X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=fft_pass_effect.cpp;h=9c8aeb058e71fa33435901d4f5a2588a38ccb4f3;hp=a3de3792b380060f1c6effe112bfd5a1d0270344;hb=0597b51d28b5f095931c54a5e5f90ed9d5e18790;hpb=c4f0d4e876a8177db5738596f22349e030e0a1dc

diff --git a/fft_pass_effect.cpp b/fft_pass_effect.cpp
index a3de379..9c8aeb0 100644
--- a/fft_pass_effect.cpp
+++ b/fft_pass_effect.cpp
@@ -1,9 +1,16 @@
 #include <GL/glew.h>
+#include <math.h>
 
-#include "fft_pass_effect.h"
+#include "effect_chain.h"
 #include "effect_util.h"
+#include "fp16.h"
+#include "fft_pass_effect.h"
 #include "util.h"
 
+using namespace std;
+
+namespace movit {
+
 FFTPassEffect::FFTPassEffect()
 	: input_width(1280),
 	  input_height(720),
@@ -21,28 +28,26 @@ FFTPassEffect::~FFTPassEffect()
 	glDeleteTextures(1, &tex);
 }
 
-std::string FFTPassEffect::output_fragment_shader()
+string FFTPassEffect::output_fragment_shader()
 {
 	char buf[256];
 	sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
 	return buf + read_file("fft_pass_effect.frag");
 }
 
-void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
+void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
 {
 	Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
 
 	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();
@@ -74,7 +79,7 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const std::string &pre
 	// 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) {
@@ -119,30 +124,33 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const std::string &pre
 		} 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);
 	check_error();
-	glBindTexture(GL_TEXTURE_1D, tex);
+	glBindTexture(GL_TEXTURE_2D, tex);
 	check_error();
-	glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 	check_error();
-	glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 	check_error();
-	glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 	check_error();
 
 	// 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).
-	glTexImage1D(GL_TEXTURE_1D, 0, (subfft_size <= 4) ? GL_RGBA16F : GL_RGBA32F, fft_size, 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;
@@ -153,3 +161,5 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const std::string &pre
 	assert(input_size % fft_size == 0);
 	set_uniform_float(glsl_program_num, prefix, "num_repeats", input_size / fft_size);
 }
+
+}  // namespace movit