Store all twiddle factors as fp16.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 9 Mar 2014 16:41:03 +0000 (17:41 +0100)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 9 Mar 2014 17:25:37 +0000 (18:25 +0100)
Seemingly this isn't such a big deal after all, especially now that
we have proper fp16 roundoff also on ATI.

fft_pass_effect.cpp

index 51bde35..c2b47f0 100644 (file)
@@ -2,6 +2,7 @@
 #include <math.h>
 
 #include "effect_util.h"
+#include "fp16.h"
 #include "fft_pass_effect.h"
 #include "util.h"
 
@@ -79,7 +80,7 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
        // 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) {
@@ -124,10 +125,10 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
                } else {
                        support_texture_index = i;
                }
-               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((base - support_texture_index) / double(input_size));
+               tmp[support_texture_index * 4 + 1] = fp64_to_fp16((base + stride - 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);
@@ -143,11 +144,14 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
 
        // 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;