From 919101c59390dbbe380af7cc77102819e515a632 Mon Sep 17 00:00:00 2001 From: "Steinar H. Gunderson" Date: Sun, 2 Jul 2017 10:10:11 +0200 Subject: [PATCH] =?utf8?q?Inline=20combine=5Ftwo=5Fsamples=20(and=20remove?= =?utf8?q?=20an=20obsolete=20assert).=20Helps=2013=E2=80=9314%=20on=20Resa?= =?utf8?q?mpleEffect::calculate=5Ftexture.?= MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit --- util.cpp | 53 ----------------------------------------------------- util.h | 42 +++++++++++++++++++++++++++++++++++++++++- 2 files changed, 41 insertions(+), 54 deletions(-) diff --git a/util.cpp b/util.cpp index 54c815c..4016644 100644 --- a/util.cpp +++ b/util.cpp @@ -219,59 +219,6 @@ string output_glsl_vec3(const string &name, float x, float y, float z) return ss.str(); } -template -void combine_two_samples(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels, - DestFloat *offset, DestFloat *total_weight, float *sum_sq_error) -{ - assert(movit_initialized); - assert(w1 * w2 >= 0.0f); // Should not have differing signs. - float z; // Normalized 0..1 between pos1 and pos2. - if (fabs(w1 + w2) < 1e-6) { - z = 0.5f; - } else { - z = w2 / (w1 + w2); - } - - // Round to the desired precision. Note that this might take z outside the 0..1 range. - *offset = from_fp32(pos1 + z * pos1_pos2_diff); - z = (to_fp32(*offset) - pos1) * inv_pos1_pos2_diff; - - // Round to the minimum number of bits we have measured earlier. - // The card will do this for us anyway, but if we know what the real z - // is, we can pick a better total_weight below. - z = lrintf(z * num_subtexels) * inv_num_subtexels; - - // Choose total weight w so that we minimize total squared error - // for the effective weights: - // - // e = (w(1-z) - a)² + (wz - b)² - // - // Differentiating by w and setting equal to zero: - // - // 2(w(1-z) - a)(1-z) + 2(wz - b)z = 0 - // w(1-z)² - a(1-z) + wz² - bz = 0 - // w((1-z)² + z²) = a(1-z) + bz - // w = (a(1-z) + bz) / ((1-z)² + z²) - // - // If z had infinite precision, this would simply reduce to w = w1 + w2. - *total_weight = from_fp32((w1 + z * (w2 - w1)) / (z * z + (1 - z) * (1 - z))); - - if (sum_sq_error != NULL) { - float err1 = to_fp32(*total_weight) * (1 - z) - w1; - float err2 = to_fp32(*total_weight) * z - w2; - *sum_sq_error = err1 * err1 + err2 * err2; - } -} - -// Explicit instantiations. -template -void combine_two_samples(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels, - float *offset, float *total_weight, float *sum_sq_error); - -template -void combine_two_samples(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels, - fp16_int_t *offset, fp16_int_t *total_weight, float *sum_sq_error); - GLuint generate_vbo(GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data) { GLuint vbo; diff --git a/util.h b/util.h index 577140b..feefea5 100644 --- a/util.h +++ b/util.h @@ -9,6 +9,7 @@ #include #include #include "defs.h" +#include "fp16.h" #define BUFFER_OFFSET(i) ((char *)NULL + (i)) @@ -74,7 +75,46 @@ enum CombineRoundingBehavior { // and sum_sq_error. template void combine_two_samples(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels, - DestFloat *offset, DestFloat *total_weight, float *sum_sq_error); + DestFloat *offset, DestFloat *total_weight, float *sum_sq_error) +{ + assert(w1 * w2 >= 0.0f); // Should not have differing signs. + float z; // Normalized 0..1 between pos1 and pos2. + if (fabs(w1 + w2) < 1e-6) { + z = 0.5f; + } else { + z = w2 / (w1 + w2); + } + + // Round to the desired precision. Note that this might take z outside the 0..1 range. + *offset = from_fp32(pos1 + z * pos1_pos2_diff); + z = (to_fp32(*offset) - pos1) * inv_pos1_pos2_diff; + + // Round to the minimum number of bits we have measured earlier. + // The card will do this for us anyway, but if we know what the real z + // is, we can pick a better total_weight below. + z = lrintf(z * num_subtexels) * inv_num_subtexels; + + // Choose total weight w so that we minimize total squared error + // for the effective weights: + // + // e = (w(1-z) - a)² + (wz - b)² + // + // Differentiating by w and setting equal to zero: + // + // 2(w(1-z) - a)(1-z) + 2(wz - b)z = 0 + // w(1-z)² - a(1-z) + wz² - bz = 0 + // w((1-z)² + z²) = a(1-z) + bz + // w = (a(1-z) + bz) / ((1-z)² + z²) + // + // If z had infinite precision, this would simply reduce to w = w1 + w2. + *total_weight = from_fp32((w1 + z * (w2 - w1)) / (z * z + (1 - z) * (1 - z))); + + if (sum_sq_error != NULL) { + float err1 = to_fp32(*total_weight) * (1 - z) - w1; + float err2 = to_fp32(*total_weight) * z - w2; + *sum_sq_error = err1 * err1 + err2 * err2; + } +} // Create a VBO with the given data. Returns the VBO number. GLuint generate_vbo(GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data); -- 2.39.2