dump state

bin/defocus.c

@@ -233,7 +233,24 @@ int test_fn( int argc, char **argv ) {
 
     df_camera_i cam = df_create_perspective_camera(1024.0, 1024.0, 1024.0, 1024.0, 2.e-4f, 1000.f, 350.0, 21.0);
     df_ray_packet packet = cam.build_ray_packet(cam.o);
-    df_evaluate_ray_packet(&packet);
+
+    int w = 1024;
+    int h = 1024;
+    df_image *out_image;
+    df_result res = df_create_image(w, h, &out_image);
+    if (res != df_result_success) {
+        fprintf(stderr, "Failed to create output image (%d)\n", (int)res);
+        return 1;
+    }
+
+
+    df_evaluate_ray_packet(&packet, out_image);
+
+    const char *out_path = "out.png";
+    res = df_write_image(out_image, out_path);
+    if (res != df_result_success) {
+        fprintf(stderr, "Failed to write to output image %s (%d)\n", out_path, (int)res);
+    }
 
     df_release_ray_packet(&packet);
     

include/defocus/intrinsic_helper.h

@@ -35,4 +35,22 @@
 /* returns [vec1[2], vec1[3], vec2[2], vec2[3]] */
 #define DF_VEC_SHUFFLE_2323(vec1, vec2) _mm_movehl_ps(vec2, vec1)
 
+/* returns the n-th float inside the vector */
+inline float extract_float(__m128 v, unsigned int n)
+{
+    /* _mm_shuffle_ps expects a constant */
+    switch (n) {
+    case 0:
+        return _mm_cvtss_f32(_mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 0)));
+    case 1:
+        return _mm_cvtss_f32(_mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 1)));
+    case 2:
+        return _mm_cvtss_f32(_mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 2)));
+    case 3:
+        return _mm_cvtss_f32(_mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 3)));
+    }
+    int nan = 0x7F800001;
+    return *(float *)&nan;
+}
+
 #endif

include/defocus/raytracing.h

@@ -1,6 +1,8 @@
 #ifndef DF_RAYTRACING_H
 #define DF_RAYTRACING_H
 
+#include "image.h"
+
 /** @brief Stores information for rays in a SIMD friendly way */
 typedef struct
 {
@@ -24,6 +26,6 @@ typedef struct
 /** @brief Free ray packet memory */
 DF_API void df_release_ray_packet(df_ray_packet *rays);
 
-DF_API void df_evaluate_ray_packet(const df_ray_packet *rays);
+DF_API void df_evaluate_ray_packet(const df_ray_packet *rays, df_image *output);
 
 #endif

lib/camera.c

@@ -16,7 +16,11 @@
 
 DF_API void df_release_ray_packet(df_ray_packet *rays)
 {
+#ifdef _MSC_VER
+    _aligned_free(rays->simd_mem);
+#else
     free(rays->simd_mem);
+#endif
     free(rays->ray_uvs);
 
     memset(rays, 0, sizeof(*rays));

lib/raytracing.c

@@ -1,8 +1,12 @@
 #include <defocus/camera.h>
+#include <defocus/intrinsic_helper.h>
 
 #include <immintrin.h>
 
-DF_API void df_evaluate_ray_packet(const df_ray_packet *rays) {
+#include <stdio.h>
+
+DF_API void df_evaluate_ray_packet(const df_ray_packet *rays, df_image *output)
+{
     const __m128 *base_x = (const __m128 *)rays->base_x; 
     const __m128 *base_y = (const __m128 *)rays->base_y;
     const __m128 *base_z = (const __m128 *)rays->base_z;
@@ -21,12 +25,14 @@ DF_API void df_evaluate_ray_packet(const df_ray_packet *rays) {
 
     /* TODO(kevin): divide to multiple threads */
     for (size_t i = 0; i < ray_count; i += 4) {
-        __m128 rays_base_x = base_x[i];
-        __m128 rays_base_y = base_y[i];
-        __m128 rays_base_z = base_z[i];
-        __m128 rays_dir_x = dir_x[i];
-        __m128 rays_dir_y = dir_y[i];
-        __m128 rays_dir_z = dir_z[i];
+        size_t index = i / 4;
+
+        __m128 rays_base_x = base_x[index];
+        __m128 rays_base_y = base_y[index];
+        __m128 rays_base_z = base_z[index];
+        __m128 rays_dir_x = dir_x[index];
+        __m128 rays_dir_y = dir_y[index];
+        __m128 rays_dir_z = dir_z[index];
 
         /* Solve for t: base.z + t * dir.z = plane_z 
          *              t = (plane_z - base.z) / dir.z 
@@ -42,6 +48,14 @@ DF_API void df_evaluate_ray_packet(const df_ray_packet *rays) {
         sample_p_y = _mm_add_ps(sample_p_y, rays_base_y);
         sample_p_z = _mm_add_ps(sample_p_z, rays_base_z);
 
+        /* Map the sample position to a 2d (uv) point on the plane, so that we can sample a color */
+        df_color color[4];
+        for (int j = 0; j < 4; ++j) {
+            float r = extract_float(sample_p_x, j);
+            float g = extract_float(sample_p_y, j);
+            
+            printf("%f %f\n", r, g);
+        }
     }
 
     /* Handle remaining (< 4) rays */

subprojects/.gitignore

@@ -0,0 +1 @@
+munit