Fix incorrect output address calculation

src/ispc/kernels/lanczos3.ispc

@@ -136,6 +136,7 @@ void resample_with_cache(uniform uint32 num_channels, uniform uint32 src_width,
     // Accumulate only along the width for each pixel, sampling from the source image
     // Results in the source image being downsampled to src_height X target_width
     foreach_tiled(y =  0 ... src_height, x = 0 ... target_width) {
+
         uint32 src_width_start = horizontal_weight_collection->starts[x];
         uint32 num_horizontal_weights = horizontal_weight_collection->weight_counts[x];
         float* horizontal_weights = horizontal_weight_collection->values[x];
@@ -160,15 +161,13 @@ void resample_with_cache(uniform uint32 num_channels, uniform uint32 src_width,
         else
             clean_and_write_4_channels(color4, scratch_write_address, scratch_space);
     }
-
     // Accumulate the scratch space data along the height
     // Downsamples the src_height X target_width image to target_height * target_width
     foreach_tiled(y =  0 ... target_height, x = 0 ... target_width) {
 
         uint32 src_height_start = vertical_weight_collection->starts[y];
         uint32 num_vertical_weights = vertical_weight_collection->weight_counts[y];
         float* vertical_weights = vertical_weight_collection->values[y];
-
         float<3> color3 = {0.0f, 0.0f, 0.0f};
         float<4> color4 = {0.0f, 0.0f, 0.0f, 0.0f};
         for (uint32 i = 0; i < num_vertical_weights; i++) {
@@ -185,8 +184,8 @@ void resample_with_cache(uniform uint32 num_channels, uniform uint32 src_width,
                 color4 += sample_4_channels(scratch_read_address, scratch_space) * weight;
         }
 
-        uint64 out_write_address = (y * target_height + x) * num_channels;
-
+        uint64 out_write_address = (y * target_width + x) * num_channels;
+        assert(out_write_address < target_height * target_width * num_channels);
         if (num_channels == 3)
             clean_and_write_3_channels(color3, out_write_address, out_data);
         else

src/lib.rs

@@ -200,6 +200,7 @@ pub fn downsample_with_custom_scale(
         vertical_weights: height_weights.ispc_representation(),
         horizontal_weights: width_weights.ispc_representation(),
     };
+
     unsafe {
         if src.format.num_channels() == 3 {
             ispc::downsample_ispc::resample_with_cache_3(