minor update

driver/tests/tex_demo/kernel.c

@@ -52,8 +52,8 @@ int main() {
 	targ.karg        = *arg;
 	targ.tile_width  = arg->dst_width;
 	targ.tile_height = (arg->dst_height + arg->num_tasks - 1) / arg->num_tasks;    
-	targ.deltaX      = 1.0f / (((float)arg->src_width) / arg->dst_width);
+	targ.deltaX      = 1.0f / arg->dst_width;
-	targ.deltaY      = 1.0f / (((float)arg->src_height) / arg->dst_height);
+	targ.deltaY      = 1.0f / arg->dst_height;
 	
 	vx_spawn_tasks(arg->num_tasks, kernel_body, &targ);
 }

driver/tests/tex_demo/main.cpp

@@ -20,7 +20,7 @@
 ///////////////////////////////////////////////////////////////////////////////
 
 const char* kernel_file = "kernel.bin";
-const char* input_file  = "palette.tga";
+const char* input_file  = "toad.tga";
 const char* output_file = "output.tga";
 float scale = 1.0f;
 

driver/tests/tex_demo/utils.cpp

@@ -63,11 +63,13 @@ int LoadTGA(const char *filename,
         return -1;
       }
 
-      // Because the TGA is BGR instead of RGB, we must swap RG components
+      // TGA uses BGR instead of RGB, we must swap RG components
-      for (int i = 0; i < pitch; i += stride) {
+      if (stride >= 3) {
-        auto tmp = line[i];
+        for (int i = 0; i < pitch; i += stride) {
-        line[i] = line[i + 2];
+          auto tmp = line[i];
-        line[i + 2] = tmp;
+          line[i] = line[i + 2];
+          line[i + 2] = tmp;
+        }
       }
     }
     break;
@@ -116,7 +118,19 @@ int SaveTGA(const char *filename,
   for (uint32_t y = 0; y < height; ++y) {
     const uint8_t* pixel_row = pixel_bytes;
     for (uint32_t x = 0; x < width; ++x) {
-      ofs.write((const char*)pixel_row, bpp);
+      // TGA uses BGR instead of RGB, we must swap RG components
+      if (bpp == 4) {
+        ofs.write((const char*)pixel_row + 2, 1);
+        ofs.write((const char*)pixel_row + 1, 1);
+        ofs.write((const char*)pixel_row + 0, 1);
+        ofs.write((const char*)pixel_row + 3, 1);
+      } else if (bpp == 3) {
+        ofs.write((const char*)pixel_row + 2, 1);
+        ofs.write((const char*)pixel_row + 1, 1);
+        ofs.write((const char*)pixel_row + 0, 1);
+      } else{
+        ofs.write((const char*)pixel_row, bpp);
+      }
       pixel_row += bpp;
     }
     pixel_bytes -= pitch;

hw/rtl/tex_unit/VX_tex_sampler.v

@@ -46,7 +46,7 @@ module VX_tex_sampler #(
         VX_tex_format #(
             .CORE_ID (CORE_ID),
             .NUM_TEXELS (4)
-        ) tex_format_texel (
+        ) tex_format (
             .texel_data  (req_texels[i]),
             .format (req_format),