#include <epoxy/gl.h>
#include <math.h>
+#include "effect_chain.h"
#include "effect_util.h"
+#include "fp16.h"
#include "fft_pass_effect.h"
#include "util.h"
int input_size = (direction == VERTICAL) ? input_height : input_width;
- // See the comments on changes_output_size() in the .h file to see
- // why this is legal. It is _needed_ because it counteracts the
- // precision issues we get because we sample the input texture with
- // normalized coordinates (especially when the repeat count along
- // the axis is not a power of two); we very rapidly end up in narrowly
- // missing a texel center, which causes precision loss to propagate
- // throughout the FFT.
- assert(*sampler_num == 1);
- glActiveTexture(GL_TEXTURE0);
+ // This is needed because it counteracts the precision issues we get
+ // because we sample the input texture with normalized coordinates
+ // (especially when the repeat count along the axis is not a power of
+ // two); we very rapidly end up in narrowly missing a texel center,
+ // which causes precision loss to propagate throughout the FFT.
+ Node *self = chain->find_node_for_effect(this);
+ glActiveTexture(chain->get_input_sampler(self, 0));
check_error();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
check_error();
// bit, so the stride is 8, and so on.
assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
- float *tmp = new float[fft_size * 4];
+ fp16_int_t *tmp = new fp16_int_t[fft_size * 4];
int subfft_size = 1 << pass_number;
double mulfac;
if (inverse) {
// is that we can have multiple FFTs along the same line,
// and want to reuse the support texture by repeating it.
int base = k * stride * 2 + offset;
- int support_texture_index;
+ int support_texture_index = i;
+ int src1 = base;
+ int src2 = base + stride;
if (direction == FFTPassEffect::VERTICAL) {
// Compensate for OpenGL's bottom-left convention.
- support_texture_index = fft_size - i - 1;
- } else {
- support_texture_index = i;
+ support_texture_index = fft_size - support_texture_index - 1;
+ src1 = fft_size - src1 - 1;
+ src2 = fft_size - src2 - 1;
}
- tmp[support_texture_index * 4 + 0] = (base - support_texture_index) / double(input_size);
- tmp[support_texture_index * 4 + 1] = (base + stride - support_texture_index) / double(input_size);
- tmp[support_texture_index * 4 + 2] = twiddle_real;
- tmp[support_texture_index * 4 + 3] = twiddle_imag;
+ tmp[support_texture_index * 4 + 0] = fp64_to_fp16((src1 - support_texture_index) / double(input_size));
+ tmp[support_texture_index * 4 + 1] = fp64_to_fp16((src2 - support_texture_index) / double(input_size));
+ tmp[support_texture_index * 4 + 2] = fp64_to_fp16(twiddle_real);
+ tmp[support_texture_index * 4 + 3] = fp64_to_fp16(twiddle_imag);
}
glActiveTexture(GL_TEXTURE0 + *sampler_num);
// Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,
// at least according to a paper by Schatzman (see gpuwave.pdf reference [30]
- // for the full reference), so we keep them at 32-bit. However, for
- // small sizes, all components are exact anyway, so we can cheat there
- // (although noting that the source coordinates become somewhat less
- // accurate then, too).
- glTexImage2D(GL_TEXTURE_2D, 0, (subfft_size <= 4) ? GL_RGBA16F : GL_RGBA32F, fft_size, 1, 0, GL_RGBA, GL_FLOAT, tmp);
+ // for the full reference); however, practical testing indicates that it's
+ // not a problem to keep the twiddle factors at 16-bit, at least as long as
+ // we round them properly--it would seem that Schatzman were mainly talking
+ // about poor sin()/cos() approximations. Thus, we store them in 16-bit,
+ // which gives a nice speed boost.
+ //
+ // Note that the source coordinates become somewhat less accurate too, though.
+ glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F, fft_size, 1, 0, GL_RGBA, GL_HALF_FLOAT, tmp);
check_error();
delete[] tmp;