From 0a3c9320e21fc211f0c61a4bda1c6932920c6883 Mon Sep 17 00:00:00 2001 From: "Steinar H. Gunderson" Date: Wed, 3 Oct 2012 16:30:34 +0200 Subject: [PATCH] Flesh out the blur code a little. It actually blurs now, although it is not as pretty as I would like it. --- blur_effect.cpp | 157 +++++++++++++++++++++++++++++++++++++++-------- blur_effect.frag | 28 +++------ 2 files changed, 139 insertions(+), 46 deletions(-) diff --git a/blur_effect.cpp b/blur_effect.cpp index 17044f1..fd81aa8 100644 --- a/blur_effect.cpp +++ b/blur_effect.cpp @@ -8,6 +8,9 @@ #include "blur_effect.h" #include "util.h" +// Must match blur_effect.frag. +#define NUM_TAPS 16 + BlurEffect::BlurEffect() : radius(3.0f), direction(HORIZONTAL) @@ -25,16 +28,27 @@ void BlurEffect::set_uniforms(GLuint glsl_program_num, const std::string &prefix { Effect::set_uniforms(glsl_program_num, prefix, sampler_num); - // We only have 15 taps to work with, and we want that to reach out to about 2.5*sigma. - // Bump up the mipmap levels (giving us box blurs) until we have what we need. + int base_texture_size, texture_size; + if (direction == HORIZONTAL) { + base_texture_size = texture_size = 1280; // FIXME + } else if (direction == VERTICAL) { + base_texture_size = texture_size = 720; // FIXME + } else { + assert(false); + } + + // We only have 16 taps to work with on each side, and we want that to + // reach out to about 2.5*sigma. Bump up the mipmap levels (giving us + // box blurs) until we have what we need. + // + // FIXME: we really need to pick the same mipmap level for both horizontal and vertical! unsigned base_mipmap_level = 0; float adjusted_radius = radius; - float pixel_size = 1.0f; - while (adjusted_radius * 2.5f > 7.0f) { + while (texture_size > 1 && adjusted_radius * 2.5f > NUM_TAPS / 2) { ++base_mipmap_level; - adjusted_radius *= 0.5f; - pixel_size *= 2.0f; - } + texture_size /= 2; // Rounding down. + adjusted_radius = radius * float(texture_size) / float(base_texture_size); + } // In the second pass, we do the same, but don't sample from a mipmap; // that would re-blur the other direction in an ugly fashion, and we already @@ -47,32 +61,121 @@ void BlurEffect::set_uniforms(GLuint glsl_program_num, const std::string &prefix } glActiveTexture(GL_TEXTURE0); + check_error(); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, base_mipmap_level); check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, base_mipmap_level); + check_error(); - // FIXME - if (direction == HORIZONTAL) { - float ps[] = { pixel_size / 1280.0f, 0.0f }; - set_uniform_vec2(glsl_program_num, prefix, "pixel_offset", ps); - } else if (direction == VERTICAL) { - float ps[] = { 0.0f, pixel_size / 720.0f }; - set_uniform_vec2(glsl_program_num, prefix, "pixel_offset", ps); + // Compute the weights; they will be symmetrical, so we only compute + // the right side. + float weight[NUM_TAPS + 1]; + if (radius < 1e-3) { + weight[0] = 1.0f; + for (unsigned i = 1; i < NUM_TAPS + 1; ++i) { + weight[i] = 0.0f; + } } else { - assert(false); + float sum = 0.0f; + for (unsigned i = 0; i < NUM_TAPS + 1; ++i) { + float z = i / adjusted_radius; + + // Gaussian blur is a common, but maybe not the prettiest choice; + // it can feel a bit too blurry in the fine detail and too little + // long-tail. This is a simple logistic distribution, which has + // a narrower peak but longer tails. + weight[i] = 1.0f / (cosh(z) * cosh(z)); + + if (i == 0) { + sum += weight[i]; + } else { + sum += 2.0f * weight[i]; + } + } + for (unsigned i = 0; i < NUM_TAPS + 1; ++i) { + weight[i] /= sum; + } } - // Simple Gaussian weights for now. - float weight[15], total = 0.0f; - for (unsigned i = 0; i < 15; ++i) { - float z = (i - 7.0f) / adjusted_radius; - weight[i] = exp(-(z*z)); - total += weight[i]; +#if 0 + // NOTE: This is currently broken. + + // Since the GPU gives us bilinear sampling for free, we can get two + // samples for the price of one (for every but the center sample, + // in which case this trick doesn't buy us anything). Simply sample + // between the two pixel centers, and we can do with fewer weights. + // (This is right even in the vertical pass where we don't actually + // sample between the pixels, because we have linear interpolation + // there too.) + // + // We pack the parameters into a float4: The relative sample coordinates + // in (x,y), and the weight in z. w is unused. + float samples[4 * (NUM_TAPS / 2 + 1)]; + + // Center sample. + samples[4 * 0 + 0] = 0.0f; + samples[4 * 0 + 1] = 0.0f; + samples[4 * 0 + 2] = weight[0]; + samples[4 * 0 + 3] = 0.0f; + + // All other samples. + for (unsigned i = 1; i < NUM_TAPS / 2 + 1; ++i) { + unsigned base_pos = i * 2 - 1; + float w1 = weight[base_pos]; + float w2 = weight[base_pos + 1]; + + float offset, total_weight; + if (w1 + w2 < 1e-6) { + offset = 0.5f; + total_weight = 0.0f; + } else { + offset = w2 / (w1 + w2); + total_weight = w1 + w2; + } +#if 0 + // hack for easier visualization + offset = 0.5f; + total_weight = 8.0f; +#endif + float x = 0.0f, y = 0.0f; + + if (direction == HORIZONTAL) { + x = (base_pos + offset) / (float)texture_size; + } else if (direction == VERTICAL) { + y = (base_pos + offset) / (float)texture_size; + } else { + assert(false); + } + + samples[4 * i + 0] = x; + samples[4 * i + 1] = y; + samples[4 * i + 2] = total_weight; + samples[4 * i + 3] = 0.0f; } - printf("[mip level %d] ", base_mipmap_level); - for (unsigned i = 0; i < 15; ++i) { - weight[i] /= total; - printf("%f ", weight[i]); + + set_uniform_vec4_array(glsl_program_num, prefix, "samples", samples, NUM_TAPS / 2 + 1); +#else + // Boring, at-whole-pixels sampling. + float samples[4 * NUM_TAPS]; + + // All other samples. + for (unsigned i = 0; i < NUM_TAPS + 1; ++i) { + float x = 0.0f, y = 0.0f; + + if (direction == HORIZONTAL) { + x = i / (float)texture_size; + } else if (direction == VERTICAL) { + y = i / (float)texture_size; + } else { + assert(false); + } + + samples[4 * i + 0] = x; + samples[4 * i + 1] = y; + samples[4 * i + 2] = weight[i]; + samples[4 * i + 3] = 0.0f; } - printf("\n"); - set_uniform_float_array(glsl_program_num, prefix, "weight", weight, 15); + + set_uniform_vec4_array(glsl_program_num, prefix, "samples", samples, NUM_TAPS + 1); +#endif } diff --git a/blur_effect.frag b/blur_effect.frag index 8c0a01d..7e5a424 100644 --- a/blur_effect.frag +++ b/blur_effect.frag @@ -1,24 +1,14 @@ // A simple unidirectional blur. -uniform vec2 PREFIX(pixel_offset); -uniform float PREFIX(weight)[15]; +#define NUM_TAPS 16 + +uniform vec4 PREFIX(samples)[NUM_TAPS + 1]; vec4 FUNCNAME(vec2 tc) { - vec4 x = LAST_INPUT(tc); - return - vec4(PREFIX(weight)[ 0]) * LAST_INPUT(tc - 7.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 1]) * LAST_INPUT(tc - 6.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 2]) * LAST_INPUT(tc - 5.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 3]) * LAST_INPUT(tc - 4.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 4]) * LAST_INPUT(tc - 3.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 5]) * LAST_INPUT(tc - 2.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 6]) * LAST_INPUT(tc - PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 7]) * LAST_INPUT(tc) + - vec4(PREFIX(weight)[ 8]) * LAST_INPUT(tc + PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[ 9]) * LAST_INPUT(tc + 2.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[10]) * LAST_INPUT(tc + 3.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[11]) * LAST_INPUT(tc + 4.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[12]) * LAST_INPUT(tc + 5.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[13]) * LAST_INPUT(tc + 6.0 * PREFIX(pixel_offset)) + - vec4(PREFIX(weight)[14]) * LAST_INPUT(tc + 7.0 * PREFIX(pixel_offset)); + vec4 sum = vec4(PREFIX(samples)[0].z) * LAST_INPUT(tc); + for (int i = 1; i < NUM_TAPS + 1; ++i) { + vec4 sample = PREFIX(samples)[i]; + sum += vec4(sample.z) * (LAST_INPUT(tc - sample.xy) + LAST_INPUT(tc + sample.xy)); + } + return sum; } -- 2.39.2