decoder: handle width_padding

added by the commit a12554c (2024-09-24) for encoder

Author: MartinPulec

src/gpujpeg_common.c

@@ -1,6 +1,6 @@
 /**
  * @file
- * Copyright (c) 2011-2024, CESNET
+ * Copyright (c) 2011-2025, CESNET
  * Copyright (c) 2011, Silicon Genome, LLC.
  *
  * All rights reserved.
@@ -1131,7 +1131,7 @@ gpujpeg_image_calculate_size(struct gpujpeg_image_parameters* param)
 {
     int bpp = gpujpeg_pixel_format_get_unit_size(param->pixel_format);
     if (bpp != 0) {
-        return (size_t) param->width * param->height * bpp;
+        return ((size_t) param->width + param->width_padding) * param->height * bpp;
     }
     const int comp_count = gpujpeg_pixel_format_get_comp_count(param->pixel_format);
     switch ( param->pixel_format ) {

src/gpujpeg_decoder.c

@@ -194,6 +194,7 @@ gpujpeg_decoder_init(struct gpujpeg_decoder* decoder, const struct gpujpeg_param
     // Check if (re)inialization is needed
     int change = 0;
     change |= coder->param_image.width != param_image->width;
+    change |= coder->param_image.width_padding != param_image->width_padding;
     change |= coder->param_image.height != param_image->height;
     change |= coder->param.comp_count != param->comp_count;
     change |= coder->param.restart_interval != param->restart_interval;

src/gpujpeg_postprocessor.cu

@@ -457,7 +457,8 @@ gpujpeg_preprocessor_decoder_copy_planar_data(struct gpujpeg_coder * coder, cuda
     size_t data_raw_offset = 0;
     bool needs_stride = false; // true if width is not divisible by MCU width
     for ( int i = 0; i < coder->param.comp_count; ++i ) {
-        needs_stride = needs_stride || coder->component[i].width != coder->component[i].data_width;
+        int component_width = coder->component[i].width + coder->param_image.width_padding;
+        needs_stride = needs_stride || component_width != coder->component[i].data_width;
     }
     if (!needs_stride) {
             for ( int i = 0; i < coder->param.comp_count; ++i ) {
@@ -490,7 +491,7 @@ gpujpeg_postprocessor_decode(struct gpujpeg_coder* coder, cudaStream_t stream)
     gpujpeg_preprocessor_decode_kernel kernel = (gpujpeg_preprocessor_decode_kernel)coder->preprocessor;
     assert(kernel != NULL);
 
-    int image_width = coder->param_image.width;
+    int image_width = coder->param_image.width + coder->param_image.width_padding;
     int image_height = coder->param_image.height;
 
     // When saving 4:2:2 data of odd width, the data should have even width, so round it