// Copyright (c) 2013 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include <algorithm>
#include <cmath>
#include <vector>
#include "base/logging.h"
#include "skia/ext/recursive_gaussian_convolution.h"
namespace skia {
namespace {
// Takes the value produced by accumulating element-wise product of image with
// a kernel and brings it back into range.
// All of the filter scaling factors are in fixed point with kShiftBits bits of
// fractional part.
template<bool take_absolute>
inline unsigned char FloatTo8(float f) {
int a = static_cast<int>(f + 0.5f);
if (take_absolute)
a = std::abs(a);
else if (a < 0)
return 0;
if (a < 256)
return a;
return 255;
}
template<RecursiveFilter::Order order>
inline float ForwardFilter(float in_n_1,
float in_n,
float in_n1,
const std::vector<float>& w,
int n,
const float* b) {
switch (order) {
case RecursiveFilter::FUNCTION:
return b[0] * in_n + b[1] * w[n-1] + b[2] * w[n-2] + b[3] * w[n-3];
case RecursiveFilter::FIRST_DERIVATIVE:
return b[0] * 0.5f * (in_n1 - in_n_1) +
b[1] * w[n-1] + b[2] * w[n-2] + b[3] * w[n-3];
case RecursiveFilter::SECOND_DERIVATIVE:
return b[0] * (in_n - in_n_1) +
b[1] * w[n-1] + b[2] * w[n-2] + b[3] * w[n-3];
}
NOTREACHED();
return 0.0f;
}
template<RecursiveFilter::Order order>
inline float BackwardFilter(const std::vector<float>& out,
int n,
float w_n,
float w_n1,
const float* b) {
switch (order) {
case RecursiveFilter::FUNCTION:
case RecursiveFilter::FIRST_DERIVATIVE:
return b[0] * w_n +
b[1] * out[n + 1] + b[2] * out[n + 2] + b[3] * out[n + 3];
case RecursiveFilter::SECOND_DERIVATIVE:
return b[0] * (w_n1 - w_n) +
b[1] * out[n + 1] + b[2] * out[n + 2] + b[3] * out[n + 3];
}
NOTREACHED();
return 0.0f;
}
template<RecursiveFilter::Order order, bool absolute_values>
unsigned char SingleChannelRecursiveFilter(
const unsigned char* const source_data,
int source_pixel_stride,
int source_row_stride,
int row_width,
int row_count,
unsigned char* const output,
int output_pixel_stride,
int output_row_stride,
const float* b) {
const int intermediate_buffer_size = row_width + 6;
std::vector<float> w(intermediate_buffer_size);
const unsigned char* in = source_data;
unsigned char* out = output;
unsigned char max_output = 0;
for (int r = 0; r < row_count;
++r, in += source_row_stride, out += output_row_stride) {
// Compute forward filter.
// First initialize start of the w (temporary) vector.
if (order == RecursiveFilter::FUNCTION)
w[0] = w[1] = w[2] = in[0];
else
w[0] = w[1] = w[2] = 0.0f;
// Note that special-casing of w[3] is needed because of derivatives.
w[3] = ForwardFilter<order>(
in[0], in[0], in[source_pixel_stride], w, 3, b);
int n = 4;
int c = 1;
int byte_index = source_pixel_stride;
for (; c < row_width - 1; ++c, ++n, byte_index += source_pixel_stride) {
w[n] = ForwardFilter<order>(in[byte_index - source_pixel_stride],
in[byte_index],
in[byte_index + source_pixel_stride],
w, n, b);
}
// The value of w corresponding to the last image pixel needs to be computed
// separately, again because of derivatives.
w[n] = ForwardFilter<order>(in[byte_index - source_pixel_stride],
in[byte_index],
in[byte_index],
w, n, b);
// Now three trailing bytes set to the same value as current w[n].
w[n + 1] = w[n];
w[n + 2] = w[n];
w[n + 3] = w[n];
// Now apply the back filter.
float w_n1 = w[n + 1];
int output_index = (row_width - 1) * output_pixel_stride;
for (; c >= 0; output_index -= output_pixel_stride, --c, --n) {
float w_n = BackwardFilter<order>(w, n, w[n], w_n1, b);
w_n1 = w[n];
w[n] = w_n;
out[output_index] = FloatTo8<absolute_values>(w_n);
max_output = std::max(max_output, out[output_index]);
}
}
return max_output;
}
unsigned char SingleChannelRecursiveFilter(
const unsigned char* const source_data,
int source_pixel_stride,
int source_row_stride,
int row_width,
int row_count,
unsigned char* const output,
int output_pixel_stride,
int output_row_stride,
const float* b,
RecursiveFilter::Order order,
bool absolute_values) {
if (absolute_values) {
switch (order) {
case RecursiveFilter::FUNCTION:
return SingleChannelRecursiveFilter<RecursiveFilter::FUNCTION, true>(
source_data, source_pixel_stride, source_row_stride,
row_width, row_count,
output, output_pixel_stride, output_row_stride, b);
case RecursiveFilter::FIRST_DERIVATIVE:
return SingleChannelRecursiveFilter<
RecursiveFilter::FIRST_DERIVATIVE, true>(
source_data, source_pixel_stride, source_row_stride,
row_width, row_count,
output, output_pixel_stride, output_row_stride, b);
case RecursiveFilter::SECOND_DERIVATIVE:
return SingleChannelRecursiveFilter<
RecursiveFilter::SECOND_DERIVATIVE, true>(
source_data, source_pixel_stride, source_row_stride,
row_width, row_count,
output, output_pixel_stride, output_row_stride, b);
}
} else {
switch (order) {
case RecursiveFilter::FUNCTION:
return SingleChannelRecursiveFilter<RecursiveFilter::FUNCTION, false>(
source_data, source_pixel_stride, source_row_stride,
row_width, row_count,
output, output_pixel_stride, output_row_stride, b);
case RecursiveFilter::FIRST_DERIVATIVE:
return SingleChannelRecursiveFilter<
RecursiveFilter::FIRST_DERIVATIVE, false>(
source_data, source_pixel_stride, source_row_stride,
row_width, row_count,
output, output_pixel_stride, output_row_stride, b);
case RecursiveFilter::SECOND_DERIVATIVE:
return SingleChannelRecursiveFilter<
RecursiveFilter::SECOND_DERIVATIVE, false>(
source_data, source_pixel_stride, source_row_stride,
row_width, row_count,
output, output_pixel_stride, output_row_stride, b);
}
}
NOTREACHED();
return 0;
}
}
float RecursiveFilter::qFromSigma(float sigma) {
DCHECK_GE(sigma, 0.5f);
if (sigma <= 2.5f)
return 3.97156f - 4.14554f * std::sqrt(1.0f - 0.26891f * sigma);
return 0.98711f * sigma - 0.96330f;
}
void RecursiveFilter::computeCoefficients(float q, float b[4]) {
b[0] = 1.57825f + 2.44413f * q + 1.4281f * q * q + 0.422205f * q * q * q;
b[1] = 2.4413f * q + 2.85619f * q * q + 1.26661f * q * q * q;
b[2] = - 1.4281f * q * q - 1.26661f * q * q * q;
b[3] = 0.422205f * q * q * q;
// The above is exactly like in the paper. To cut down on computations,
// we can fix up these numbers a bit now.
float b_norm = 1.0f - (b[1] + b[2] + b[3]) / b[0];
b[1] /= b[0];
b[2] /= b[0];
b[3] /= b[0];
b[0] = b_norm;
}
RecursiveFilter::RecursiveFilter(float sigma, Order order)
: order_(order), q_(qFromSigma(sigma)) {
computeCoefficients(q_, b_);
}
unsigned char SingleChannelRecursiveGaussianX(const unsigned char* source_data,
int source_byte_row_stride,
int input_channel_index,
int input_channel_count,
const RecursiveFilter& filter,
const SkISize& image_size,
unsigned char* output,
int output_byte_row_stride,
int output_channel_index,
int output_channel_count,
bool absolute_values) {
return SingleChannelRecursiveFilter(source_data + input_channel_index,
input_channel_count,
source_byte_row_stride,
image_size.width(),
image_size.height(),
output + output_channel_index,
output_channel_count,
output_byte_row_stride,
filter.b(),
filter.order(),
absolute_values);
}
unsigned char SingleChannelRecursiveGaussianY(const unsigned char* source_data,
int source_byte_row_stride,
int input_channel_index,
int input_channel_count,
const RecursiveFilter& filter,
const SkISize& image_size,
unsigned char* output,
int output_byte_row_stride,
int output_channel_index,
int output_channel_count,
bool absolute_values) {
return SingleChannelRecursiveFilter(source_data + input_channel_index,
source_byte_row_stride,
input_channel_count,
image_size.height(),
image_size.width(),
output + output_channel_index,
output_byte_row_stride,
output_channel_count,
filter.b(),
filter.order(),
absolute_values);
}
} // namespace skia