#define BLUR_FUNCTIONS ///////////////////////////////// MIT LICENSE //////////////////////////////// // Copyright (C) 2014 TroggleMonkey // // Permission is hereby granted, free of charge, to any person obtaining a copy // of this software and associated documentation files (the "Software"), to // deal in the Software without restriction, including without limitation the // rights to use, copy, modify, merge, publish, distribute, sublicense, and/or // sell copies of the Software, and to permit persons to whom the Software is // furnished to do so, subject to the following conditions: // // The above copyright notice and this permission notice shall be included in // all copies or substantial portions of the Software. // // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING // FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS // IN THE SOFTWARE. ///////////////////////////////// DESCRIPTION //////////////////////////////// // This file provides reusable one-pass and separable (two-pass) blurs. // Requires: All blurs share these requirements (dxdy requirement is split): // 1.) All requirements of gamma-management.h must be satisfied! // 2.) filter_linearN must == "true" in your .cgp preset unless // you're using tex2DblurNresize at 1x scale. // 3.) mipmap_inputN must == "true" in your .cgp preset if // IN.output_size < IN.video_size. // 4.) IN.output_size == IN.video_size / pow(2, M), where M is some // positive integer. tex2Dblur*resize can resize arbitrarily // (and the blur will be done after resizing), but arbitrary // resizes "fail" with other blurs due to the way they mix // static weights with bilinear sample exploitation. // 5.) In general, dxdy should contain the uv pixel spacing: // dxdy = (IN.video_size/IN.output_size)/IN.texture_size // 6.) For separable blurs (tex2DblurNresize and tex2DblurNfast), // zero out the dxdy component in the unblurred dimension: // dxdy = vec2(dxdy.x, 0.0) or vec2(0.0, dxdy.y) // Many blurs share these requirements: // 1.) One-pass blurs require scale_xN == scale_yN or scales > 1.0, // or they will blur more in the lower-scaled dimension. // 2.) One-pass shared sample blurs require ddx(), ddy(), and // tex2Dlod() to be supported by the current Cg profile, and // the drivers must support high-quality derivatives. // 3.) One-pass shared sample blurs require: // tex_uv.w == log2(IN.video_size/IN.output_size).y; // Non-wrapper blurs share this requirement: // 1.) sigma is the intended standard deviation of the blur // Wrapper blurs share this requirement, which is automatically // met (unless OVERRIDE_BLUR_STD_DEVS is #defined; see below): // 1.) blurN_std_dev must be global static const float values // specifying standard deviations for Nx blurs in units // of destination pixels // Optional: 1.) The including file (or an earlier included file) may // optionally #define USE_BINOMIAL_BLUR_STD_DEVS to replace // default standard deviations with those matching a binomial // distribution. (See below for details/properties.) // 2.) The including file (or an earlier included file) may // optionally #define OVERRIDE_BLUR_STD_DEVS and override: // static const float blur3_std_dev // static const float blur4_std_dev // static const float blur5_std_dev // static const float blur6_std_dev // static const float blur7_std_dev // static const float blur8_std_dev // static const float blur9_std_dev // static const float blur10_std_dev // static const float blur11_std_dev // static const float blur12_std_dev // static const float blur17_std_dev // static const float blur25_std_dev // static const float blur31_std_dev // static const float blur43_std_dev // 3.) The including file (or an earlier included file) may // optionally #define OVERRIDE_ERROR_BLURRING and override: // static const float error_blurring // This tuning value helps mitigate weighting errors from one- // pass shared-sample blurs sharing bilinear samples between // fragments. Values closer to 0.0 have "correct" blurriness // but allow more artifacts, and values closer to 1.0 blur away // artifacts by sampling closer to halfway between texels. // UPDATE 6/21/14: The above static constants may now be overridden // by non-static uniform constants. This permits exposing blur // standard deviations as runtime GUI shader parameters. However, // using them keeps weights from being statically computed, and the // speed hit depends on the blur: On my machine, uniforms kill over // 53% of the framerate with tex2Dblur12x12shared, but they only // drop the framerate by about 18% with tex2Dblur11fast. // Quality and Performance Comparisons: // For the purposes of the following discussion, "no sRGB" means // GAMMA_ENCODE_EVERY_FBO is #defined, and "sRGB" means it isn't. // 1.) tex2DblurNfast is always faster than tex2DblurNresize. // 2.) tex2DblurNresize functions are the only ones that can arbitrarily resize // well, because they're the only ones that don't exploit bilinear samples. // This also means they're the only functions which can be truly gamma- // correct without linear (or sRGB FBO) input, but only at 1x scale. // 3.) One-pass shared sample blurs only have a speed advantage without sRGB. // They also have some inaccuracies due to their shared-[bilinear-]sample // design, which grow increasingly bothersome for smaller blurs and higher- // frequency source images (relative to their resolution). I had high // hopes for them, but their most realistic use case is limited to quickly // reblurring an already blurred input at full resolution. Otherwise: // a.) If you're blurring a low-resolution source, you want a better blur. // b.) If you're blurring a lower mipmap, you want a better blur. // c.) If you're blurring a high-resolution, high-frequency source, you // want a better blur. // 4.) The one-pass blurs without shared samples grow slower for larger blurs, // but they're competitive with separable blurs at 5x5 and smaller, and // even tex2Dblur7x7 isn't bad if you're wanting to conserve passes. // Here are some framerates from a GeForce 8800GTS. The first pass resizes to // viewport size (4x in this test) and linearizes for sRGB codepaths, and the // remaining passes perform 6 full blurs. Mipmapped tests are performed at the // same scale, so they just measure the cost of mipmapping each FBO (only every // other FBO is mipmapped for separable blurs, to mimic realistic usage). // Mipmap Neither sRGB+Mipmap sRGB Function // 76.0 92.3 131.3 193.7 tex2Dblur3fast // 63.2 74.4 122.4 175.5 tex2Dblur3resize // 93.7 121.2 159.3 263.2 tex2Dblur3x3 // 59.7 68.7 115.4 162.1 tex2Dblur3x3resize // 63.2 74.4 122.4 175.5 tex2Dblur5fast // 49.3 54.8 100.0 132.7 tex2Dblur5resize // 59.7 68.7 115.4 162.1 tex2Dblur5x5 // 64.9 77.2 99.1 137.2 tex2Dblur6x6shared // 55.8 63.7 110.4 151.8 tex2Dblur7fast // 39.8 43.9 83.9 105.8 tex2Dblur7resize // 40.0 44.2 83.2 104.9 tex2Dblur7x7 // 56.4 65.5 71.9 87.9 tex2Dblur8x8shared // 49.3 55.1 99.9 132.5 tex2Dblur9fast // 33.3 36.2 72.4 88.0 tex2Dblur9resize // 27.8 29.7 61.3 72.2 tex2Dblur9x9 // 37.2 41.1 52.6 60.2 tex2Dblur10x10shared // 44.4 49.5 91.3 117.8 tex2Dblur11fast // 28.8 30.8 63.6 75.4 tex2Dblur11resize // 33.6 36.5 40.9 45.5 tex2Dblur12x12shared // TODO: Fill in benchmarks for new untested blurs. // tex2Dblur17fast // tex2Dblur25fast // tex2Dblur31fast // tex2Dblur43fast // tex2Dblur3x3resize ///////////////////////////// SETTINGS MANAGEMENT //////////////////////////// // Set static standard deviations, but allow users to override them with their // own constants (even non-static uniforms if they're okay with the speed hit): #ifndef OVERRIDE_BLUR_STD_DEVS // blurN_std_dev values are specified in terms of dxdy strides. #ifdef USE_BINOMIAL_BLUR_STD_DEVS // By request, we can define standard deviations corresponding to a // binomial distribution with p = 0.5 (related to Pascal's triangle). // This distribution works such that blurring multiple times should // have the same result as a single larger blur. These values are // larger than default for blurs up to 6x and smaller thereafter. const float blur3_std_dev = 0.84931640625; const float blur4_std_dev = 0.84931640625; const float blur5_std_dev = 1.0595703125; const float blur6_std_dev = 1.06591796875; const float blur7_std_dev = 1.17041015625; const float blur8_std_dev = 1.1720703125; const float blur9_std_dev = 1.2259765625; const float blur10_std_dev = 1.21982421875; const float blur11_std_dev = 1.25361328125; const float blur12_std_dev = 1.2423828125; const float blur17_std_dev = 1.27783203125; const float blur25_std_dev = 1.2810546875; const float blur31_std_dev = 1.28125; const float blur43_std_dev = 1.28125; #else // The defaults are the largest values that keep the largest unused // blur term on each side <= 1.0/256.0. (We could get away with more // or be more conservative, but this compromise is pretty reasonable.) const float blur3_std_dev = 0.62666015625; const float blur4_std_dev = 0.66171875; const float blur5_std_dev = 0.9845703125; const float blur6_std_dev = 1.02626953125; const float blur7_std_dev = 1.36103515625; const float blur8_std_dev = 1.4080078125; const float blur9_std_dev = 1.7533203125; const float blur10_std_dev = 1.80478515625; const float blur11_std_dev = 2.15986328125; const float blur12_std_dev = 2.215234375; const float blur17_std_dev = 3.45535583496; const float blur25_std_dev = 5.3409576416; const float blur31_std_dev = 6.86488037109; const float blur43_std_dev = 10.1852050781; #endif // USE_BINOMIAL_BLUR_STD_DEVS #endif // OVERRIDE_BLUR_STD_DEVS #ifndef OVERRIDE_ERROR_BLURRING // error_blurring should be in [0.0, 1.0]. Higher values reduce ringing // in shared-sample blurs but increase blurring and feature shifting. const float error_blurring = 0.5; #endif // Make a length squared helper macro (for usage with static constants): #define LENGTH_SQ(vec) (dot(vec, vec)) /////////////////////////////////// HELPERS ////////////////////////////////// vec4 uv2_to_uv4(vec2 tex_uv) { // Make a vec2 uv offset safe for adding to vec4 tex2Dlod coords: return vec4(tex_uv, 0.0, 0.0); } // Make a length squared helper macro (for usage with static constants): #define LENGTH_SQ(vec) (dot(vec, vec)) float get_fast_gaussian_weight_sum_inv(const float sigma) { // We can use the Gaussian integral to calculate the asymptotic weight for // the center pixel. Since the unnormalized center pixel weight is 1.0, // the normalized weight is the same as the weight sum inverse. Given a // large enough blur (9+), the asymptotic weight sum is close and faster: // center_weight = 0.5 * // (erf(0.5/(sigma*sqrt(2.0))) - erf(-0.5/(sigma*sqrt(2.0)))) // erf(-x) == -erf(x), so we get 0.5 * (2.0 * erf(blah blah)): // However, we can get even faster results with curve-fitting. These are // also closer than the asymptotic results, because they were constructed // from 64 blurs sizes from [3, 131) and 255 equally-spaced sigmas from // (0, blurN_std_dev), so the results for smaller sigmas are biased toward // smaller blurs. The max error is 0.0031793913. // Relative FPS: 134.3 with erf, 135.8 with curve-fitting. //static const float temp = 0.5/sqrt(2.0); //return erf(temp/sigma); return min(exp(exp(0.348348412457428/ (sigma - 0.0860587260734721))), 0.399334576340352/sigma); } //////////////////// ARBITRARILY RESIZABLE ONE-PASS BLURS //////////////////// vec3 tex2Dblur3x3resize(const sampler2D tex, const vec2 tex_uv, const vec2 dxdy, const float sigma) { // Requires: Global requirements must be met (see file description). // Returns: A 3x3 Gaussian blurred mipmapped texture lookup of the // resized input. // Description: // This is the only arbitrarily resizable one-pass blur; tex2Dblur5x5resize // would perform like tex2Dblur9x9, MUCH slower than tex2Dblur5resize. const float denom_inv = 0.5/(sigma*sigma); // Load each sample. We need all 3x3 samples. Quad-pixel communication // won't help either: This should perform like tex2Dblur5x5, but sharing a // 4x4 sample field would perform more like tex2Dblur8x8shared (worse). const vec2 sample4_uv = tex_uv; const vec2 dx = vec2(dxdy.x, 0.0); const vec2 dy = vec2(0.0, dxdy.y); const vec2 sample1_uv = sample4_uv - dy; const vec2 sample7_uv = sample4_uv + dy; const vec3 sample0 = tex2D_linearize(tex, sample1_uv - dx).rgb; const vec3 sample1 = tex2D_linearize(tex, sample1_uv).rgb; const vec3 sample2 = tex2D_linearize(tex, sample1_uv + dx).rgb; const vec3 sample3 = tex2D_linearize(tex, sample4_uv - dx).rgb; const vec3 sample4 = tex2D_linearize(tex, sample4_uv).rgb; const vec3 sample5 = tex2D_linearize(tex, sample4_uv + dx).rgb; const vec3 sample6 = tex2D_linearize(tex, sample7_uv - dx).rgb; const vec3 sample7 = tex2D_linearize(tex, sample7_uv).rgb; const vec3 sample8 = tex2D_linearize(tex, sample7_uv + dx).rgb; // Statically compute Gaussian sample weights: const float w4 = 1.0; const float w1_3_5_7 = exp(-LENGTH_SQ(vec2(1.0, 0.0)) * denom_inv); const float w0_2_6_8 = exp(-LENGTH_SQ(vec2(1.0, 1.0)) * denom_inv); const float weight_sum_inv = 1.0/(w4 + 4.0 * (w1_3_5_7 + w0_2_6_8)); // Weight and sum the samples: const vec3 sum = w4 * sample4 + w1_3_5_7 * (sample1 + sample3 + sample5 + sample7) + w0_2_6_8 * (sample0 + sample2 + sample6 + sample8); return sum * weight_sum_inv; } // Resizable one-pass blurs: vec3 tex2Dblur3x3resize(const sampler2D texture, const vec2 tex_uv, const vec2 dxdy) { return tex2Dblur3x3resize(texture, tex_uv, dxdy, blur3_std_dev); }