Add SSE/AVX2 version of rgb_to_xyz().sse2

1 files changed, 81 insertions, 27 deletions

diff --git a/test/rgb_xyz_test.cc b/test/rgb_xyz_test.cc
index 5dcfe673..961edc95 100644
--- a/test/rgb_xyz_test.cc
+++ b/test/rgb_xyz_test.cc
@@ -45,25 +45,46 @@ using std::cout;
 using boost::shared_ptr;
 using boost::optional;
 using boost::scoped_array;
+using boost::shared_array;
 
-/** Convert a test image from sRGB to XYZ and check that the transforms are right */
-BOOST_AUTO_TEST_CASE (rgb_xyz_test)
+template <class T>
+void
+input_gamma (T& v)
+{
+	if (v < 0.04045) {
+		v /= 12.92;
+	} else {
+		v = pow ((v + 0.055) / 1.055, 2.4);
+	}
+}
+
+
+shared_array<uint8_t>
+random_data (dcp::Size size, int components)
 {
 	srand (0);
-	dcp::Size const size (640, 480);
+	shared_array<uint8_t> data (new uint8_t[size.width * size.height * components * 2]);
 
-	scoped_array<uint8_t> rgb (new uint8_t[size.width * size.height * 6]);
 	for (int y = 0; y < size.height; ++y) {
-		uint16_t* p = reinterpret_cast<uint16_t*> (rgb.get() + y * size.width * 6);
+		uint16_t* p = reinterpret_cast<uint16_t*> (data.get() + y * size.width * components * 2);
 		for (int x = 0; x < size.width; ++x) {
 			/* Write a 12-bit random number for each component */
-			for (int c = 0; c < 3; ++c) {
-				*p = (rand () & 0xfff) << 4;
-				++p;
+			for (int c = 0; c < components; ++c) {
+				*p++ = (rand () & 0xfff) << 4;
 			}
 		}
 	}
 
+	return data;
+}
+
+
+/** Convert a test image from sRGB to XYZ and check that the transforms are right */
+BOOST_AUTO_TEST_CASE (rgb_xyz_test)
+{
+	dcp::Size const size (640, 480);
+
+	shared_array<uint8_t> rgb = random_data (size, 3);
 	shared_ptr<dcp::OpenJPEGImage> xyz = dcp::rgb_to_xyz (rgb.get(), size, size.width * 6, dcp::ColourConversion::srgb_to_xyz ());
 
 	for (int y = 0; y < size.height; ++y) {
@@ -74,25 +95,9 @@ BOOST_AUTO_TEST_CASE (rgb_xyz_test)
 			double cg = *p++ / 65535.0;
 			double cb = *p++ / 65535.0;
 
-			/* Input gamma */
-
-			if (cr < 0.04045) {
-				cr /= 12.92;
-			} else {
-				cr = pow ((cr + 0.055) / 1.055, 2.4);
-			}
-
-			if (cg < 0.04045) {
-				cg /= 12.92;
-			} else {
-				cg = pow ((cg + 0.055) / 1.055, 2.4);
-			}
-
-			if (cb < 0.04045) {
-				cb /= 12.92;
-			} else {
-				cb = pow ((cb + 0.055) / 1.055, 2.4);
-			}
+			input_gamma (cr);
+			input_gamma (cg);
+			input_gamma (cb);
 
 			/* Matrix */
 
@@ -119,6 +124,55 @@ BOOST_AUTO_TEST_CASE (rgb_xyz_test)
 	}
 }
 
+
+/** Convert a test image from sRGB to XYZ using the SSE/AVX code and check that the transforms are right */
+BOOST_AUTO_TEST_CASE (rgb_xyz_test_avx2)
+{
+	srand (0);
+	dcp::Size const size (647, 481);
+
+	shared_array<uint8_t> rgb = random_data (size, 4);
+	shared_ptr<dcp::OpenJPEGImage> xyz = dcp::rgb_to_xyz_avx2 (rgb.get(), size, dcp::ColourConversion::srgb_to_xyz());
+
+	uint16_t* p = reinterpret_cast<uint16_t*> (rgb.get());
+	for (int y = 0; y < size.height; ++y) {
+		for (int x = 0; x < size.width; ++x) {
+
+			float cr = *p++ / 65535.0;
+			float cg = *p++ / 65535.0;
+			float cb = *p++ / 65535.0;
+			p++;
+
+			input_gamma (cr);
+			input_gamma (cg);
+			input_gamma (cb);
+
+			/* Matrix */
+
+			float cx = cr * 0.4124564 + cg * 0.3575761 + cb * 0.1804375;
+			float cy = cr * 0.2126729 + cg * 0.7151522 + cb * 0.0721750;
+			float cz = cr * 0.0193339 + cg * 0.1191920 + cb * 0.9503041;
+
+			/* Compand */
+
+			cx *= 48 / 52.37;
+			cy *= 48 / 52.37;
+			cz *= 48 / 52.37;
+
+			/* Output gamma */
+
+			cx = pow (cx, 1 / 2.6);
+			cy = pow (cy, 1 / 2.6);
+			cz = pow (cz, 1 / 2.6);
+
+			BOOST_REQUIRE_CLOSE (cx * 4095, xyz->data(0)[y * size.width + x], 3);
+			BOOST_REQUIRE_CLOSE (cy * 4095, xyz->data(1)[y * size.width + x], 3);
+			BOOST_REQUIRE_CLOSE (cz * 4095, xyz->data(2)[y * size.width + x], 3);
+		}
+	}
+}
+
+
 static list<string> notes;
 
 static void