding_top, padding_left,
                 padding_bot, padding_right);
-
+    }
     // Phase 2: Process border part with more complex algorithm
     // (usually significantly smaller part of computations)
     //=======================================================================
@@ -680,8 +682,10 @@ static __attribute__ ((always_inline)) void convolution2D_nhwc_krnpad(
     perception_area_nopad.row_end = out_height - CEIL_DIV(padding_bot, stride_height);
     perception_area_nopad.clmn_beg = CEIL_DIV(padding_left, stride_width);
     perception_area_nopad.clmn_end = out_width - CEIL_DIV(padding_right, stride_width);
-    
-    convolution2D_nhwc_nopad<int8_t, int8_t, int32_t, mli_acc32_t>(
+
+    if ((perception_area_nopad.row_end - perception_area_nopad.row_beg > 0)
+        && (perception_area_nopad.clmn_end - perception_area_nopad.clmn_beg > 0)){
+        convolution2D_nhwc_nopad<int8_t, int8_t, int32_t, mli_acc32_t>(
                 in_ftrs, weights, biases, out_ftrs, &perception_area_nopad, quant_params,
                 val_min_limit, val_max_limit,
                 in_ch, in_width, in_height,
@@ -690,6 +694,7 @@ static __attribute__ ((always_inline)) void convolution2D_nhwc_krnpad(
                 stride_height, stride_width,
                 padding_top, padding_left,
                 padding_bot, padding_right);
+    }
 
     // Phase 2: Process border part with more complex algorithm
     // (usually significantly smaller part of computations)

lib/src/kernels/pooling/mli_krn_avepool_hwc.h

@@ -194,12 +194,13 @@ static inline void __attribute__((always_inline)) avepool_hwc_krnpad(
         const int padding_bot) {
     // Phase 1: Process central part (without border effects - padding free)
     //=======================================================================
-    {
-        const int row_beg = CEIL_DIV(padding_top, stride_height);
-        const int row_end = out_height - CEIL_DIV(padding_bot, stride_height);
-        const int clmn_beg = CEIL_DIV(padding_left, stride_width);
-        const int clmn_end = out_width - CEIL_DIV(padding_right, stride_width);
 
+    const int row_beg = CEIL_DIV(padding_top, stride_height);
+    const int row_end = out_height - CEIL_DIV(padding_bot, stride_height);
+    const int clmn_beg = CEIL_DIV(padding_left, stride_width);
+    const int clmn_end = out_width - CEIL_DIV(padding_right, stride_width);
+
+    if ((row_end - row_beg > 0) && (clmn_end - clmn_beg > 0)) {
         avepool_hwc_nopad(
                 row_beg, row_end, clmn_beg, clmn_end,
                 in_ftrs, out_ftrs, channels, in_width,

lib/src/kernels/pooling/mli_krn_maxpool_hwc.h

@@ -191,13 +191,20 @@ static inline void __attribute__((always_inline)) maxpool_hwc_pad(
 
     // Phase 1: Process central part (without border effects - padding free)
     //=======================================================================
-    maxpool_hwc_nopad<io_T>(
-        row_beg, row_end, clmn_beg, clmn_end,
-        stride_width, stride_height, padding_top,
-        padding_bot, padding_left, padding_right, 
-        in_ftrs, out_ftrs, channels_num, kernel_height,
-        kernel_width, in_height, in_width, out_width,
-        out_height);
+    int row_beg = CEIL_DIV(padding_top, stride_height);
+    int row_end = out_height - CEIL_DIV(padding_bot, stride_height);
+    int clmn_beg = CEIL_DIV(padding_left, stride_width);
+    int clmn_end = out_width - CEIL_DIV(padding_right, stride_width);
+
+    if ((row_end - row_beg > 0) && (clmn_end - clmn_beg > 0)) {
+        maxpool_hwc_nopad<io_T>(
+            row_beg, row_end, clmn_beg, clmn_end,
+            stride_width, stride_height, padding_top,
+            padding_bot, padding_left, padding_right,
+            in_ftrs, out_ftrs, channels_num, kernel_height,
+            kernel_width, in_height, in_width, out_width,
+            out_height);
+    }
     // Phase 2: Process border part with more complex algorithm
     // (usually significantly smaller part of computations)
     //=======================================================================