You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
466 lines
13 KiB
C++
466 lines
13 KiB
C++
/*
|
|
Copyright 2007 nVidia, Inc.
|
|
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
|
|
|
|
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
See the License for the specific language governing permissions and limitations under the License.
|
|
*/
|
|
|
|
// Utility and common routines
|
|
|
|
#include "utils.h"
|
|
#include "nvmath/Vector.inl"
|
|
#include <math.h>
|
|
|
|
using namespace nv;
|
|
|
|
static int denom7_weights_64[] = {0, 9, 18, 27, 37, 46, 55, 64}; // divided by 64
|
|
static int denom15_weights_64[] = {0, 4, 9, 13, 17, 21, 26, 30, 34, 38, 43, 47, 51, 55, 60, 64}; // divided by 64
|
|
|
|
/*static*/ Format Utils::FORMAT;
|
|
|
|
int Utils::lerp(int a, int b, int i, int denom)
|
|
{
|
|
nvDebugCheck (denom == 3 || denom == 7 || denom == 15);
|
|
nvDebugCheck (i >= 0 && i <= denom);
|
|
|
|
int round = 32, shift = 6, *weights;
|
|
|
|
switch(denom)
|
|
{
|
|
case 3: denom *= 5; i *= 5; // fall through to case 15
|
|
case 15: weights = denom15_weights_64; break;
|
|
case 7: weights = denom7_weights_64; break;
|
|
default: nvDebugCheck(0);
|
|
}
|
|
|
|
return (a*weights[denom-i] +b*weights[i] + round) >> shift;
|
|
}
|
|
|
|
Vector3 Utils::lerp(const Vector3& a, const Vector3 &b, int i, int denom)
|
|
{
|
|
nvDebugCheck (denom == 3 || denom == 7 || denom == 15);
|
|
nvDebugCheck (i >= 0 && i <= denom);
|
|
|
|
int shift = 6, *weights;
|
|
|
|
switch(denom)
|
|
{
|
|
case 3: denom *= 5; i *= 5; // fall through to case 15
|
|
case 15: weights = denom15_weights_64; break;
|
|
case 7: weights = denom7_weights_64; break;
|
|
default: nvAssume(0);
|
|
}
|
|
|
|
// no need to round these as this is an exact division
|
|
return (a*weights[denom-i] +b*weights[i]) / float(1 << shift);
|
|
}
|
|
|
|
|
|
/*
|
|
For unsigned f16, clamp the input to [0,F16MAX]. Thus u15.
|
|
For signed f16, clamp the input to [-F16MAX,F16MAX]. Thus s16.
|
|
|
|
The conversions proceed as follows:
|
|
|
|
unsigned f16: get bits. if high bit set, clamp to 0, else clamp to F16MAX.
|
|
signed f16: get bits. extract exp+mantissa and clamp to F16MAX. return -value if sign bit was set, else value
|
|
unsigned int: get bits. return as a positive value.
|
|
signed int. get bits. return as a value in -32768..32767.
|
|
|
|
The inverse conversions are just the inverse of the above.
|
|
*/
|
|
|
|
// clamp the 3 channels of the input vector to the allowable range based on FORMAT
|
|
// note that each channel is a float storing the allowable range as a bit pattern converted to float
|
|
// that is, for unsigned f16 say, we would clamp each channel to the range [0, F16MAX]
|
|
|
|
void Utils::clamp(Vector3 &v)
|
|
{
|
|
for (int i=0; i<3; ++i)
|
|
{
|
|
switch(Utils::FORMAT)
|
|
{
|
|
case UNSIGNED_F16:
|
|
if (v.component[i] < 0.0) v.component[i] = 0;
|
|
else if (v.component[i] > F16MAX) v.component[i] = F16MAX;
|
|
break;
|
|
|
|
case SIGNED_F16:
|
|
if (v.component[i] < -F16MAX) v.component[i] = -F16MAX;
|
|
else if (v.component[i] > F16MAX) v.component[i] = F16MAX;
|
|
break;
|
|
|
|
default:
|
|
nvAssume (0);
|
|
}
|
|
}
|
|
}
|
|
|
|
// convert a u16 value to s17 (represented as an int) based on the format expected
|
|
int Utils::ushort_to_format(unsigned short input)
|
|
{
|
|
int out, s;
|
|
|
|
// clamp to the valid range we are expecting
|
|
switch (Utils::FORMAT)
|
|
{
|
|
case UNSIGNED_F16:
|
|
if (input & F16S_MASK) out = 0;
|
|
else if (input > F16MAX) out = F16MAX;
|
|
else out = input;
|
|
break;
|
|
|
|
case SIGNED_F16:
|
|
s = input & F16S_MASK;
|
|
input &= F16EM_MASK;
|
|
if (input > F16MAX) out = F16MAX;
|
|
else out = input;
|
|
out = s ? -out : out;
|
|
break;
|
|
}
|
|
return out;
|
|
}
|
|
|
|
// convert a s17 value to u16 based on the format expected
|
|
unsigned short Utils::format_to_ushort(int input)
|
|
{
|
|
unsigned short out;
|
|
|
|
// clamp to the valid range we are expecting
|
|
switch (Utils::FORMAT)
|
|
{
|
|
case UNSIGNED_F16:
|
|
nvDebugCheck (input >= 0 && input <= F16MAX);
|
|
out = input;
|
|
break;
|
|
|
|
case SIGNED_F16:
|
|
nvDebugCheck (input >= -F16MAX && input <= F16MAX);
|
|
// convert to sign-magnitude
|
|
int s;
|
|
if (input < 0) { s = F16S_MASK; input = -input; }
|
|
else { s = 0; }
|
|
out = s | input;
|
|
break;
|
|
}
|
|
return out;
|
|
}
|
|
|
|
// quantize the input range into equal-sized bins
|
|
int Utils::quantize(float value, int prec)
|
|
{
|
|
int q, ivalue, s;
|
|
|
|
nvDebugCheck (prec > 1); // didn't bother to make it work for 1
|
|
|
|
value = (float)floor(value + 0.5);
|
|
|
|
int bias = (prec > 10) ? ((1<<(prec-1))-1) : 0; // bias precisions 11..16 to get a more accurate quantization
|
|
|
|
switch (Utils::FORMAT)
|
|
{
|
|
case UNSIGNED_F16:
|
|
nvDebugCheck (value >= 0 && value <= F16MAX);
|
|
ivalue = (int)value;
|
|
q = ((ivalue << prec) + bias) / (F16MAX+1);
|
|
nvDebugCheck (q >= 0 && q < (1 << prec));
|
|
break;
|
|
|
|
case SIGNED_F16:
|
|
nvDebugCheck (value >= -F16MAX && value <= F16MAX);
|
|
// convert to sign-magnitude
|
|
ivalue = (int)value;
|
|
if (ivalue < 0) { s = 1; ivalue = -ivalue; } else s = 0;
|
|
|
|
q = ((ivalue << (prec-1)) + bias) / (F16MAX+1);
|
|
if (s)
|
|
q = -q;
|
|
nvDebugCheck (q > -(1 << (prec-1)) && q < (1 << (prec-1)));
|
|
break;
|
|
}
|
|
|
|
return q;
|
|
}
|
|
|
|
int Utils::finish_unquantize(int q, int prec)
|
|
{
|
|
if (Utils::FORMAT == UNSIGNED_F16)
|
|
return (q * 31) >> 6; // scale the magnitude by 31/64
|
|
else if (Utils::FORMAT == SIGNED_F16)
|
|
return (q < 0) ? -(((-q) * 31) >> 5) : (q * 31) >> 5; // scale the magnitude by 31/32
|
|
else
|
|
return q;
|
|
}
|
|
|
|
// unquantize each bin to midpoint of original bin range, except
|
|
// for the end bins which we push to an endpoint of the bin range.
|
|
// we do this to ensure we can represent all possible original values.
|
|
// the asymmetric end bins do not affect PSNR for the test images.
|
|
//
|
|
// code this function assuming an arbitrary bit pattern as the encoded block
|
|
int Utils::unquantize(int q, int prec)
|
|
{
|
|
int unq, s;
|
|
|
|
nvDebugCheck (prec > 1); // not implemented for prec 1
|
|
|
|
switch (Utils::FORMAT)
|
|
{
|
|
// modify this case to move the multiplication by 31 after interpolation.
|
|
// Need to use finish_unquantize.
|
|
|
|
// since we have 16 bits available, let's unquantize this to 16 bits unsigned
|
|
// thus the scale factor is [0-7c00)/[0-10000) = 31/64
|
|
case UNSIGNED_F16:
|
|
if (prec >= 15)
|
|
unq = q;
|
|
else if (q == 0)
|
|
unq = 0;
|
|
else if (q == ((1<<prec)-1))
|
|
unq = U16MAX;
|
|
else
|
|
unq = (q * (U16MAX+1) + (U16MAX+1)/2) >> prec;
|
|
break;
|
|
|
|
// here, let's stick with S16 (no apparent quality benefit from going to S17)
|
|
// range is (-7c00..7c00)/(-8000..8000) = 31/32
|
|
case SIGNED_F16:
|
|
// don't remove this test even though it appears equivalent to the code below
|
|
// as it isn't -- the code below can overflow for prec = 16
|
|
if (prec >= 16)
|
|
unq = q;
|
|
else
|
|
{
|
|
if (q < 0) { s = 1; q = -q; } else s = 0;
|
|
|
|
if (q == 0)
|
|
unq = 0;
|
|
else if (q >= ((1<<(prec-1))-1))
|
|
unq = s ? -S16MAX : S16MAX;
|
|
else
|
|
{
|
|
unq = (q * (S16MAX+1) + (S16MAX+1)/2) >> (prec-1);
|
|
if (s)
|
|
unq = -unq;
|
|
}
|
|
}
|
|
break;
|
|
}
|
|
return unq;
|
|
}
|
|
|
|
static int clamp(double r, double low, double high)
|
|
{
|
|
if (r < low) return low;
|
|
else if (r > high) return high;
|
|
else return r;
|
|
}
|
|
|
|
|
|
// match the tonemapping function used by exrdisplay
|
|
/*static void tonemap(const Vector3 &in, double exposure, Vector3 &out)
|
|
{
|
|
double r,g,b;
|
|
unsigned short h;
|
|
|
|
// convert from bit pattern back to half and then to double
|
|
h = in.x; r = h;
|
|
h = in.y; g = h;
|
|
h = in.z; b = h;
|
|
|
|
// 1) Compensate for fogging by subtracting defog
|
|
// from the raw pixel values.
|
|
// Response: We work with defog of 0.0, so this is a no-op
|
|
|
|
// 2) Multiply the defogged pixel values by
|
|
// 2^(exposure + 2.47393).
|
|
double exposure_scale = pow(2.0, exposure + 2.47393);
|
|
r *= exposure_scale;
|
|
g *= exposure_scale;
|
|
b *= exposure_scale;
|
|
|
|
// 3) Values, which are now 1.0, are called "middle gray".
|
|
// If defog and exposure are both set to 0.0, then
|
|
// middle gray corresponds to a raw pixel value of 0.18.
|
|
// In step 6, middle gray values will be mapped to an
|
|
// intensity 3.5 f-stops below the display's maximum
|
|
// intensity.
|
|
// Response: no apparent content.
|
|
|
|
// 4) Apply a knee function. The knee function has two
|
|
// parameters, kneeLow and kneeHigh. Pixel values
|
|
// below 2^kneeLow are not changed by the knee
|
|
// function. Pixel values above kneeLow are lowered
|
|
// according to a logarithmic curve, such that the
|
|
// value 2^kneeHigh is mapped to 2^3.5 (in step 6,
|
|
// this value will be mapped to the the display's
|
|
// maximum intensity).
|
|
// Response: kneeLow = 0.0 (2^0.0 => 1); kneeHigh = 5.0 (2^5 =>32)
|
|
if (r > 1.0)
|
|
r = 1.0 + log ((r-1.0) * 0.184874 + 1) / 0.184874;
|
|
if (g > 1.0)
|
|
g = 1.0 + log ((g-1.0) * 0.184874 + 1) / 0.184874;
|
|
if (b > 1.0)
|
|
b = 1.0 + log ((b-1.0) * 0.184874 + 1) / 0.184874;
|
|
//
|
|
// 5) Gamma-correct the pixel values, assuming that the
|
|
// screen's gamma is 0.4545 (or 1/2.2).
|
|
r = pow (r, 0.4545);
|
|
g = pow (g, 0.4545);
|
|
b = pow (b, 0.4545);
|
|
|
|
// 6) Scale the values such that pixels middle gray
|
|
// pixels are mapped to 84.66 (or 3.5 f-stops below
|
|
// the display's maximum intensity).
|
|
//
|
|
// 7) Clamp the values to [0, 255].
|
|
r *= 84.66f;
|
|
g *= 84.66f;
|
|
b *= 84.66f;
|
|
|
|
out.x = clamp (r, 0, 255);
|
|
out.y = clamp (g, 0, 255);
|
|
out.z = clamp (b, 0, 255);
|
|
}*/
|
|
|
|
static void mpsnrmap(const Vector3 &in, int exposure, Vector3 &out)
|
|
{
|
|
double r,g,b;
|
|
uint16 h;
|
|
|
|
// convert from bit pattern back to half and then to double
|
|
h = in.x; r = h;
|
|
h = in.y; g = h;
|
|
h = in.z; b = h;
|
|
|
|
nvDebugCheck (exposure > -32 && exposure < 32);
|
|
if (exposure > 0)
|
|
{
|
|
r *= 1 << exposure;
|
|
g *= 1 << exposure;
|
|
b *= 1 << exposure;
|
|
}
|
|
else if (exposure < 0)
|
|
{
|
|
exposure = -exposure;
|
|
r /= 1 << exposure;
|
|
g /= 1 << exposure;
|
|
b /= 1 << exposure;
|
|
}
|
|
r = 255 * pow (r, 0.4545);
|
|
g = 255 * pow (g, 0.4545);
|
|
b = 255 * pow (b, 0.4545);
|
|
|
|
out.x = clamp (r, 0, 255);
|
|
out.y = clamp (g, 0, 255);
|
|
out.z = clamp (b, 0, 255);
|
|
}
|
|
|
|
// pick a norm!
|
|
#define NORM_EUCLIDEAN 1
|
|
|
|
double Utils::norm(const Vector3 &a, const Vector3 &b)
|
|
{
|
|
#ifdef NORM_EUCLIDEAN
|
|
return lengthSquared(a - b);
|
|
#endif
|
|
#ifdef NORM_ABS
|
|
Vector3 err = a - b;
|
|
return fabs(err.x) + fabs(err.y) + fabs(err.z);
|
|
#endif
|
|
#ifdef NORM_EUCLIDEAN_EXPOSURE_UNWEIGHED
|
|
double toterr = 0;
|
|
Vector3 mapa, mapb, err;
|
|
for (int i=-6; i <= 6; i += 3) // figure how many exposure samples needed. I'd argue if you take too many it's same as euclidean
|
|
{
|
|
tonemap(a, i, mapa);
|
|
tonemap(b, i, mapb);
|
|
err = mapa - mapb;
|
|
toterr += err * err;
|
|
}
|
|
return toterr;
|
|
#endif
|
|
#ifdef NORM_EUCLIDEAN_EXPOSURE_WEIGHED
|
|
double toterr = 0;
|
|
Vector3 mapa, mapb, err;
|
|
double rwt = 0.299;
|
|
double gwt = 0.587;
|
|
double bwt = 0.114;
|
|
for (int i=-6; i <= 6; i += 3) // figure how many exposure samples needed. I'd argue if you take too many it's same as euclidean
|
|
{
|
|
tonemap(a, i, mapa);
|
|
tonemap(b, i, mapb);
|
|
mapa.x *= rwt; mapa.y *= gwt; mapa.z *= bwt;
|
|
mapb.x *= rwt; mapb.y *= gwt; mapb.z *= bwt;
|
|
err = mapa - mapb;
|
|
toterr += err * err;
|
|
}
|
|
return toterr;
|
|
#endif
|
|
}
|
|
|
|
double Utils::mpsnr_norm(const Vector3 &a, int exposure, const Vector3 &b)
|
|
{
|
|
Vector3 mapa, mapb;
|
|
|
|
mpsnrmap(a, exposure, mapa);
|
|
mpsnrmap(b, exposure, mapb);
|
|
|
|
return lengthSquared(mapa - mapb);
|
|
}
|
|
|
|
// parse <name>[<start>{:<end>}]{,}
|
|
// the pointer starts here ^
|
|
// name is 1 or 2 chars and matches field names. start and end are decimal numbers
|
|
void Utils::parse(const char *encoding, int &ptr, Field &field, int &endbit, int &len)
|
|
{
|
|
if (ptr <= 0) return;
|
|
--ptr;
|
|
if (encoding[ptr] == ',') --ptr;
|
|
nvDebugCheck (encoding[ptr] == ']');
|
|
--ptr;
|
|
endbit = 0;
|
|
int scale = 1;
|
|
while (encoding[ptr] != ':' && encoding[ptr] != '[')
|
|
{
|
|
nvDebugCheck(encoding[ptr] >= '0' && encoding[ptr] <= '9');
|
|
endbit += (encoding[ptr--] - '0') * scale;
|
|
scale *= 10;
|
|
}
|
|
int startbit = 0; scale = 1;
|
|
if (encoding[ptr] == '[')
|
|
startbit = endbit;
|
|
else
|
|
{
|
|
ptr--;
|
|
while (encoding[ptr] != '[')
|
|
{
|
|
nvDebugCheck(encoding[ptr] >= '0' && encoding[ptr] <= '9');
|
|
startbit += (encoding[ptr--] - '0') * scale;
|
|
scale *= 10;
|
|
}
|
|
}
|
|
len = startbit - endbit + 1; // startbit>=endbit note
|
|
--ptr;
|
|
if (encoding[ptr] == 'm') field = FIELD_M;
|
|
else if (encoding[ptr] == 'd') field = FIELD_D;
|
|
else {
|
|
// it's wxyz
|
|
nvDebugCheck (encoding[ptr] >= 'w' && encoding[ptr] <= 'z');
|
|
int foo = encoding[ptr--] - 'w';
|
|
// now it is r g or b
|
|
if (encoding[ptr] == 'r') foo += 10;
|
|
else if (encoding[ptr] == 'g') foo += 20;
|
|
else if (encoding[ptr] == 'b') foo += 30;
|
|
else nvDebugCheck(0);
|
|
field = (Field) foo;
|
|
}
|
|
}
|
|
|
|
|