func log2f4(x x86.M128) x86.M128 {
exp := sse2.Set1Epi32(exp_mask)
mant := sse2.Set1Epi32(mantissa_mask)
one := sse.Set1Ps(1.0)
i := sse2.CastpsSi128(x)
e := sse2.Cvtepi32Ps(sse2.SubEpi32(sse2.SrliEpi32(sse2.AndSi128(i, exp), 23), sse2.Set1Epi32(127)))
m := sse.OrPs(sse2.Castsi128Ps(sse2.AndSi128(i, mant)), one)
var p x86.M128
/* Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[ */
if LOG_poly_DEGREE == 6 {
p = poly5(m, log_p5_0, log_p5_1, log_p5_2, log_p5_3, log_p5_4, log_p5_5)
} else if LOG_poly_DEGREE == 5 {
p = poly4(m, log_p4_0, log_p4_1, log_p4_2, log_p4_3, log_p4_4)
} else if LOG_poly_DEGREE == 4 {
p = poly3(m, log_p3_0, log_p3_1, log_p3_2, log_p3_3)
} else if LOG_poly_DEGREE == 3 {
p = poly2(m, log_p2_0, log_p2_1, log_p2_2)
} else {
panic("unsupported poly degree")
}
/* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
p = sse.MulPs(p, sse.SubPs(m, one))
return sse.AddPs(p, e)
}
// vector of 4 32-bit unsigned integers
func (u Uint32Div) DivSSE4(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
t7 := sse4.BlendvEpi8(t2, t4, t5) // blend two results
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
}
// 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
}
// Try some complex intrinsics (SSE2)
// Doesn't test any values (since intrisics are unimplemented)
// Converted from https://github.com/klauspost/rawspeed/blob/develop/RawSpeed/RawImageDataU16.cpp#L152
func TestComplex(t *testing.T) {
full_scale_fp := 1000
half_scale_fp := 500
mDitherScale := true
var sub_mul [4]x86.M128i
var rand_mul x86.M128i
sseround := sse2.SetEpi32(512, 512, 512, 512)
ssesub2 := sse2.SetEpi32(32768, 32768, 32768, 32768)
ssesign := sse2.SetEpi32(0x80008000, 0x80008000, 0x80008000, 0x80008000)
sse_full_scale_fp := sse2.Set1Epi32(full_scale_fp | (full_scale_fp << 16))
sse_half_scale_fp := sse2.Set1Epi32(half_scale_fp >> 4)
if mDitherScale {
rand_mul = sse2.Set1Epi32(0x4d9f1d32)
} else {
rand_mul = sse2.SetzeroSi128()
}
rand_mask := sse2.Set1Epi32(0x00ff00ff) // 8 random bits
width := 1024
height := 1024
// Emulate 1024 x 1024 x 16bpp
input := make([]byte, 1024*1024*2)
for y := 0; y < height; y++ {
// Convert current line to []M128i
line := x86.BytesToM128i(input[y*width*2 : y*width*2+width*2])
var sserandom x86.M128i
if mDitherScale {
sserandom = sse2.SetEpi32(width*1676+y*18000, width*2342+y*34311, width*4272+y*12123, width*1234+y*23464)
} else {
sserandom = sse2.SetzeroSi128()
}
var ssescale, ssesub x86.M128i
if (y & 1) == 0 {
ssesub = sub_mul[0]
ssescale = sub_mul[1]
} else {
ssesub = sub_mul[2]
ssescale = sub_mul[3]
}
for x, pix_low := range line {
// Subtract black
pix_low = sse2.SubsEpu16(pix_low, ssesub)
// Multiply the two unsigned shorts and combine it to 32 bit result
pix_high := sse2.MulhiEpu16(pix_low, ssescale)
temp := sse2.MulloEpi16(pix_low, ssescale)
pix_low = sse2.UnpackloEpi16(temp, pix_high)
pix_high = sse2.UnpackhiEpi16(temp, pix_high)
// Add rounder
pix_low = sse2.AddEpi32(pix_low, sseround)
pix_high = sse2.AddEpi32(pix_high, sseround)
sserandom = sse2.XorSi128(sse2.MulhiEpi16(sserandom, rand_mul), sse2.MulloEpi16(sserandom, rand_mul))
rand_masked := sse2.AndSi128(sserandom, rand_mask) // Get 8 random bits
rand_masked = sse2.MulloEpi16(rand_masked, sse_full_scale_fp)
zero := sse2.SetzeroSi128()
rand_lo := sse2.SubEpi32(sse_half_scale_fp, sse2.UnpackloEpi16(rand_masked, zero))
rand_hi := sse2.SubEpi32(sse_half_scale_fp, sse2.UnpackhiEpi16(rand_masked, zero))
pix_low = sse2.AddEpi32(pix_low, rand_lo)
pix_high = sse2.AddEpi32(pix_high, rand_hi)
// Shift down
pix_low = sse2.SraiEpi32(pix_low, 10)
pix_high = sse2.SraiEpi32(pix_high, 10)
// Subtract to avoid clipping
pix_low = sse2.SubEpi32(pix_low, ssesub2)
pix_high = sse2.SubEpi32(pix_high, ssesub2)
// Pack
pix_low = sse2.PacksEpi32(pix_low, pix_high)
// Shift sign off
pix_low = sse2.XorSi128(pix_low, ssesign)
line[x] = pix_low
}
}
}