Loosen up a bound in ResampleEffectTest a bit.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Fri, 29 Dec 2017 23:18:03 +0000 (00:18 +0100)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Fri, 29 Dec 2017 23:18:03 +0000 (00:18 +0100)
resample_effect_test.cpp

index 84f61e7..4faa810 100644 (file)
@@ -164,8 +164,13 @@ TEST(ResampleEffectTest, UpscaleByThreeGetsCorrectPixelCenters) {
        tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
 
        // We only bother checking that the middle pixel is still correct,
-       // and that symmetry holds.
-       EXPECT_FLOAT_EQ(1.0, out_data[7 * (size * 3) + 7]);
+       // and that symmetry holds. Note that the middle weight in practice
+       // becomes something like 0.99999 due to the normalization
+       // (some supposedly zero weights become 1e-6 or so), and then after
+       // squaring, the error compounds. Ironically, less texture precision
+       // here will give a more accurate result, since the weight can get
+       // rounded towards 1.0.
+       EXPECT_NEAR(1.0, out_data[7 * (size * 3) + 7], 1e-3);
        for (unsigned y = 0; y < size * 3; ++y) {
                for (unsigned x = 0; x < size * 3; ++x) {
                        EXPECT_NEAR(out_data[y * (size * 3) + x], out_data[(size * 3 - y - 1) * (size * 3) + x], 1e-6);