tex_unit update

driver/tests/tex_demo/common.h

@@ -9,7 +9,7 @@ struct kernel_arg_t {
   uint32_t src_height;
   uint32_t src_stride;
   uint32_t src_pitch;
-  uint32_t src_ptr;
+  uint8_t src_ptr;
   uint32_t dst_width;
   uint32_t dst_height;
   uint32_t dst_stride;  

driver/tests/tex_demo/kernel.c

@@ -16,18 +16,19 @@ struct tile_arg_t {
 void kernel_body(int task_id, void* arg) {
 	struct tile_arg_t* _arg = (struct tile_arg_t*)(arg);
 	
-	uint32_t xoffset = task_id * _arg->tile_width;
+	uint32_t xoffset = 0;
 	uint32_t yoffset = task_id * _arg->tile_height;	
-	uint32_t* dst_ptr = (uint32_t*)_arg->karg.dst_ptr + xoffset + yoffset * _arg->karg.dst_pitch;
+	uint8_t* dst_ptr = (uint8_t*)(_arg->karg.dst_ptr + xoffset * _arg->karg.dst_stride + yoffset * _arg->karg.dst_pitch);
 
 	float fu = xoffset * _arg->deltaX;
 	float fv = yoffset * _arg->deltaY;
 
 	for (uint32_t y = 0; y < _arg->tile_height; ++y) {
+		uint32_t* dst_row = (uint32_t*)dst_ptr;
 		for (uint32_t x = 0; x < _arg->tile_width; ++x) {
 			int32_t u = (int32_t)(fu * (1<<20));
 			int32_t v = (int32_t)(fv * (1<<20));
-			dst_ptr[x] = vx_tex(0, u, v, 0x0);
+			dst_row[x] = vx_tex(0, u, v, 0x0);
 			fu += _arg->deltaX;
 		}
 		dst_ptr += _arg->karg.dst_pitch;

driver/tests/tex_demo/main.cpp

@@ -20,7 +20,7 @@
 ///////////////////////////////////////////////////////////////////////////////
 
 const char* kernel_file = "kernel.bin";
-const char* input_file  = "sample.tga";
+const char* input_file  = "palette.tga";
 const char* output_file = "output.tga";
 float scale = 1.0f;
 
@@ -69,7 +69,7 @@ void cleanup() {
   }
 }
 
-int run_test(const kernel_arg_t& kernel_arg, uint32_t buf_size, uint32_t width, uint32_t height, uint32_t dst_bpp) {
+int run_test(const kernel_arg_t& kernel_arg, uint32_t buf_size, uint32_t width, uint32_t height, uint32_t bpp) {
   // start device
   std::cout << "start device" << std::endl;
   RT_CHECK(vx_start(device));
@@ -83,14 +83,15 @@ int run_test(const kernel_arg_t& kernel_arg, uint32_t buf_size, uint32_t width,
   RT_CHECK(vx_copy_from_dev(buffer, kernel_arg.dst_ptr, buf_size, 0));
 
   std::vector<uint8_t> dst_pixels(buf_size);
-  auto buf_ptr = (int8_t*)vx_host_ptr(buffer);
+  auto buf_ptr = (uint8_t*)vx_host_ptr(buffer);
   for (uint32_t i = 0; i < buf_size; ++i) {
     dst_pixels[i] = buf_ptr[i];
   } 
 
   // save output image
   std::cout << "save output image" << std::endl;  
-  RT_CHECK(SaveTGA(output_file, dst_pixels, width, height, dst_bpp));
+  dump_image(dst_pixels, width, height, bpp);
+  RT_CHECK(SaveTGA(output_file, dst_pixels, width, height, bpp));
 
   return 0;
 }
@@ -106,6 +107,7 @@ int main(int argc, char *argv[]) {
   parse_args(argc, argv);
 
   RT_CHECK(LoadTGA(input_file, src_pixels, &src_width, &src_height, &src_bpp));
+  dump_image(src_pixels, src_width, src_height, src_bpp);
   uint32_t src_bufsize = src_bpp * src_width * src_height;
 
   uint32_t dst_width   = (uint32_t)(src_width * scale);
@@ -122,9 +124,9 @@ int main(int argc, char *argv[]) {
   RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_WARPS, &max_warps));
   RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_THREADS, &max_threads));
 
-  uint32_t num_tasks = max_cores * max_warps * max_threads;
+  uint32_t num_tasks = max_cores * max_warps * max_threads / 4;
 
-  std::cout << "number of tasks: " << num_tasks << std::endl;
+  std::cout << "number of tasks: " << std::dec << num_tasks << std::endl;
   std::cout << "source buffer: width=" << src_width << ", heigth=" << src_height << ", size=" << src_bufsize << " bytes" << std::endl;
   std::cout << "destination buffer: width=" << dst_width << ", heigth=" << dst_height << ", size=" << dst_bufsize << " bytes" << std::endl;
 
@@ -138,8 +140,8 @@ int main(int argc, char *argv[]) {
   RT_CHECK(vx_alloc_dev_mem(device, src_bufsize, &src_addr));
   RT_CHECK(vx_alloc_dev_mem(device, dst_bufsize, &dst_addr));
 
-  std::cout << "src_addr=" << std::hex << src_addr << std::endl;
-  std::cout << "dst_addr=" << std::hex << dst_addr << std::endl;
+  std::cout << "src_addr=0x" << std::hex << src_addr << std::endl;
+  std::cout << "dst_addr=0x" << std::hex << dst_addr << std::endl;
 
   // allocate staging shared memory  
   std::cout << "allocate shared memory" << std::endl;    
@@ -154,13 +156,13 @@ int main(int argc, char *argv[]) {
     kernel_arg.src_width   = src_width;
     kernel_arg.src_height  = src_height;
     kernel_arg.src_stride  = src_bpp;
-    kernel_arg.src_pitch   = src_bpp * src_width * src_height;
+    kernel_arg.src_pitch   = src_bpp * src_width;
     kernel_arg.src_ptr     = src_addr;
 
     kernel_arg.dst_width   = dst_width;
     kernel_arg.dst_height  = dst_height;
     kernel_arg.dst_stride  = dst_bpp;
-    kernel_arg.dst_pitch   = dst_bpp * dst_width * dst_height;    
+    kernel_arg.dst_pitch   = dst_bpp * dst_width;    
     kernel_arg.dst_ptr     = dst_addr;
 
     auto buf_ptr = (int*)vx_host_ptr(buffer);

driver/tests/tex_demo/utils.cpp

@@ -1,5 +1,6 @@
 #include "utils.h"
 #include <fstream>
+#include <assert.h>
 
 struct __attribute__((__packed__)) tga_header_t {
   int8_t idlength;
@@ -108,8 +109,35 @@ int SaveTGA(const char *filename,
   header.bitsperpixel = bpp * 8;
   header.imagedescriptor = 0;
 
-  ofs.write(reinterpret_cast<char *>(&header), sizeof(tga_header_t));  
-  ofs.write((const char*)pixels.data(), pixels.size());
+  ofs.write(reinterpret_cast<char *>(&header), sizeof(tga_header_t));
+
+  uint32_t pitch = bpp * width;
+  const uint8_t* pixel_bytes = pixels.data() + (height - 1) * pitch;
+  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);
+      pixel_row += bpp;
+    }
+    pixel_bytes -= pitch;
+  }
 
   return 0;
+}
+
+void dump_image(const std::vector<uint8_t>& pixels, uint32_t width, uint32_t height, uint32_t bpp) {
+  assert(width * height * bpp == pixels.size());
+  const uint8_t* pixel_bytes = pixels.data();
+  for (uint32_t y = 0; y < height; ++y) {
+    for (uint32_t x = 0; x < width; ++x) {
+      uint32_t pixel32 = 0;
+      for (uint32_t b = 0; b < bpp; ++b) {
+        uint32_t pixel8 = *pixel_bytes++;
+        pixel32 |= pixel8 << (b * 8);
+      }
+      if (x) std::cout << ", ";
+      std::cout << std::hex << pixel32;
+    }
+    std::cout << std::endl;
+  }
 }

driver/tests/tex_demo/utils.h

@@ -12,4 +12,6 @@ int SaveTGA(const char *filename,
             const std::vector<uint8_t> &pixels, 
             uint32_t width,
             uint32_t height, 
-            uint32_t bpp);
+            uint32_t bpp);
+
+void dump_image(const std::vector<uint8_t>& pixels, uint32_t width, uint32_t height, uint32_t bpp);