graphics: Rendering a scaled image might miss edge pixels

[hajimehoshi]

This test fails on MacBook Pro:

func TestImageStretch(t *testing.T) {
        img0, _ := NewImage(16, 17, FilterDefault)

        pix := make([]byte, 4*16*17)
        for i := 0; i < 16*16; i++ {
                pix[4*i] = 0xff
                pix[4*i+3] = 0xff
        }
        for i := 0; i < 16; i++ {
                pix[4*(16*16+i)+1] = 0xff
                pix[4*(16*16+i)+3] = 0xff
        }
        img0.ReplacePixels(pix)

        const h = 4096
        img1, _ := NewImage(16, h, FilterDefault)
        for i := 1; i < h; i++ {
                img1.Clear()
                op := &DrawImageOptions{}
                op.GeoM.Scale(1, float64(i)/16)
                r := image.Rect(0, 0, 16, 16)
                op.SourceRect = &r
                img1.DrawImage(img0, op)
                for j := -1; j <= 1; j++ {
                        got := img1.At(0, i+j).(color.RGBA)
                        want := color.RGBA{}
                        if j < 0 {
                                want = color.RGBA{0xff, 0, 0, 0xff}
                        }
                        if got != want {
                                t.Errorf("At(%d, %d) (i=%d): got: %#v, want: %#v", 0, i+j, i, got, want)
                        }
                }
        }
}

[hajimehoshi]

--- FAIL: TestImageStretch (2.43s)
image_test.go:830: At(0, 4093) (i=4094): got: color.RGBA{R:0x0, G:0x0, B:0x0, A:0x0}, want: color.RGBA{R:0xff, G:0x0, B:0x0, A:0xff}
image_test.go:830: At(0, 4094) (i=4095): got: color.RGBA{R:0x0, G:0x0, B:0x0, A:0x0}, want: color.RGBA{R:0xff, G:0x0, B:0x0, A:0xff}
FAIL
FAIL    github.com/hajimehoshi/ebiten   2.970s

[hajimehoshi]

The situation was more complicated than I thought. On MacBook Pro 2018, the result depends on what kind of tests are executed at the same time. I think this is a bug on the GPU, but am not sure.