minor update

driver/tests/tex_demo/common.h

@@ -5,17 +5,18 @@
 
 struct kernel_arg_t {
   uint32_t num_tasks;
-  uint32_t format;
+  uint8_t format;
-  uint32_t filter;
+  uint8_t filter;
-  uint32_t wrap;
+  uint8_t wrap;
-  uint32_t src_logWidth;
+  uint8_t use_sw;
-  uint32_t src_logHeight;
+  uint8_t src_logWidth;
-  uint32_t src_stride;
+  uint8_t src_logHeight;
-  uint32_t src_pitch;
+  uint8_t src_stride;
-  uint8_t src_ptr;
+  uint8_t src_pitch;
+  uint32_t src_ptr;
   uint32_t dst_width;
   uint32_t dst_height;
-  uint32_t dst_stride;  
+  uint8_t dst_stride;  
   uint32_t dst_pitch;
   uint32_t dst_ptr;  
 };

driver/tests/tex_demo/kernel.c

@@ -3,9 +3,7 @@
 #include "common.h"
 #include "texsw.h"
 
-uint32_t ilog2 (uint32_t value) {
+#define ENABLE_SW
-  	return (uint32_t)(sizeof(uint32_t) * 8UL) - (uint32_t)__builtin_clzl((value << 1) - 1UL) - 1;
-}
 
 struct tile_arg_t {
   	struct kernel_arg_t* state;	
@@ -30,8 +28,15 @@ void kernel_body(int task_id, void* arg) {
 		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_row[x] = tex_sw(state, 0, u, v, 0x0);
+		#ifdef ENABLE_SW
+			if (state->use_sw) {
+				dst_row[x] = tex_sw(state, 0, u, v, 0x0);
+			} else {
+		#endif
 			dst_row[x] = vx_tex(0, u, v, 0x0);
+		#ifdef ENABLE_SW
+			}
+		#endif
 			fu += _arg->deltaX;
 		}
 		dst_ptr += state->dst_pitch;

driver/tests/tex_demo/main.cpp

@@ -3,6 +3,7 @@
 #include <unistd.h>
 #include <string.h>
 #include <chrono>
+#include <cmath>
 #include <assert.h>
 #include <vortex.h>
 #include "common.h"
@@ -27,6 +28,7 @@ int wrap    = 0;
 int filter  = 0;
 float scale = 1.0f;
 int format  = 0;
+bool use_sw = false;
 ePixelFormat eformat = FORMAT_A8R8G8B8;
 
 vx_device_h device = nullptr;
@@ -34,7 +36,7 @@ vx_buffer_h buffer = nullptr;
 
 static void show_usage() {
    std::cout << "Vortex Texture Test." << std::endl;
-   std::cout << "Usage: [-k: kernel] [-i image] [-o image] [-s scale] [-w wrap] [-f format] [-g filter] [-h: help]" << std::endl;
+   std::cout << "Usage: [-k: kernel] [-i image] [-o image] [-s scale] [-w wrap] [-f format] [-g filter] [-z no_hw] [-h: help]" << std::endl;
 }
 
 static void parse_args(int argc, char **argv) {
@@ -53,6 +55,9 @@ static void parse_args(int argc, char **argv) {
     case 'w':
       wrap = std::atoi(optarg);
       break;
+    case 'z':
+      use_sw = std::atoi(optarg);
+      break;
     case 'f': {
       format  = std::atoi(optarg);
       switch (format) {
@@ -124,7 +129,7 @@ int run_test(const kernel_arg_t& kernel_arg,
 
   // save output image
   std::cout << "save output image" << std::endl;  
-  //dump_image(dst_pixels, width, height, bpp);
+  dump_image(dst_pixels, width, height, bpp);
   RT_CHECK(SaveTGA(output_file, dst_pixels, width, height, bpp));
 
   return 0;
@@ -201,9 +206,10 @@ int main(int argc, char *argv[]) {
     kernel_arg.format     = format;
     kernel_arg.filter     = filter;
     kernel_arg.wrap       = wrap;
+    kernel_arg.use_sw     = use_sw;
     
-    kernel_arg.src_logWidth  = ilog2(src_width);
+    kernel_arg.src_logWidth  = (uint32_t)std::log2(src_width);
-    kernel_arg.src_logHeight = ilog2(src_height);
+    kernel_arg.src_logHeight = (uint32_t)std::log2(src_height);
     kernel_arg.src_stride = src_bpp;
     kernel_arg.src_pitch  = src_bpp * src_width;
     kernel_arg.src_ptr    = src_addr;