// vector of 4 32-bit unsigned integers
func (d Uint32Div) Div(a M128i) M128i {
t1 := sse2.MulEpu32(a, u.multiplier) // 32x32->64 bit unsigned multiplication of a[0] and a[2]
t2 := sse2.SrliEpi64(t1, 32) // high dword of result 0 and 2
t3 := sse2.SrliEpi64(a, 32) // get a[1] and a[3] into position for multiplication
t4 := sse2.MulEpu32(t3, u.multiplier) // 32x32->64 bit unsigned multiplication of a[1] and a[3]
t5 := sse2.SetEpi32(-1, 0, -1, 0) // mask of dword 1 and 3
t6 := sse2.AndSi128(t4, t5) // high dword of result 1 and 3
t7 := sse2.OrSi128(t2, t6) // combine all four results into one vector
t8 := sse2.SubEpi32(a, t7) // subtract
t9 := sse2.SrlEpi32(t8, u.shift1) // shift right logical
t10 := sse2.AddEpi32(t7, t9) // add
return sse2.SrlEpi32(t10, d.u.shift2) // shift right logical
}
// Mixed sse+sse2
// Applies a 3x3 matrix and sRGB gamma corrects input
// and writes it to output.
// https://github.com/rawstudio/rawstudio/blob/master/plugins/colorspace-transform/colorspace_transform_sse2.c#L205
func Transform8sRGB() {
w := 1048
h := 1024
var matrix [3][3]float32
// Input is a w * h * 4 components 16 bit per component image
var input = make([]byte, w*h*8)
// Output is 4 byte/pixel, RGBA image
var output = make([]byte, w*h*4)
// The matrix with values splatted to all registers
var matPs = make([]x86.M128, 3*3)
// Fill with values
matPs[0] = sse.Set1Ps(matrix[0][0])
matPs[1] = sse.Set1Ps(matrix[0][1])
matPs[2] = sse.Set1Ps(matrix[0][2])
matPs[3] = sse.Set1Ps(matrix[1][0])
matPs[4] = sse.Set1Ps(matrix[1][1])
matPs[5] = sse.Set1Ps(matrix[1][2])
matPs[6] = sse.Set1Ps(matrix[2][0])
matPs[7] = sse.Set1Ps(matrix[2][1])
matPs[8] = sse.Set1Ps(matrix[2][2])
for y := 0; y < h; y++ {
i := x86.BytesToM128i(input[y*w*8 : y*w*8+w*8])
o := x86.BytesToM128i(output[y*w*4 : y*w*4+w*4])
// Converts 4 pixels per loop
for x := 0; x < w/4; x++ {
/* Load and convert to float */
zero := sse2.SetzeroSi128()
in := i[x*2] // Load two pixels
in2 := i[x*2+1] // Load two pixels
p1 := sse2.UnpackloEpi16(in, zero)
p2 := sse2.UnpackhiEpi16(in, zero)
p3 := sse2.UnpackloEpi16(in2, zero)
p4 := sse2.UnpackhiEpi16(in2, zero)
p1f := sse2.Cvtepi32Ps(p1)
p2f := sse2.Cvtepi32Ps(p2)
p3f := sse2.Cvtepi32Ps(p3)
p4f := sse2.Cvtepi32Ps(p4)
/* Convert to planar */
g1g0r1r0 := sse.UnpackloPs(p1f, p2f)
b1b0 := sse.UnpackhiPs(p1f, p2f)
g3g2r3r2 := sse.UnpackloPs(p3f, p4f)
b3b2 := sse.UnpackhiPs(p3f, p4f)
r := sse.MovelhPs(g1g0r1r0, g3g2r3r2)
g := sse.MovehlPs(g3g2r3r2, g1g0r1r0)
b := sse.MovelhPs(b1b0, b3b2)
/* Apply matrix to convert to sRGB */
r = sseMatrix3Mul(matPs[0:3], r, g, b)
g = sseMatrix3Mul(matPs[3:6], r, g, b)
b = sseMatrix3Mul(matPs[6:9], r, g, b)
/* Normalize to 0->1 and clamp */
normalize := sse.Set1Ps(1.0 / 65535.0)
max_val := sse.Set1Ps(1.0)
min_val := sse.SetzeroPs()
r = sse.MinPs(max_val, sse.MaxPs(min_val, sse.MulPs(normalize, r)))
g = sse.MinPs(max_val, sse.MaxPs(min_val, sse.MulPs(normalize, g)))
b = sse.MinPs(max_val, sse.MaxPs(min_val, sse.MulPs(normalize, b)))
/* Apply Gamma */
/* Calculate values to be used if larger than junction point */
mul_over := sse.Set1Ps(1.055)
sub_over := sse.Set1Ps(0.055)
pow_over := sse.Set1Ps(1.0 / 2.4)
r_gam := sse.SubPs(sse.MulPs(mul_over, FastPowPs(r, pow_over)), sub_over)
g_gam := sse.SubPs(sse.MulPs(mul_over, FastPowPs(g, pow_over)), sub_over)
b_gam := sse.SubPs(sse.MulPs(mul_over, FastPowPs(b, pow_over)), sub_over)
/* Create mask for values smaller than junction point */
junction := sse.Set1Ps(0.0031308)
mask_r := sse.CmpltPs(r, junction)
mask_g := sse.CmpltPs(g, junction)
mask_b := sse.CmpltPs(b, junction)
/* Calculate value to be used if under junction */
mul_under := sse.Set1Ps(12.92)
r_mul := sse.AndPs(mask_r, sse.MulPs(mul_under, r))
g_mul := sse.AndPs(mask_g, sse.MulPs(mul_under, g))
b_mul := sse.AndPs(mask_b, sse.MulPs(mul_under, b))
/* Select the value to be used based on the junction mask and scale to 8 bit */
upscale := sse.Set1Ps(255.5)
r = sse.MulPs(upscale, sse.OrPs(r_mul, sse.AndnotPs(mask_r, r_gam)))
g = sse.MulPs(upscale, sse.OrPs(g_mul, sse.AndnotPs(mask_g, g_gam)))
b = sse.MulPs(upscale, sse.OrPs(b_mul, sse.AndnotPs(mask_b, b_gam)))
/* Convert to 8 bit unsigned and interleave*/
r_i := sse2.CvtpsEpi32(r)
g_i := sse2.CvtpsEpi32(g)
b_i := sse2.CvtpsEpi32(b)
r_i = sse2.PacksEpi32(r_i, r_i)
g_i = sse2.PacksEpi32(g_i, g_i)
b_i = sse2.PacksEpi32(b_i, b_i)
/* Set alpha value to 255 and store */
alpha_mask := sse2.Set1Epi32(0xff000000)
rg_i := sse2.UnpackloEpi16(r_i, g_i)
bb_i := sse2.UnpackloEpi16(b_i, b_i)
p1 = sse2.UnpackloEpi32(rg_i, bb_i)
p2 = sse2.UnpackhiEpi32(rg_i, bb_i)
p1 = sse2.OrSi128(alpha_mask, sse2.PackusEpi16(p1, p2))
o[x] = p1
}
}
}