path: root/libavfilter/vf_deshake_opencl.c

/*
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 *
 * Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
 * Copyright (C) 2009, Willow Garage Inc., all rights reserved.
 * Copyright (C) 2013, OpenCV Foundation, all rights reserved.
 * Third party copyrights are property of their respective owners.
 *
 * Redistribution and use in source and binary forms, with or without modification,
 * are permitted provided that the following conditions are met:
 *
 *   * Redistribution's of source code must retain the above copyright notice,
 *     this list of conditions and the following disclaimer.
 *
 *   * Redistribution's in binary form must reproduce the above copyright notice,
 *     this list of conditions and the following disclaimer in the documentation
 *     and/or other materials provided with the distribution.
 *
 *   * The name of the copyright holders may not be used to endorse or promote products
 *     derived from this software without specific prior written permission.
 *
 * This software is provided by the copyright holders and contributors "as is" and
 * any express or implied warranties, including, but not limited to, the implied
 * warranties of merchantability and fitness for a particular purpose are disclaimed.
 * In no event shall the Intel Corporation or contributors be liable for any direct,
 * indirect, incidental, special, exemplary, or consequential damages
 * (including, but not limited to, procurement of substitute goods or services;
 * loss of use, data, or profits; or business interruption) however caused
 * and on any theory of liability, whether in contract, strict liability,
 * or tort (including negligence or otherwise) arising in any way out of
 * the use of this software, even if advised of the possibility of such damage.
 */

#include <float.h>
#include <libavutil/lfg.h>
#include "libavutil/opt.h"
#include "libavutil/mem.h"
#include "libavutil/fifo.h"
#include "libavutil/common.h"
#include "libavutil/avassert.h"
#include "libavutil/pixdesc.h"
#include "libavutil/pixfmt.h"
#include "avfilter.h"
#include "framequeue.h"
#include "filters.h"
#include "transform.h"
#include "internal.h"
#include "opencl.h"
#include "opencl_source.h"
#include "video.h"

/*
This filter matches feature points between frames (dealing with outliers) and then
uses the matches to estimate an affine transform between frames. This transform is
decomposed into various values (translation, scale, rotation) and the values are
summed relative to the start of the video to obtain on absolute camera position
for each frame. This "camera path" is then smoothed via a gaussian filter, resulting
in a new path that is turned back into an affine transform and applied to each
frame to render it.

High-level overview:

All of the work to extract motion data from frames occurs in queue_frame. Motion data
is buffered in a smoothing window, so queue_frame simply computes the absolute camera
positions and places them in ringbuffers.

filter_frame is responsible for looking at the absolute camera positions currently
in the ringbuffers, applying the gaussian filter, and then transforming the frames.
*/

// Number of bits for BRIEF descriptors
#define BREIFN 512
// Size of the patch from which a BRIEF descriptor is extracted
// This is the size used in OpenCV
#define BRIEF_PATCH_SIZE 31
#define BRIEF_PATCH_SIZE_HALF (BRIEF_PATCH_SIZE / 2)

#define MATCHES_CONTIG_SIZE 2000

#define ROUNDED_UP_DIV(a, b) ((a + (b - 1)) / b)

typedef struct PointPair {
    // Previous frame
    cl_float2 p1;
    // Current frame
    cl_float2 p2;
} PointPair;

typedef struct MotionVector {
    PointPair p;
    // Used to mark vectors as potential outliers
    cl_int should_consider;
} MotionVector;

// Denotes the indices for the different types of motion in the ringbuffers array
enum RingbufferIndices {
    RingbufX,
    RingbufY,
    RingbufRot,
    RingbufScaleX,
    RingbufScaleY,

    // Should always be last
    RingbufCount
};

// Struct that holds data for drawing point match debug data
typedef struct DebugMatches {
    MotionVector *matches;
    // The points used to calculate the affine transform for a frame
    MotionVector model_matches[3];

    int num_matches;
    // For cases where we couldn't calculate a model
    int num_model_matches;
} DebugMatches;

// Groups together the ringbuffers that store absolute distortion / position values
// for each frame
typedef struct AbsoluteFrameMotion {
    // Array with the various ringbuffers, indexed via the RingbufferIndices enum
    AVFifo *ringbuffers[RingbufCount];

    // Offset to get to the current frame being processed
    // (not in bytes)
    int curr_frame_offset;
    // Keeps track of where the start and end of contiguous motion data is (to
    // deal with cases where no motion data is found between two frames)
    int data_start_offset;
    int data_end_offset;

    AVFifo *debug_matches;
} AbsoluteFrameMotion;

// Takes care of freeing the arrays within the DebugMatches inside of the
// debug_matches ringbuffer and then freeing the buffer itself.
static void free_debug_matches(AbsoluteFrameMotion *afm) {
    DebugMatches dm;

    if (!afm->debug_matches) {
        return;
    }

    while (av_fifo_read(afm->debug_matches, &dm, 1) >= 0)
        av_freep(&dm.matches);

    av_fifo_freep2(&afm->debug_matches);
}

// Stores the translation, scale, rotation, and skew deltas between two frames
typedef struct FrameDelta {
    cl_float2 translation;
    float rotation;
    cl_float2 scale;
    cl_float2 skew;
} FrameDelta;

typedef struct SimilarityMatrix {
    // The 2x3 similarity matrix
    double matrix[6];
} SimilarityMatrix;

typedef struct CropInfo {
    // The top left corner of the bounding box for the crop
    cl_float2 top_left;
    // The bottom right corner of the bounding box for the crop
    cl_float2 bottom_right;
} CropInfo;

// Returned from function that determines start and end values for iteration
// around the current frame in a ringbuffer
typedef struct IterIndices {
    int start;
    int end;
} IterIndices;

typedef struct DeshakeOpenCLContext {
    OpenCLFilterContext ocf;
    // Whether or not the above `OpenCLFilterContext` has been initialized
    int initialized;

    // These variables are used in the activate callback
    int64_t duration;
    int eof;

    // State for random number generation
    AVLFG alfg;

    // FIFO frame queue used to buffer future frames for processing
    FFFrameQueue fq;
    // Ringbuffers for frame positions
    AbsoluteFrameMotion abs_motion;

    // The number of frames' motion to consider before and after the frame we are
    // smoothing
    int smooth_window;
    // The number of the frame we are currently processing
    int curr_frame;

    // Stores a 1d array of normalised gaussian kernel values for convolution
    float *gauss_kernel;

    // Buffer for error values used in RANSAC code
    float *ransac_err;

    // Information regarding how to crop the smoothed luminance (or RGB) planes
    CropInfo crop_y;
    // Information regarding how to crop the smoothed chroma planes
    CropInfo crop_uv;

    // Whether or not we are processing YUV input (as oppposed to RGB)
    int is_yuv;
    // The underlying format of the hardware surfaces
    int sw_format;

    // Buffer to copy `matches` into for the CPU to work with
    MotionVector *matches_host;
    MotionVector *matches_contig_host;

    MotionVector *inliers;

    cl_command_queue command_queue;
    cl_kernel kernel_grayscale;
    cl_kernel kernel_harris_response;
    cl_kernel kernel_refine_features;
    cl_kernel kernel_brief_descriptors;
    cl_kernel kernel_match_descriptors;
    cl_kernel kernel_transform;
    cl_kernel kernel_crop_upscale;

    // Stores a frame converted to grayscale
    cl_mem grayscale;
    // Stores the harris response for a frame (measure of "cornerness" for each pixel)
    cl_mem harris_buf;

    // Detected features after non-maximum suppression and sub-pixel refinement
    cl_mem refined_features;
    // Saved from the previous frame
    cl_mem prev_refined_features;

    // BRIEF sampling pattern that is randomly initialized
    cl_mem brief_pattern;
    // Feature point descriptors for the current frame
    cl_mem descriptors;
    // Feature point descriptors for the previous frame
    cl_mem prev_descriptors;
    // Vectors between points in current and previous frame
    cl_mem matches;
    cl_mem matches_contig;
    // Holds the matrix to transform luminance (or RGB) with
    cl_mem transform_y;
    // Holds the matrix to transform chroma with
    cl_mem transform_uv;

    // Configurable options

    int tripod_mode;
    int debug_on;
    int should_crop;

    // Whether or not feature points should be refined at a sub-pixel level
    cl_int refine_features;
    // If the user sets a value other than the default, 0, this percentage is
    // translated into a sigma value ranging from 0.5 to 40.0
    float smooth_percent;
    // This number is multiplied by the video frame rate to determine the size
    // of the smooth window
    float smooth_window_multiplier;

    // Debug stuff

    cl_kernel kernel_draw_debug_info;
    cl_mem debug_matches;
    cl_mem debug_model_matches;

    // These store the total time spent executing the different kernels in nanoseconds
    unsigned long long grayscale_time;
    unsigned long long harris_response_time;
    unsigned long long refine_features_time;
    unsigned long long brief_descriptors_time;
    unsigned long long match_descriptors_time;
    unsigned long long transform_time;
    unsigned long long crop_upscale_time;

    // Time spent copying matched features from the device to the host
    unsigned long long read_buf_time;
} DeshakeOpenCLContext;

// Returns a random uniformly-distributed number in [low, high]
static int rand_in(int low, int high, AVLFG *alfg) {
    return (av_lfg_get(alfg) % (high - low)) + low;
}

// Returns the average execution time for an event given the total time and the
// number of frames processed.
static double averaged_event_time_ms(unsigned long long total_time, int num_frames) {
    return (double)total_time / (double)num_frames / 1000000.0;
}

// The following code is loosely ported from OpenCV

// Estimates affine transform from 3 point pairs
// model is a 2x3 matrix:
//      a b c
//      d e f
static void run_estimate_kernel(const MotionVector *point_pairs, double *model)
{
    // src points
    double x1 = point_pairs[0].p.p1.s[0];
    double y1 = point_pairs[0].p.p1.s[1];
    double x2 = point_pairs[1].p.p1.s[0];
    double y2 = point_pairs[1].p.p1.s[1];
    double x3 = point_pairs[2].p.p1.s[0];
    double y3 = point_pairs[2].p.p1.s[1];

    // dest points
    double X1 = point_pairs[0].p.p2.s[0];
    double Y1 = point_pairs[0].p.p2.s[1];
    double X2 = point_pairs[1].p.p2.s[0];
    double Y2 = point_pairs[1].p.p2.s[1];
    double X3 = point_pairs[2].p.p2.s[0];
    double Y3 = point_pairs[2].p.p2.s[1];

    double d = 1.0 / ( x1*(y2-y3) + x2*(y3-y1) + x3*(y1-y2) );

    model[0] = d * ( X1*(y2-y3) + X2*(y3-y1) + X3*(y1-y2) );
    model[1] = d * ( X1*(x3-x2) + X2*(x1-x3) + X3*(x2-x1) );
    model[2] = d * ( X1*(x2*y3 - x3*y2) + X2*(x3*y1 - x1*y3) + X3*(x1*y2 - x2*y1) );

    model[3] = d * ( Y1*(y2-y3) + Y2*(y3-y1) + Y3*(y1-y2) );
    model[4] = d * ( Y1*(x3-x2) + Y2*(x1-x3) + Y3*(x2-x1) );
    model[5] = d * ( Y1*(x2*y3 - x3*y2) + Y2*(x3*y1 - x1*y3) + Y3*(x1*y2 - x2*y1) );
}

// Checks that the 3 points in the given array are not collinear
static int points_not_collinear(const cl_float2 **points)
{
    int j, k, i = 2;

    for (j = 0; j < i; j++) {
        double dx1 = points[j]->s[0] - points[i]->s[0];
        double dy1 = points[j]->s[1] - points[i]->s[1];

        for (k = 0; k < j; k++) {
            double dx2 = points[k]->s[0] - points[i]->s[0];
            double dy2 = points[k]->s[1] - points[i]->s[1];

            // Assuming a 3840 x 2160 video with a point at (0, 0) and one at
            // (3839, 2159), this prevents a third point from being within roughly
            // 0.5 of a pixel of the line connecting the two on both axes
            if (fabs(dx2*dy1 - dy2*dx1) <= 1.0) {
                return 0;
            }
        }
    }

    return 1;
}

// Checks a subset of 3 point pairs to make sure that the points are not collinear
// and not too close to each other
static int check_subset(const MotionVector *pairs_subset)
{
    const cl_float2 *prev_points[] = {
        &pairs_subset[0].p.p1,
        &pairs_subset[1].p.p1,
        &pairs_subset[2].p.p1
    };

    const cl_float2 *curr_points[] = {
        &pairs_subset[0].p.p2,
        &pairs_subset[1].p.p2,
        &pairs_subset[2].p.p2
    };

    return points_not_collinear(prev_points) && points_not_collinear(curr_points);
}

// Selects a random subset of 3 points from point_pairs and places them in pairs_subset
static int get_subset(
    AVLFG *alfg,
    const MotionVector *point_pairs,
    const int num_point_pairs,
    MotionVector *pairs_subset,
    int max_attempts
) {
    int idx[3];
    int i = 0, j, iters = 0;

    for (; iters < max_attempts; iters++) {
        for (i = 0; i < 3 && iters < max_attempts;) {
            int idx_i = 0;

            for (;;) {
                idx_i = idx[i] = rand_in(0, num_point_pairs, alfg);

                for (j = 0; j < i; j++) {
                    if (idx_i == idx[j]) {
                        break;
                    }
                }

                if (j == i) {
                    break;
                }
            }

            pairs_subset[i] = point_pairs[idx[i]];
            i++;
        }

        if (i == 3 && !check_subset(pairs_subset)) {
            continue;
        }
        break;
    }

    return i == 3 && iters < max_attempts;
}

// Computes the error for each of the given points based on the given model.
static void compute_error(
    const MotionVector *point_pairs,
    const int num_point_pairs,
    const double *model,
    float *err
) {
    double F0 = model[0], F1 = model[1], F2 = model[2];
    double F3 = model[3], F4 = model[4], F5 = model[5];

    for (int i = 0; i < num_point_pairs; i++) {
        const cl_float2 *f = &point_pairs[i].p.p1;
        const cl_float2 *t = &point_pairs[i].p.p2;

        double a = F0*f->s[0] + F1*f->s[1] + F2 - t->s[0];
        double b = F3*f->s[0] + F4*f->s[1] + F5 - t->s[1];

        err[i] = a*a + b*b;
    }
}

// Determines which of the given point matches are inliers for the given model
// based on the specified threshold.
//
// err must be an array of num_point_pairs length
static int find_inliers(
    MotionVector *point_pairs,
    const int num_point_pairs,
    const double *model,
    float *err,
    double thresh
) {
    float t = (float)(thresh * thresh);
    int i, n = num_point_pairs, num_inliers = 0;

    compute_error(point_pairs, num_point_pairs, model, err);

    for (i = 0; i < n; i++) {
        if (err[i] <= t) {
            // This is an inlier
            point_pairs[i].should_consider = 1;
            num_inliers += 1;
        } else {
            point_pairs[i].should_consider = 0;
        }
    }

    return num_inliers;
}

// Determines the number of iterations required to achieve the desired confidence level.
//
// The equation used to determine the number of iterations to do is:
// 1 - confidence = (1 - inlier_probability^num_points)^num_iters
//
// Solving for num_iters:
//
// num_iters = log(1 - confidence) / log(1 - inlier_probability^num_points)
//
// A more in-depth explanation can be found at https://en.wikipedia.org/wiki/Random_sample_consensus
// under the 'Parameters' heading
static int ransac_update_num_iters(double confidence, double num_outliers, int max_iters)
{
    double num, denom;

    confidence   = av_clipd(confidence, 0.0, 1.0);
    num_outliers = av_clipd(num_outliers, 0.0, 1.0);

    // avoid inf's & nan's
    num = FFMAX(1.0 - confidence, DBL_MIN);
    denom = 1.0 - pow(1.0 - num_outliers, 3);
    if (denom < DBL_MIN) {
        return 0;
    }

    num = log(num);
    denom = log(denom);

    return denom >= 0 || -num >= max_iters * (-denom) ? max_iters : (int)round(num / denom);
}

// Estimates an affine transform between the given pairs of points using RANdom
// SAmple Consensus
static int estimate_affine_2d(
    DeshakeOpenCLContext *deshake_ctx,
    MotionVector *point_pairs,
    DebugMatches *debug_matches,
    const int num_point_pairs,
    double *model_out,
    const double threshold,
    const int max_iters,
    const double confidence
) {
    int result = 0;
    double best_model[6], model[6];
    MotionVector pairs_subset[3], best_pairs[3];

    int iter, niters = FFMAX(max_iters, 1);
    int good_count, max_good_count = 0;

    // We need at least 3 points to build a model from
    if (num_point_pairs < 3) {
        return 0;
    } else if (num_point_pairs == 3) {
        // There are only 3 points, so RANSAC doesn't apply here
        run_estimate_kernel(point_pairs, model_out);

        for (int i = 0; i < 3; ++i) {
            point_pairs[i].should_consider = 1;
        }

        return 1;
    }

    for (iter = 0; iter < niters; ++iter) {
        int found = get_subset(&deshake_ctx->alfg, point_pairs, num_point_pairs, pairs_subset, 10000);

        if (!found) {
            if (iter == 0) {
                return 0;
            }

            break;
        }

        run_estimate_kernel(pairs_subset, model);
        good_count = find_inliers(point_pairs, num_point_pairs, model, deshake_ctx->ransac_err, threshold);

        if (good_count > FFMAX(max_good_count, 2)) {
            for (int mi = 0; mi < 6; ++mi) {
                best_model[mi] = model[mi];
            }

            for (int pi = 0; pi < 3; pi++) {
                best_pairs[pi] = pairs_subset[pi];
            }

            max_good_count = good_count;
            niters = ransac_update_num_iters(
                confidence,
                (double)(num_point_pairs - good_count) / num_point_pairs,
                niters
            );
        }
    }

    if (max_good_count > 0) {
        for (int mi = 0; mi < 6; ++mi) {
            model_out[mi] = best_model[mi];
        }

        for (int pi = 0; pi < 3; ++pi) {
            debug_matches->model_matches[pi] = best_pairs[pi];
        }
        debug_matches->num_model_matches = 3;

        // Find the inliers again for the best model for debugging
        find_inliers(point_pairs, num_point_pairs, best_model, deshake_ctx->ransac_err, threshold);
        result = 1;
    }

    return result;
}

// "Wiggles" the first point in best_pairs around a tiny bit in order to decrease the
// total error
static void optimize_model(
    DeshakeOpenCLContext *deshake_ctx,
    MotionVector *best_pairs,
    MotionVector *inliers,
    const int num_inliers,
    float best_err,
    double *model_out
) {
    float move_x_val = 0.01;
    float move_y_val = 0.01;
    int move_x = 1;
    float old_move_x_val = 0;
    double model[6];
    int last_changed = 0;

    for (int iters = 0; iters < 200; iters++) {
        float total_err = 0;

        if (move_x) {
            best_pairs[0].p.p2.s[0] += move_x_val;
        } else {
            best_pairs[0].p.p2.s[0] += move_y_val;
        }

        run_estimate_kernel(best_pairs, model);
        compute_error(inliers, num_inliers, model, deshake_ctx->ransac_err);

        for (int j = 0; j < num_inliers; j++) {
            total_err += deshake_ctx->ransac_err[j];
        }

        if (total_err < best_err) {
            for (int mi = 0; mi < 6; ++mi) {
                model_out[mi] = model[mi];
            }

            best_err = total_err;
            last_changed = iters;
        } else {
            // Undo the change
            if (move_x) {
                best_pairs[0].p.p2.s[0] -= move_x_val;
            } else {
                best_pairs[0].p.p2.s[0] -= move_y_val;
            }

            if (iters - last_changed > 4) {
                // We've already improved the model as much as we can
                break;
            }

            old_move_x_val = move_x_val;

            if (move_x) {
                move_x_val *= -1;
            } else {
                move_y_val *= -1;
            }

            if (old_move_x_val < 0) {
                move_x = 0;
            } else {
                move_x = 1;
            }
        }
    }
}

// Uses a process similar to that of RANSAC to find a transform that minimizes
// the total error for a set of point matches determined to be inliers
//
// (Pick random subsets, compute model, find total error, iterate until error
// is minimized.)
static int minimize_error(
    DeshakeOpenCLContext *deshake_ctx,
    MotionVector *inliers,
    DebugMatches *debug_matches,
    const int num_inliers,
    double *model_out,
    const int max_iters
) {
    int result = 0;
    float best_err = FLT_MAX;
    double best_model[6], model[6];
    MotionVector pairs_subset[3], best_pairs[3];

    for (int i = 0; i < max_iters; i++) {
        float total_err = 0;
        int found = get_subset(&deshake_ctx->alfg, inliers, num_inliers, pairs_subset, 10000);

        if (!found) {
            if (i == 0) {
                return 0;
            }

            break;
        }

        run_estimate_kernel(pairs_subset, model);
        compute_error(inliers, num_inliers, model, deshake_ctx->ransac_err);

        for (int j = 0; j < num_inliers; j++) {
            total_err += deshake_ctx->ransac_err[j];
        }

        if (i == 0 || total_err < best_err) {
            for (int mi = 0; mi < 6; ++mi) {
                best_model[mi] = model[mi];
            }

            for (int pi = 0; pi < 3; pi++) {
                best_pairs[pi] = pairs_subset[pi];
            }

            best_err = total_err;
        }
    }

    for (int mi = 0; mi < 6; ++mi) {
        model_out[mi] = best_model[mi];
    }

    for (int pi = 0; pi < 3; ++pi) {
        debug_matches->model_matches[pi] = best_pairs[pi];
    }
    debug_matches->num_model_matches = 3;
    result = 1;

    optimize_model(deshake_ctx, best_pairs, inliers, num_inliers, best_err, model_out);
    return result;
}

// End code from OpenCV

// Decomposes a similarity matrix into translation, rotation, scale, and skew
//
// See http://frederic-wang.fr/decomposition-of-2d-transform-matrices.html
static FrameDelta decompose_transform(double *model)
{
    FrameDelta ret;

    double a = model[0];
    double c = model[1];
    double e = model[2];
    double b = model[3];
    double d = model[4];
    double f = model[5];
    double delta = a * d - b * c;

    memset(&ret, 0, sizeof(ret));

    ret.translation.s[0] = e;
    ret.translation.s[1] = f;

    // This is the QR method
    if (a != 0 || b != 0) {
        double r = hypot(a, b);

        ret.rotation = FFSIGN(b) * acos(a / r);
        ret.scale.s[0] = r;
        ret.scale.s[1] = delta / r;
        ret.skew.s[0] = atan((a * c + b * d) / (r * r));
        ret.skew.s[1] = 0;
    } else if (c != 0 || d != 0) {
        double s = sqrt(c * c + d * d);

        ret.rotation = M_PI / 2 - FFSIGN(d) * acos(-c / s);
        ret.scale.s[0] = delta / s;
        ret.scale.s[1] = s;
        ret.skew.s[0] = 0;
        ret.skew.s[1] = atan((a * c + b * d) / (s * s));
    } // otherwise there is only translation

    return ret;
}

// Move valid vectors from the 2d buffer into a 1d buffer where they are contiguous
static int make_vectors_contig(
    DeshakeOpenCLContext *deshake_ctx,
    int size_y,
    int size_x
) {
    int num_vectors = 0;

    for (int i = 0; i < size_y; ++i) {
        for (int j = 0; j < size_x; ++j) {
            MotionVector v = deshake_ctx->matches_host[j + i * size_x];

            if (v.should_consider) {
                deshake_ctx->matches_contig_host[num_vectors] = v;
                ++num_vectors;
            }

            // Make sure we do not exceed the amount of space we allocated for these vectors
            if (num_vectors == MATCHES_CONTIG_SIZE - 1) {
                return num_vectors;
            }
        }
    }
    return num_vectors;
}

// Returns the gaussian kernel value for the given x coordinate and sigma value
static float gaussian_for(int x, float sigma) {
    return 1.0f / expf(((float)x * (float)x) / (2.0f * sigma * sigma));
}

// Makes a normalized gaussian kernel of the given length for the given sigma
// and places it in gauss_kernel
static void make_gauss_kernel(float *gauss_kernel, float length, float sigma)
{
    float gauss_sum = 0;
    int window_half = length / 2;

    for (int i = 0; i < length; ++i) {
        float val = gaussian_for(i - window_half, sigma);

        gauss_sum += val;
        gauss_kernel[i] = val;
    }

    // Normalize the gaussian values
    for (int i = 0; i < length; ++i) {
        gauss_kernel[i] /= gauss_sum;
    }
}

// Returns indices to start and end iteration at in order to iterate over a window
// of length size centered at the current frame in a ringbuffer
//
// Always returns numbers that result in a window of length size, even if that
// means specifying negative indices or indices past the end of the values in the
// ringbuffers. Make sure you clip indices appropriately within your loop.
static IterIndices start_end_for(DeshakeOpenCLContext *deshake_ctx, int length) {
    IterIndices indices;

    indices.start = deshake_ctx->abs_motion.curr_frame_offset - (length / 2);
    indices.end = deshake_ctx->abs_motion.curr_frame_offset + (length / 2) + (length % 2);

    return indices;
}

// Sets val to the value in the given ringbuffer at the given offset, taking care of
// clipping the offset into the appropriate range
static void ringbuf_float_at(
    DeshakeOpenCLContext *deshake_ctx,
    AVFifo *values,
    float *val,
    int offset
) {
    int clip_start, clip_end, offset_clipped;
    if (deshake_ctx->abs_motion.data_end_offset != -1) {
        clip_end = deshake_ctx->abs_motion.data_end_offset;
    } else {
        // This expression represents the last valid index in the buffer,
        // which we use repeatedly at the end of the video.
        clip_end = deshake_ctx->smooth_window - av_fifo_can_write(values) - 1;
    }

    if (deshake_ctx->abs_motion.data_start_offset != -1) {
        clip_start = deshake_ctx->abs_motion.data_start_offset;
    } else {
        // Negative indices will occur at the start of the video, and we want
        // them to be clipped to 0 in order to repeatedly use the position of
        // the first frame.
        clip_start = 0;
    }

    offset_clipped = av_clip(
        offset,
        clip_start,
        clip_end
    );

    av_fifo_peek(values, val, 1, offset_clipped);
}

// Returns smoothed current frame value of the given buffer of floats based on the
// given Gaussian kernel and its length (also the window length, centered around the
// current frame) and the "maximum value" of the motion.
//
// This "maximum value" should be the width / height of the image in the case of
// translation and an empirically chosen constant for rotation / scale.
//
// The sigma chosen to generate the final gaussian kernel with used to smooth the
// camera path is either hardcoded (set by user, deshake_ctx->smooth_percent) or
// adaptively chosen.
static float smooth(
    DeshakeOpenCLContext *deshake_ctx,
    float *gauss_kernel,
    int length,
    float max_val,
    AVFifo *values
) {
    float new_large_s = 0, new_small_s = 0, new_best = 0, old, diff_between,
          percent_of_max, inverted_percent;
    IterIndices indices = start_end_for(deshake_ctx, length);
    float large_sigma = 40.0f;
    float small_sigma = 2.0f;
    float best_sigma;

    if (deshake_ctx->smooth_percent) {
        best_sigma = (large_sigma - 0.5f) * deshake_ctx->smooth_percent + 0.5f;
    } else {
        // Strategy to adaptively smooth trajectory:
        //
        // 1. Smooth path with large and small sigma values
        // 2. Take the absolute value of the difference between them
        // 3. Get a percentage by putting the difference over the "max value"
        // 4, Invert the percentage
        // 5. Calculate a new sigma value weighted towards the larger sigma value
        // 6. Determine final smoothed trajectory value using that sigma

        make_gauss_kernel(gauss_kernel, length, large_sigma);
        for (int i = indices.start, j = 0; i < indices.end; ++i, ++j) {
            ringbuf_float_at(deshake_ctx, values, &old, i);
            new_large_s += old * gauss_kernel[j];
        }

        make_gauss_kernel(gauss_kernel, length, small_sigma);
        for (int i = indices.start, j = 0; i < indices.end; ++i, ++j) {
            ringbuf_float_at(deshake_ctx, values, &old, i);
            new_small_s += old * gauss_kernel[j];
        }

        diff_between = fabsf(new_large_s - new_small_s);
        percent_of_max = diff_between / max_val;
        inverted_percent = 1 - percent_of_max;
        best_sigma = large_sigma * powf(inverted_percent, 40);
    }

    make_gauss_kernel(gauss_kernel, length, best_sigma);
    for (int i = indices.start, j = 0; i < indices.end; ++i, ++j) {
        ringbuf_float_at(deshake_ctx, values, &old, i);
        new_best += old * gauss_kernel[j];
    }

    return new_best;
}

// Returns the position of the given point after the transform is applied
static cl_float2 transformed_point(float x, float y, float *transform) {
    cl_float2 ret;

    ret.s[0] = x * transform[0] + y * transform[1] + transform[2];
    ret.s[1] = x * transform[3] + y * transform[4] + transform[5];

    return ret;
}

// Creates an affine transform that scales from the center of a frame
static void transform_center_scale(
    float x_shift,
    float y_shift,
    float angle,
    float scale_x,
    float scale_y,
    float center_w,
    float center_h,
    float *matrix
) {
    cl_float2 center_s;
    float center_s_w, center_s_h;

    ff_get_matrix(
        0,
        0,
        0,
        scale_x,
        scale_y,
        matrix
    );

    center_s = transformed_point(center_w, center_h, matrix);
    center_s_w = center_w - center_s.s[0];
    center_s_h = center_h - center_s.s[1];

    ff_get_matrix(
        x_shift + center_s_w,
        y_shift + center_s_h,
        angle,
        scale_x,
        scale_y,
        matrix
    );
}

// Determines the crop necessary to eliminate black borders from a smoothed frame
// and updates target crop accordingly
static void update_needed_crop(
    CropInfo* crop,
    float *transform,
    float frame_width,
    float frame_height
) {
    float new_width, new_height, adjusted_width, adjusted_height, adjusted_x, adjusted_y;

    cl_float2 top_left = transformed_point(0, 0, transform);
    cl_float2 top_right = transformed_point(frame_width, 0, transform);
    cl_float2 bottom_left = transformed_point(0, frame_height, transform);
    cl_float2 bottom_right = transformed_point(frame_width, frame_height, transform);
    float ar_h = frame_height / frame_width;
    float ar_w = frame_width / frame_height;

    if (crop->bottom_right.s[0] == 0) {
        // The crop hasn't been set to the original size of the plane
        crop->bottom_right.s[0] = frame_width;
        crop->bottom_right.s[1] = frame_height;
    }

    crop->top_left.s[0] = FFMAX3(
        crop->top_left.s[0],
        top_left.s[0],
        bottom_left.s[0]
    );

    crop->top_left.s[1] = FFMAX3(
        crop->top_left.s[1],
        top_left.s[1],
        top_right.s[1]
    );

    crop->bottom_right.s[0] = FFMIN3(
        crop->bottom_right.s[0],
        bottom_right.s[0],
        top_right.s[0]
    );

    crop->bottom_right.s[1] = FFMIN3(
        crop->bottom_right.s[1],
        bottom_right.s[1],
        bottom_left.s[1]
    );

    // Make sure our potentially new bounding box has the same aspect ratio
    new_height = crop->bottom_right.s[1] - crop->top_left.s[1];
    new_width = crop->bottom_right.s[0] - crop->top_left.s[0];

    adjusted_width = new_height * ar_w;
    adjusted_x = crop->bottom_right.s[0] - adjusted_width;

    if (adjusted_x >= crop->top_left.s[0]) {
        crop->top_left.s[0] = adjusted_x;
    } else {
        adjusted_height = new_width * ar_h;
        adjusted_y = crop->bottom_right.s[1] - adjusted_height;
        crop->top_left.s[1] = adjusted_y;
    }
}

static av_cold void deshake_opencl_uninit(AVFilterContext *avctx)
{
    DeshakeOpenCLContext *ctx = avctx->priv;
    cl_int cle;

    for (int i = 0; i < RingbufCount; i++)
        av_fifo_freep2(&ctx->abs_motion.ringbuffers[i]);

    if (ctx->debug_on)
        free_debug_matches(&ctx->abs_motion);

    if (ctx->gauss_kernel)
        av_freep(&ctx->gauss_kernel);

    if (ctx->ransac_err)
        av_freep(&ctx->ransac_err);

    if (ctx->matches_host)
        av_freep(&ctx->matches_host);

    if (ctx->matches_contig_host)
        av_freep(&ctx->matches_contig_host);

    if (ctx->inliers)
        av_freep(&ctx->inliers);

    ff_framequeue_free(&ctx->fq);

    CL_RELEASE_KERNEL(ctx->kernel_grayscale);
    CL_RELEASE_KERNEL(ctx->kernel_harris_response);
    CL_RELEASE_KERNEL(ctx->kernel_refine_features);
    CL_RELEASE_KERNEL(ctx->kernel_brief_descriptors);
    CL_RELEASE_KERNEL(ctx->kernel_match_descriptors);
    CL_RELEASE_KERNEL(ctx->kernel_crop_upscale);
    if (ctx->debug_on)
        CL_RELEASE_KERNEL(ctx->kernel_draw_debug_info);

    CL_RELEASE_QUEUE(ctx->command_queue);

    if (!ctx->is_yuv)
        CL_RELEASE_MEMORY(ctx->grayscale);
    CL_RELEASE_MEMORY(ctx->harris_buf);
    CL_RELEASE_MEMORY(ctx->refined_features);
    CL_RELEASE_MEMORY(ctx->prev_refined_features);
    CL_RELEASE_MEMORY(ctx->brief_pattern);
    CL_RELEASE_MEMORY(ctx->descriptors);
    CL_RELEASE_MEMORY(ctx->prev_descriptors);
    CL_RELEASE_MEMORY(ctx->matches);
    CL_RELEASE_MEMORY(ctx->matches_contig);
    CL_RELEASE_MEMORY(ctx->transform_y);
    CL_RELEASE_MEMORY(ctx->transform_uv);
    if (ctx->debug_on) {
        CL_RELEASE_MEMORY(ctx->debug_matches);
        CL_RELEASE_MEMORY(ctx->debug_model_matches);
    }

    ff_opencl_filter_uninit(avctx);
}

static int deshake_opencl_init(AVFilterContext *avctx)
{
    DeshakeOpenCLContext *ctx = avctx->priv;
    AVFilterLink *outlink = avctx->outputs[0];
    AVFilterLink *inlink = avctx->inputs[0];
    // Pointer to the host-side pattern buffer to be initialized and then copied
    // to the GPU
    PointPair *pattern_host = NULL;
    cl_int cle;
    int err;
    cl_ulong8 zeroed_ulong8;
    FFFrameQueueGlobal fqg;
    cl_image_format grayscale_format;
    cl_image_desc grayscale_desc;
    cl_command_queue_properties queue_props;

    const enum AVPixelFormat disallowed_formats[14] = {
        AV_PIX_FMT_GBRP,
        AV_PIX_FMT_GBRP9BE,
        AV_PIX_FMT_GBRP9LE,
        AV_PIX_FMT_GBRP10BE,
        AV_PIX_FMT_GBRP10LE,
        AV_PIX_FMT_GBRP16BE,
        AV_PIX_FMT_GBRP16LE,
        AV_PIX_FMT_GBRAP,
        AV_PIX_FMT_GBRAP16BE,
        AV_PIX_FMT_GBRAP16LE,
        AV_PIX_FMT_GBRAP12BE,
        AV_PIX_FMT_GBRAP12LE,
        AV_PIX_FMT_GBRAP10BE,
        AV_PIX_FMT_GBRAP10LE
    };

    // Number of elements for an array
    const int image_grid_32 = ROUNDED_UP_DIV(outlink->h, 32) * ROUNDED_UP_DIV(outlink->w, 32);

    const int descriptor_buf_size = image_grid_32 * (BREIFN / 8);
    const int features_buf_size = image_grid_32 * sizeof(cl_float2);

    const AVHWFramesContext *hw_frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hw_frames_ctx->sw_format);

    av_assert0(hw_frames_ctx);
    av_assert0(desc);

    ff_framequeue_global_init(&fqg);
    ff_framequeue_init(&ctx->fq, &fqg);
    ctx->eof = 0;
    ctx->smooth_window = (int)(av_q2d(avctx->inputs[0]->frame_rate) * ctx->smooth_window_multiplier);
    ctx->curr_frame = 0;

    memset(&zeroed_ulong8, 0, sizeof(cl_ulong8));

    ctx->gauss_kernel = av_malloc_array(ctx->smooth_window, sizeof(float));
    if (!ctx->gauss_kernel) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    ctx->ransac_err = av_malloc_array(MATCHES_CONTIG_SIZE, sizeof(float));
    if (!ctx->ransac_err) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    for (int i = 0; i < RingbufCount; i++) {
        ctx->abs_motion.ringbuffers[i] = av_fifo_alloc2(ctx->smooth_window,
            sizeof(float), 0);

        if (!ctx->abs_motion.ringbuffers[i]) {
            err = AVERROR(ENOMEM);
            goto fail;
        }
    }

    if (ctx->debug_on) {
        ctx->abs_motion.debug_matches = av_fifo_alloc2(
            ctx->smooth_window / 2,
            sizeof(DebugMatches), 0
        );

        if (!ctx->abs_motion.debug_matches) {
            err = AVERROR(ENOMEM);
            goto fail;
        }
    }

    ctx->abs_motion.curr_frame_offset = 0;
    ctx->abs_motion.data_start_offset = -1;
    ctx->abs_motion.data_end_offset = -1;

    pattern_host = av_malloc_array(BREIFN, sizeof(PointPair));
    if (!pattern_host) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    ctx->matches_host = av_malloc_array(image_grid_32, sizeof(MotionVector));
    if (!ctx->matches_host) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    ctx->matches_contig_host = av_malloc_array(MATCHES_CONTIG_SIZE, sizeof(MotionVector));
    if (!ctx->matches_contig_host) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    ctx->inliers = av_malloc_array(MATCHES_CONTIG_SIZE, sizeof(MotionVector));
    if (!ctx->inliers) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    // Initializing the patch pattern for building BREIF descriptors with
    av_lfg_init(&ctx->alfg, 234342424);
    for (int i = 0; i < BREIFN; ++i) {
        PointPair pair;

        for (int j = 0; j < 2; ++j) {
            pair.p1.s[j] = rand_in(-BRIEF_PATCH_SIZE_HALF, BRIEF_PATCH_SIZE_HALF + 1, &ctx->alfg);
            pair.p2.s[j] = rand_in(-BRIEF_PATCH_SIZE_HALF, BRIEF_PATCH_SIZE_HALF + 1, &ctx->alfg);
        }

        pattern_host[i] = pair;
    }

    for (int i = 0; i < 14; i++) {
        if (ctx->sw_format == disallowed_formats[i]) {
            av_log(avctx, AV_LOG_ERROR, "unsupported format in deshake_opencl.\n");
            err = AVERROR(ENOSYS);
            goto fail;
        }
    }

    if (desc->flags & AV_PIX_FMT_FLAG_RGB) {
        ctx->is_yuv = 0;
    } else {
        ctx->is_yuv = 1;
    }
    ctx->sw_format = hw_frames_ctx->sw_format;

    err = ff_opencl_filter_load_program(avctx, &ff_source_deshake_cl, 1);
    if (err < 0)
        goto fail;

    if (ctx->debug_on) {
        queue_props = CL_QUEUE_PROFILING_ENABLE;
    } else {
        queue_props = 0;
    }
    ctx->command_queue = clCreateCommandQueue(
        ctx->ocf.hwctx->context,
        ctx->ocf.hwctx->device_id,
        queue_props,
        &cle
    );
    CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create OpenCL command queue %d.\n", cle);

    CL_CREATE_KERNEL(ctx, grayscale);
    CL_CREATE_KERNEL(ctx, harris_response);
    CL_CREATE_KERNEL(ctx, refine_features);
    CL_CREATE_KERNEL(ctx, brief_descriptors);
    CL_CREATE_KERNEL(ctx, match_descriptors);
    CL_CREATE_KERNEL(ctx, transform);
    CL_CREATE_KERNEL(ctx, crop_upscale);
    if (ctx->debug_on)
        CL_CREATE_KERNEL(ctx, draw_debug_info);

    if (!ctx->is_yuv) {
        grayscale_format.image_channel_order = CL_R;
        grayscale_format.image_channel_data_type = CL_FLOAT;

        grayscale_desc = (cl_image_desc) {
            .image_type = CL_MEM_OBJECT_IMAGE2D,
            .image_width = outlink->w,
            .image_height = outlink->h,
            .image_depth = 0,
            .image_array_size = 0,
            .image_row_pitch = 0,
            .image_slice_pitch = 0,
            .num_mip_levels = 0,
            .num_samples = 0,
            .buffer = NULL,
        };

        ctx->grayscale = clCreateImage(
            ctx->ocf.hwctx->context,
            0,
            &grayscale_format,
            &grayscale_desc,
            NULL,
            &cle
        );
        CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create grayscale image: %d.\n", cle);
    }

    CL_CREATE_BUFFER(ctx, harris_buf, outlink->h * outlink->w * sizeof(float));
    CL_CREATE_BUFFER(ctx, refined_features, features_buf_size);
    CL_CREATE_BUFFER(ctx, prev_refined_features, features_buf_size);
    CL_CREATE_BUFFER_FLAGS(
        ctx,
        brief_pattern,
        CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
        BREIFN * sizeof(PointPair),
        pattern_host
    );
    CL_CREATE_BUFFER(ctx, descriptors, descriptor_buf_size);
    CL_CREATE_BUFFER(ctx, prev_descriptors, descriptor_buf_size);
    CL_CREATE_BUFFER(ctx, matches, image_grid_32 * sizeof(MotionVector));
    CL_CREATE_BUFFER(ctx, matches_contig, MATCHES_CONTIG_SIZE * sizeof(MotionVector));
    CL_CREATE_BUFFER(ctx, transform_y, 9 * sizeof(float));
    CL_CREATE_BUFFER(ctx, transform_uv, 9 * sizeof(float));
    if (ctx->debug_on) {
        CL_CREATE_BUFFER(ctx, debug_matches, MATCHES_CONTIG_SIZE * sizeof(MotionVector));
        CL_CREATE_BUFFER(ctx, debug_model_matches, 3 * sizeof(MotionVector));
    }

    ctx->initialized = 1;
    av_freep(&pattern_host);

    return 0;

fail:
    av_freep(&pattern_host);
    return err;
}

// Logs debug information about the transform data
static void transform_debug(AVFilterContext *avctx, float *new_vals, float *old_vals, int curr_frame) {
    av_log(avctx, AV_LOG_VERBOSE,
        "Frame %d:\n"
        "\tframe moved from: %f x, %f y\n"
        "\t              to: %f x, %f y\n"
        "\t    rotated from: %f degrees\n"
        "\t              to: %f degrees\n"
        "\t     scaled from: %f x, %f y\n"
        "\t              to: %f x, %f y\n"
        "\n"
        "\tframe moved by: %f x, %f y\n"
        "\t    rotated by: %f degrees\n"
        "\t     scaled by: %f x, %f y\n",
        curr_frame,
        old_vals[RingbufX], old_vals[RingbufY],
        new_vals[RingbufX], new_vals[RingbufY],
        old_vals[RingbufRot] * (180.0 / M_PI),
        new_vals[RingbufRot] * (180.0 / M_PI),
        old_vals[RingbufScaleX], old_vals[RingbufScaleY],
        new_vals[RingbufScaleX], new_vals[RingbufScaleY],
        old_vals[RingbufX] - new_vals[RingbufX], old_vals[RingbufY] - new_vals[RingbufY],
        old_vals[RingbufRot] * (180.0 / M_PI) - new_vals[RingbufRot] * (180.0 / M_PI),
        new_vals[RingbufScaleX] / old_vals[RingbufScaleX], new_vals[RingbufScaleY] / old_vals[RingbufScaleY]
    );
}

// Uses the buffered motion information to determine a transform that smooths the
// given frame and applies it
static int filter_frame(AVFilterLink *link, AVFrame *input_frame)
{
    AVFilterContext *avctx = link->dst;
    AVFilterLink *outlink = avctx->outputs[0];
    DeshakeOpenCLContext *deshake_ctx = avctx->priv;
    AVFrame *cropped_frame = NULL, *transformed_frame = NULL;
    int err;
    cl_int cle;
    float new_vals[RingbufCount];
    float old_vals[RingbufCount];
    // Luma (in the case of YUV) transform, or just the transform in the case of RGB
    float transform_y[9];
    // Chroma transform
    float transform_uv[9];
    // Luma crop transform (or RGB)
    float transform_crop_y[9];
    // Chroma crop transform
    float transform_crop_uv[9];
    float transform_debug_rgb[9];
    size_t global_work[2];
    int64_t duration;
    cl_mem src, transformed, dst;
    cl_mem transforms[AV_VIDEO_MAX_PLANES];
    CropInfo crops[AV_VIDEO_MAX_PLANES];
    cl_event transform_event, crop_upscale_event;
    DebugMatches debug_matches;
    cl_int num_model_matches;

    const float center_w = (float)input_frame->width / 2;
    const float center_h = (float)input_frame->height / 2;

    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(deshake_ctx->sw_format);
    const int chroma_width  = AV_CEIL_RSHIFT(input_frame->width, desc->log2_chroma_w);
    const int chroma_height = AV_CEIL_RSHIFT(input_frame->height, desc->log2_chroma_h);

    const float center_w_chroma = (float)chroma_width / 2;
    const float center_h_chroma = (float)chroma_height / 2;

    const float luma_w_over_chroma_w = ((float)input_frame->width / (float)chroma_width);
    const float luma_h_over_chroma_h = ((float)input_frame->height / (float)chroma_height);

    if (deshake_ctx->debug_on) {
        av_fifo_read(
            deshake_ctx->abs_motion.debug_matches,
            &debug_matches, 1);
    }

    if (input_frame->duration) {
        duration = input_frame->duration;
    } else {
        duration = av_rescale_q(1, av_inv_q(outlink->frame_rate), outlink->time_base);
    }
    deshake_ctx->duration = input_frame->pts + duration;

    // Get the absolute transform data for this frame
    for (int i = 0; i < RingbufCount; i++) {
        av_fifo_peek(deshake_ctx->abs_motion.ringbuffers[i],
                     &old_vals[i], 1,
                     deshake_ctx->abs_motion.curr_frame_offset);
    }

    if (deshake_ctx->tripod_mode) {
        // If tripod mode is turned on we simply undo all motion relative to the
        // first frame

        new_vals[RingbufX] = 0.0f;
        new_vals[RingbufY] = 0.0f;
        new_vals[RingbufRot] = 0.0f;
        new_vals[RingbufScaleX] = 1.0f;
        new_vals[RingbufScaleY] = 1.0f;
    } else {
        // Tripod mode is off and we need to smooth a moving camera

        new_vals[RingbufX] = smooth(
            deshake_ctx,
            deshake_ctx->gauss_kernel,
            deshake_ctx->smooth_window,
            input_frame->width,
            deshake_ctx->abs_motion.ringbuffers[RingbufX]
        );
        new_vals[RingbufY] = smooth(
            deshake_ctx,
            deshake_ctx->gauss_kernel,
            deshake_ctx->smooth_window,
            input_frame->height,
            deshake_ctx->abs_motion.ringbuffers[RingbufY]
        );
        new_vals[RingbufRot] = smooth(
            deshake_ctx,
            deshake_ctx->gauss_kernel,
            deshake_ctx->smooth_window,
            M_PI / 4,
            deshake_ctx->abs_motion.ringbuffers[RingbufRot]
        );
        new_vals[RingbufScaleX] = smooth(
            deshake_ctx,
            deshake_ctx->gauss_kernel,
            deshake_ctx->smooth_window,
            2.0f,
            deshake_ctx->abs_motion.ringbuffers[RingbufScaleX]
        );
        new_vals[RingbufScaleY] = smooth(
            deshake_ctx,
            deshake_ctx->gauss_kernel,
            deshake_ctx->smooth_window,
            2.0f,
            deshake_ctx->abs_motion.ringbuffers[RingbufScaleY]
        );
    }

    transform_center_scale(
        old_vals[RingbufX] - new_vals[RingbufX],
        old_vals[RingbufY] - new_vals[RingbufY],
        old_vals[RingbufRot] - new_vals[RingbufRot],
        new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
        new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
        center_w,
        center_h,
        transform_y
    );

    transform_center_scale(
        (old_vals[RingbufX] - new_vals[RingbufX]) / luma_w_over_chroma_w,
        (old_vals[RingbufY] - new_vals[RingbufY]) / luma_h_over_chroma_h,
        old_vals[RingbufRot] - new_vals[RingbufRot],
        new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
        new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
        center_w_chroma,
        center_h_chroma,
        transform_uv
    );

    CL_BLOCKING_WRITE_BUFFER(deshake_ctx->command_queue, deshake_ctx->transform_y, 9 * sizeof(float), transform_y, NULL);
    CL_BLOCKING_WRITE_BUFFER(deshake_ctx->command_queue, deshake_ctx->transform_uv, 9 * sizeof(float), transform_uv, NULL);

    if (deshake_ctx->debug_on)
        transform_debug(avctx, new_vals, old_vals, deshake_ctx->curr_frame);

    cropped_frame = ff_get_video_buffer(outlink, outlink->w, outlink->h);
    if (!cropped_frame) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    transformed_frame = ff_get_video_buffer(outlink, outlink->w, outlink->h);
    if (!transformed_frame) {
        err = AVERROR(ENOMEM);
        goto fail;
    }

    transforms[0] = deshake_ctx->transform_y;
    transforms[1] = transforms[2] = deshake_ctx->transform_uv;

    for (int p = 0; p < AV_VIDEO_MAX_PLANES; p++) {
        // Transform all of the planes appropriately
        src = (cl_mem)input_frame->data[p];
        transformed = (cl_mem)transformed_frame->data[p];

        if (!transformed)
            break;

        err = ff_opencl_filter_work_size_from_image(avctx, global_work, input_frame, p, 0);
        if (err < 0)
            goto fail;

        CL_RUN_KERNEL_WITH_ARGS(
            deshake_ctx->command_queue,
            deshake_ctx->kernel_transform,
            global_work,
            NULL,
            &transform_event,
            { sizeof(cl_mem), &src },
            { sizeof(cl_mem), &transformed },
            { sizeof(cl_mem), &transforms[p] },
        );
    }

    if (deshake_ctx->debug_on && !deshake_ctx->is_yuv && debug_matches.num_matches > 0) {
        CL_BLOCKING_WRITE_BUFFER(
            deshake_ctx->command_queue,
            deshake_ctx->debug_matches,
            debug_matches.num_matches * sizeof(MotionVector),
            debug_matches.matches,
            NULL
        );

        CL_BLOCKING_WRITE_BUFFER(
            deshake_ctx->command_queue,
            deshake_ctx->debug_model_matches,
            debug_matches.num_model_matches * sizeof(MotionVector),
            debug_matches.model_matches,
            NULL
        );

        num_model_matches = debug_matches.num_model_matches;

        // Invert the transform
        transform_center_scale(
            new_vals[RingbufX] - old_vals[RingbufX],
            new_vals[RingbufY] - old_vals[RingbufY],
            new_vals[RingbufRot] - old_vals[RingbufRot],
            old_vals[RingbufScaleX] / new_vals[RingbufScaleX],
            old_vals[RingbufScaleY] / new_vals[RingbufScaleY],
            center_w,
            center_h,
            transform_debug_rgb
        );

        CL_BLOCKING_WRITE_BUFFER(deshake_ctx->command_queue, deshake_ctx->transform_y, 9 * sizeof(float), transform_debug_rgb, NULL);

        transformed = (cl_mem)transformed_frame->data[0];
        CL_RUN_KERNEL_WITH_ARGS(
            deshake_ctx->command_queue,
            deshake_ctx->kernel_draw_debug_info,
            (size_t[]){ debug_matches.num_matches },
            NULL,
            NULL,
            { sizeof(cl_mem), &transformed },
            { sizeof(cl_mem), &deshake_ctx->debug_matches },
            { sizeof(cl_mem), &deshake_ctx->debug_model_matches },
            { sizeof(cl_int), &num_model_matches },
            { sizeof(cl_mem), &deshake_ctx->transform_y }
        );
    }

    if (deshake_ctx->should_crop) {
        // Generate transforms for cropping
        transform_center_scale(
            (old_vals[RingbufX] - new_vals[RingbufX]) / 5,
            (old_vals[RingbufY] - new_vals[RingbufY]) / 5,
            (old_vals[RingbufRot] - new_vals[RingbufRot]) / 5,
            new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
            new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
            center_w,
            center_h,
            transform_crop_y
        );
        update_needed_crop(&deshake_ctx->crop_y, transform_crop_y, input_frame->width, input_frame->height);

        transform_center_scale(
            (old_vals[RingbufX] - new_vals[RingbufX]) / (5 * luma_w_over_chroma_w),
            (old_vals[RingbufY] - new_vals[RingbufY]) / (5 * luma_h_over_chroma_h),
            (old_vals[RingbufRot] - new_vals[RingbufRot]) / 5,
            new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
            new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
            center_w_chroma,
            center_h_chroma,
            transform_crop_uv
        );
        update_needed_crop(&deshake_ctx->crop_uv, transform_crop_uv, chroma_width, chroma_height);

        crops[0] = deshake_ctx->crop_y;
        crops[1] = crops[2] = deshake_ctx->crop_uv;

        for (int p = 0; p < AV_VIDEO_MAX_PLANES; p++) {
            // Crop all of the planes appropriately
            dst = (cl_mem)cropped_frame->data[p];
            transformed = (cl_mem)transformed_frame->data[p];

            if (!dst)
                break;

            err = ff_opencl_filter_work_size_from_image(avctx, global_work, input_frame, p, 0);
            if (err < 0)
                goto fail;

            CL_RUN_KERNEL_WITH_ARGS(
                deshake_ctx->command_queue,
                deshake_ctx->kernel_crop_upscale,
                global_work,
                NULL,
                &crop_upscale_event,
                { sizeof(cl_mem), &transformed },
                { sizeof(cl_mem), &dst },
                { sizeof(cl_float2), &crops[p].top_left },
                { sizeof(cl_float2), &crops[p].bottom_right },
            );
        }
    }

    if (deshake_ctx->curr_frame < deshake_ctx->smooth_window / 2) {
        // This means we are somewhere at the start of the video. We need to
        // increment the current frame offset until it reaches the center of
        // the ringbuffers (as the current frame will be located there for
        // the rest of the video).
        //
        // The end of the video is taken care of by draining motion data
        // one-by-one out of the buffer, causing the (at that point fixed)
        // offset to move towards later frames' data.
        ++deshake_ctx->abs_motion.curr_frame_offset;
    }

    if (deshake_ctx->abs_motion.data_end_offset != -1) {
        // Keep the end offset in sync with the frame it's supposed to be
        // positioned at
        --deshake_ctx->abs_motion.data_end_offset;

        if (deshake_ctx->abs_motion.data_end_offset == deshake_ctx->abs_motion.curr_frame_offset - 1) {
            // The end offset would be the start of the new video sequence; flip to
            // start offset
            deshake_ctx->abs_motion.data_end_offset = -1;
            deshake_ctx->abs_motion.data_start_offset = deshake_ctx->abs_motion.curr_frame_offset;
        }
    } else if (deshake_ctx->abs_motion.data_start_offset != -1) {
        // Keep the start offset in sync with the frame it's supposed to be
        // positioned at
        --deshake_ctx->abs_motion.data_start_offset;
    }

    if (deshake_ctx->debug_on) {
        deshake_ctx->transform_time += ff_opencl_get_event_time(transform_event);
        if (deshake_ctx->should_crop) {
            deshake_ctx->crop_upscale_time += ff_opencl_get_event_time(crop_upscale_event);
        }
    }

    ++deshake_ctx->curr_frame;

    if (deshake_ctx->debug_on)
        av_freep(&debug_matches.matches);

    if (deshake_ctx->should_crop) {
        err = av_frame_copy_props(cropped_frame, input_frame);
        if (err < 0)
            goto fail;

        av_frame_free(&transformed_frame);
        av_frame_free(&input_frame);
        return ff_filter_frame(outlink, cropped_frame);

    } else {
        err = av_frame_copy_props(transformed_frame, input_frame);
        if (err < 0)
            goto fail;

        av_frame_free(&cropped_frame);
        av_frame_free(&input_frame);
        return ff_filter_frame(outlink, transformed_frame);
    }

fail:
    clFinish(deshake_ctx->command_queue);

    if (deshake_ctx->debug_on)
        if (debug_matches.matches)
            av_freep(&debug_matches.matches);

    av_frame_free(&input_frame);
    av_frame_free(&transformed_frame);
    av_frame_free(&cropped_frame);
    return err;
}

// Add the given frame to the frame queue to eventually be processed.
//
// Also determines the motion from the previous frame and updates the stored
// motion information accordingly.
static int queue_frame(AVFilterLink *link, AVFrame *input_frame)
{
    AVFilterContext *avctx = link->dst;
    DeshakeOpenCLContext *deshake_ctx = avctx->priv;
    int err;
    int num_vectors;
    int num_inliers = 0;
    cl_int cle;
    FrameDelta relative;
    SimilarityMatrix model;
    size_t global_work[2];
    size_t harris_global_work[2];
    size_t grid_32_global_work[2];
    int grid_32_h, grid_32_w;
    size_t local_work[2];
    cl_mem src, temp;
    float prev_vals[5];
    float new_vals[5];
    cl_event grayscale_event, harris_response_event, refine_features_event,
             brief_event, match_descriptors_event, read_buf_event;
    DebugMatches debug_matches;

    num_vectors = 0;

    local_work[0] = 8;
    local_work[1] = 8;

    err = ff_opencl_filter_work_size_from_image(avctx, global_work, input_frame, 0, 0);
    if (err < 0)
        goto fail;

    err = ff_opencl_filter_work_size_from_image(avctx, harris_global_work, input_frame, 0, 8);
    if (err < 0)
        goto fail;

    err = ff_opencl_filter_work_size_from_image(avctx, grid_32_global_work, input_frame, 0, 32);
    if (err < 0)
        goto fail;

    // We want a single work-item for each 32x32 block of pixels in the input frame
    grid_32_global_work[0] /= 32;
    grid_32_global_work[1] /= 32;

    grid_32_h = ROUNDED_UP_DIV(input_frame->height, 32);
    grid_32_w = ROUNDED_UP_DIV(input_frame->width, 32);

    if (deshake_ctx->is_yuv) {
        deshake_ctx->grayscale = (cl_mem)input_frame->data[0];
    } else {
        src = (cl_mem)input_frame->data[0];

        CL_RUN_KERNEL_WITH_ARGS(
            deshake_ctx->command_queue,
            deshake_ctx->kernel_grayscale,
            global_work,
            NULL,
            &grayscale_event,
            { sizeof(cl_mem), &src },
            { sizeof(cl_mem), &deshake_ctx->grayscale }
        );
    }

    CL_RUN_KERNEL_WITH_ARGS(
        deshake_ctx->command_queue,
        deshake_ctx->kernel_harris_response,
        harris_global_work,
        local_work,
        &harris_response_event,
        { sizeof(cl_mem), &deshake_ctx->grayscale },
        { sizeof(cl_mem), &deshake_ctx->harris_buf }
    );

    CL_RUN_KERNEL_WITH_ARGS(
        deshake_ctx->command_queue,
        deshake_ctx->kernel_refine_features,
        grid_32_global_work,
        NULL,
        &refine_features_event,
        { sizeof(cl_mem), &deshake_ctx->grayscale },
        { sizeof(cl_mem), &deshake_ctx->harris_buf },
        { sizeof(cl_mem), &deshake_ctx->refined_features },
        { sizeof(cl_int), &deshake_ctx->refine_features }
    );

    CL_RUN_KERNEL_WITH_ARGS(
        deshake_ctx->command_queue,
        deshake_ctx->kernel_brief_descriptors,
        grid_32_global_work,
        NULL,
        &brief_event,
        { sizeof(cl_mem), &deshake_ctx->grayscale },
        { sizeof(cl_mem), &deshake_ctx->refined_features },
        { sizeof(cl_mem), &deshake_ctx->descriptors },
        { sizeof(cl_mem), &deshake_ctx->brief_pattern}
    );

    if (!av_fifo_can_read(deshake_ctx->abs_motion.ringbuffers[RingbufX])) {
        // This is the first frame we've been given to queue, meaning there is
        // no previous frame to match descriptors to

        goto no_motion_data;
    }

    CL_RUN_KERNEL_WITH_ARGS(
        deshake_ctx->command_queue,
        deshake_ctx->kernel_match_descriptors,
        grid_32_global_work,
        NULL,
        &match_descriptors_event,
        { sizeof(cl_mem), &deshake_ctx->prev_refined_features },
        { sizeof(cl_mem), &deshake_ctx->refined_features },
        { sizeof(cl_mem), &deshake_ctx->descriptors },
        { sizeof(cl_mem), &deshake_ctx->prev_descriptors },
        { sizeof(cl_mem), &deshake_ctx->matches }
    );

    cle = clEnqueueReadBuffer(
        deshake_ctx->command_queue,
        deshake_ctx->matches,
        CL_TRUE,
        0,
        grid_32_h * grid_32_w * sizeof(MotionVector),
        deshake_ctx->matches_host,
        0,
        NULL,
        &read_buf_event
    );
    CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to read matches to host: %d.\n", cle);

    num_vectors = make_vectors_contig(deshake_ctx, grid_32_h, grid_32_w);

    if (num_vectors < 10) {
        // Not enough matches to get reliable motion data for this frame
        //
        // From this point on all data is relative to this frame rather than the
        // original frame. We have to make sure that we don't mix values that were
        // relative to the original frame with the new values relative to this
        // frame when doing the gaussian smoothing. We keep track of where the old
        // values end using this data_end_offset field in order to accomplish
        // that goal.
        //
        // If no motion data is present for multiple frames in a short window of
        // time, we leave the end where it was to avoid mixing 0s in with the
        // old data (and just treat them all as part of the new values)
        if (deshake_ctx->abs_motion.data_end_offset == -1) {
            deshake_ctx->abs_motion.data_end_offset =
                av_fifo_can_read(deshake_ctx->abs_motion.ringbuffers[RingbufX]) - 1;
        }

        goto no_motion_data;
    }

    if (!estimate_affine_2d(
        deshake_ctx,
        deshake_ctx->matches_contig_host,
        &debug_matches,
        num_vectors,
        model.matrix,
        10.0,
        3000,
        0.999999999999
    )) {
        goto no_motion_data;
    }

    for (int i = 0; i < num_vectors; i++) {
        if (deshake_ctx->matches_contig_host[i].should_consider) {
            deshake_ctx->inliers[num_inliers] = deshake_ctx->matches_contig_host[i];
            num_inliers++;
        }
    }

    if (!minimize_error(
        deshake_ctx,
        deshake_ctx->inliers,
        &debug_matches,
        num_inliers,
        model.matrix,
        400
    )) {
        goto no_motion_data;
    }


    relative = decompose_transform(model.matrix);

    // Get the absolute transform data for the previous frame
    for (int i = 0; i < RingbufCount; i++) {
        av_fifo_peek(
            deshake_ctx->abs_motion.ringbuffers[i],
            &prev_vals[i], 1,
            av_fifo_can_read(deshake_ctx->abs_motion.ringbuffers[i]) - 1);
    }

    new_vals[RingbufX]      = prev_vals[RingbufX] + relative.translation.s[0];
    new_vals[RingbufY]      = prev_vals[RingbufY] + relative.translation.s[1];
    new_vals[RingbufRot]    = prev_vals[RingbufRot] + relative.rotation;
    new_vals[RingbufScaleX] = prev_vals[RingbufScaleX] / relative.scale.s[0];
    new_vals[RingbufScaleY] = prev_vals[RingbufScaleY] / relative.scale.s[1];

    if (deshake_ctx->debug_on) {
        if (!deshake_ctx->is_yuv) {
            deshake_ctx->grayscale_time     += ff_opencl_get_event_time(grayscale_event);
        }
        deshake_ctx->harris_response_time   += ff_opencl_get_event_time(harris_response_event);
        deshake_ctx->refine_features_time   += ff_opencl_get_event_time(refine_features_event);
        deshake_ctx->brief_descriptors_time += ff_opencl_get_event_time(brief_event);
        deshake_ctx->match_descriptors_time += ff_opencl_get_event_time(match_descriptors_event);
        deshake_ctx->read_buf_time          += ff_opencl_get_event_time(read_buf_event);
    }

    goto end;

no_motion_data:
    new_vals[RingbufX]      = 0.0f;
    new_vals[RingbufY]      = 0.0f;
    new_vals[RingbufRot]    = 0.0f;
    new_vals[RingbufScaleX] = 1.0f;
    new_vals[RingbufScaleY] = 1.0f;

    for (int i = 0; i < num_vectors; i++) {
        deshake_ctx->matches_contig_host[i].should_consider = 0;
    }
    debug_matches.num_model_matches = 0;

    if (deshake_ctx->debug_on) {
        av_log(avctx, AV_LOG_VERBOSE,
            "\n[ALERT] No motion data found in queue_frame, motion reset to 0\n\n"
        );
    }

    goto end;

end:
    // Swap the descriptor buffers (we don't need the previous frame's descriptors
    // again so we will use that space for the next frame's descriptors)
    temp = deshake_ctx->prev_descriptors;
    deshake_ctx->prev_descriptors = deshake_ctx->descriptors;
    deshake_ctx->descriptors = temp;

    // Same for the refined features
    temp = deshake_ctx->prev_refined_features;
    deshake_ctx->prev_refined_features = deshake_ctx->refined_features;
    deshake_ctx->refined_features = temp;

    if (deshake_ctx->debug_on) {
        if (num_vectors == 0) {
            debug_matches.matches = NULL;
        } else {
            debug_matches.matches = av_malloc_array(num_vectors, sizeof(MotionVector));

            if (!debug_matches.matches) {
                err = AVERROR(ENOMEM);
                goto fail;
            }
        }

        for (int i = 0; i < num_vectors; i++) {
            debug_matches.matches[i] = deshake_ctx->matches_contig_host[i];
        }
        debug_matches.num_matches = num_vectors;

        av_fifo_write(
            deshake_ctx->abs_motion.debug_matches,
            &debug_matches, 1);
    }

    for (int i = 0; i < RingbufCount; i++) {
        av_fifo_write(deshake_ctx->abs_motion.ringbuffers[i], &new_vals[i], 1);
    }

    return ff_framequeue_add(&deshake_ctx->fq, input_frame);

fail:
    clFinish(deshake_ctx->command_queue);
    av_frame_free(&input_frame);
    return err;
}

static int activate(AVFilterContext *ctx)
{
    AVFilterLink *inlink = ctx->inputs[0];
    AVFilterLink *outlink = ctx->outputs[0];
    DeshakeOpenCLContext *deshake_ctx = ctx->priv;
    AVFrame *frame = NULL;
    int ret, status;
    int64_t pts;

    FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);

    if (!deshake_ctx->eof) {
        ret = ff_inlink_consume_frame(inlink, &frame);
        if (ret < 0)
            return ret;
        if (ret > 0) {
            if (!frame->hw_frames_ctx)
                return AVERROR(EINVAL);

            if (!deshake_ctx->initialized) {
                ret = deshake_opencl_init(ctx);
                if (ret < 0)
                    return ret;
            }

            // If there is no more space in the ringbuffers, remove the oldest
            // values to make room for the new ones
            if (!av_fifo_can_write(deshake_ctx->abs_motion.ringbuffers[RingbufX])) {
                for (int i = 0; i < RingbufCount; i++) {
                    av_fifo_drain2(deshake_ctx->abs_motion.ringbuffers[i], 1);
                }
            }
            ret = queue_frame(inlink, frame);
            if (ret < 0)
                return ret;
            if (ret >= 0) {
                // See if we have enough buffered frames to process one
                //
                // "enough" is half the smooth window of queued frames into the future
                if (ff_framequeue_queued_frames(&deshake_ctx->fq) >= deshake_ctx->smooth_window / 2) {
                    return filter_frame(inlink, ff_framequeue_take(&deshake_ctx->fq));
                }
            }
        }
    }

    if (!deshake_ctx->eof && ff_inlink_acknowledge_status(inlink, &status, &pts)) {
        if (status == AVERROR_EOF) {
            deshake_ctx->eof = 1;
        }
    }

    if (deshake_ctx->eof) {
        // Finish processing the rest of the frames in the queue.
        while(ff_framequeue_queued_frames(&deshake_ctx->fq) != 0) {
            for (int i = 0; i < RingbufCount; i++) {
                av_fifo_drain2(deshake_ctx->abs_motion.ringbuffers[i], 1);
            }

            ret = filter_frame(inlink, ff_framequeue_take(&deshake_ctx->fq));
            if (ret < 0) {
                return ret;
            }
        }

        if (deshake_ctx->debug_on) {
            av_log(ctx, AV_LOG_VERBOSE,
                "Average kernel execution times:\n"
                "\t        grayscale: %0.3f ms\n"
                "\t  harris_response: %0.3f ms\n"
                "\t  refine_features: %0.3f ms\n"
                "\tbrief_descriptors: %0.3f ms\n"
                "\tmatch_descriptors: %0.3f ms\n"
                "\t        transform: %0.3f ms\n"
                "\t     crop_upscale: %0.3f ms\n"
                "Average buffer read times:\n"
                "\t     features buf: %0.3f ms\n",
                averaged_event_time_ms(deshake_ctx->grayscale_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->harris_response_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->refine_features_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->brief_descriptors_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->match_descriptors_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->transform_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->crop_upscale_time, deshake_ctx->curr_frame),
                averaged_event_time_ms(deshake_ctx->read_buf_time, deshake_ctx->curr_frame)
            );
        }

        ff_outlink_set_status(outlink, AVERROR_EOF, deshake_ctx->duration);
        return 0;
    }

    if (!deshake_ctx->eof) {
        FF_FILTER_FORWARD_WANTED(outlink, inlink);
    }

    return FFERROR_NOT_READY;
}

static const AVFilterPad deshake_opencl_inputs[] = {
    {
        .name = "default",
        .type = AVMEDIA_TYPE_VIDEO,
        .config_props = &ff_opencl_filter_config_input,
    },
};

static const AVFilterPad deshake_opencl_outputs[] = {
    {
        .name = "default",
        .type = AVMEDIA_TYPE_VIDEO,
        .config_props = &ff_opencl_filter_config_output,
    },
};

#define OFFSET(x) offsetof(DeshakeOpenCLContext, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM

static const AVOption deshake_opencl_options[] = {
    {
        "tripod", "simulates a tripod by preventing any camera movement whatsoever "
        "from the original frame",
        OFFSET(tripod_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS
    },
    {
        "debug", "turn on additional debugging information",
        OFFSET(debug_on), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS
    },
    {
        "adaptive_crop", "attempt to subtly crop borders to reduce mirrored content",
        OFFSET(should_crop), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, FLAGS
    },
    {
        "refine_features", "refine feature point locations at a sub-pixel level",
        OFFSET(refine_features), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, FLAGS
    },
    {
        "smooth_strength", "smoothing strength (0 attempts to adaptively determine optimal strength)",
        OFFSET(smooth_percent), AV_OPT_TYPE_FLOAT, {.dbl = 0.0f}, 0.0f, 1.0f, FLAGS
    },
    {
        "smooth_window_multiplier", "multiplier for number of frames to buffer for motion data",
        OFFSET(smooth_window_multiplier), AV_OPT_TYPE_FLOAT, {.dbl = 2.0}, 0.1, 10.0, FLAGS
    },
    { NULL }
};

AVFILTER_DEFINE_CLASS(deshake_opencl);

const AVFilter ff_vf_deshake_opencl = {
    .name           = "deshake_opencl",
    .description    = NULL_IF_CONFIG_SMALL("Feature-point based video stabilization filter"),
    .priv_size      = sizeof(DeshakeOpenCLContext),
    .priv_class     = &deshake_opencl_class,
    .init           = &ff_opencl_filter_init,
    .uninit         = &deshake_opencl_uninit,
    .activate       = activate,
    FILTER_INPUTS(deshake_opencl_inputs),
    FILTER_OUTPUTS(deshake_opencl_outputs),
    FILTER_SINGLE_PIXFMT(AV_PIX_FMT_OPENCL),
    .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
    .flags          = AVFILTER_FLAG_HWDEVICE,
};