[FFmpeg-devel] [PATCH] libavfilter: Add OpenCL convolution filter v0.2
Mark Thompson
sw at jkqxz.net
Fri Mar 23 17:06:55 EET 2018
On 22/03/18 22:52, Danil Iashchenko wrote:
> Hi there!
> Thank you for advices, I have fixed the problem when per plane matrices
> application was incorrect. Now it works as expected and behaves like the existing vf_convolution filter.
>
> Tested for yuv and nv12 formats.
> The following filters from ffmpeg documentation
> (https://ffmpeg.org/ffmpeg-filters.html#Examples-43) work correctly:
>
> * sharpen
> * blur
> * edge enchance
> * edge detect
> * laplacian edge detector which includes diagonals
> * emboss
> * custom tests
>
> I have two questions:
>
> 1. Looks like existing convolution filter implementaion ignores
> 0rdiv and 0bias parameters for 0m='0 0 0 0 1 0 0 0 0' matrix,
> so results will be the same for following sample filter runs:
>
> -vf convolution='0m=0 0 0 0 1 0 0 0 0:0rdiv=1:0bias=20'
> -vf convolution='0m=0 0 0 0 1 0 0 0 0:0rdiv=1/3:0bias=0'
> -vf convolution='0m=0 0 0 0 1 0 0 0 0:0rdiv=1/2:0bias=100'
> -vf convolution='0m=0 0 0 0 1 0 0 0 0:0rdiv=1/5:0bias=50'
>
> and will not differ from
>
> -vf convolution='0m=0 0 0 0 1 0 0 0 0:0rdiv=1:0bias=0'
>
> My implementation does not ignore 0rdiv and 0bias parameters in case of m0='0 0 0 0 1 0 0 0 0',
> and results will differ from the existing implementation so I do not know
> which one is correct: mine or exisiting.
Good question, I don't know. I feel like your answer makes more sense, but it's a pretty degenerate use of the filter so maybe it doesn't matter? Would anyone more familiar with this (Paul?) like to offer an opinion?
> 2. I have a local kernel (not included in this patch), but there is no
> significant speed difference comparing to global kernel. Shall I include
> it in the next patch?
Even with 7x7 matrices, which load every value many more times than 3x3?
How about making a two patch series, one without it and then a second patch on top of it adding the local kernel? It would be useful to see it and I may be able to test on more devices, if it's not useful then we can just not apply the second patch.
Some more review comments below.
Thanks,
- Mark
> ---
> configure | 1 +
> libavfilter/Makefile | 1 +
> libavfilter/allfilters.c | 1 +
> libavfilter/opencl/convolution.cl | 40 ++++
> libavfilter/opencl_source.h | 1 +
> libavfilter/vf_convolution_opencl.c | 365 ++++++++++++++++++++++++++++++++++++
> 6 files changed, 409 insertions(+)
> create mode 100644 libavfilter/opencl/convolution.cl
> create mode 100644 libavfilter/vf_convolution_opencl.c
>
> diff --git a/configure b/configure
> index 6916b45..bf5c312 100755
> --- a/configure
> +++ b/configure
> @@ -3210,6 +3210,7 @@ blackframe_filter_deps="gpl"
> boxblur_filter_deps="gpl"
> bs2b_filter_deps="libbs2b"
> colormatrix_filter_deps="gpl"
> +convolution_opencl_filter_deps="opencl"
> convolve_filter_deps="avcodec"
> convolve_filter_select="fft"
> coreimage_filter_deps="coreimage appkit"
> diff --git a/libavfilter/Makefile b/libavfilter/Makefile
> index 6a60836..d005934 100644
> --- a/libavfilter/Makefile
> +++ b/libavfilter/Makefile
> @@ -156,6 +156,7 @@ OBJS-$(CONFIG_COLORLEVELS_FILTER) += vf_colorlevels.o
> OBJS-$(CONFIG_COLORMATRIX_FILTER) += vf_colormatrix.o
> OBJS-$(CONFIG_COLORSPACE_FILTER) += vf_colorspace.o colorspacedsp.o
> OBJS-$(CONFIG_CONVOLUTION_FILTER) += vf_convolution.o
> +OBJS-$(CONFIG_CONVOLUTION_OPENCL_FILTER) += vf_convolution_opencl.o opencl.o opencl/convolution.o
> OBJS-$(CONFIG_CONVOLVE_FILTER) += vf_convolve.o framesync.o
> OBJS-$(CONFIG_COPY_FILTER) += vf_copy.o
> OBJS-$(CONFIG_COREIMAGE_FILTER) += vf_coreimage.o
> diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
> index 9adb109..f2dc55e 100644
> --- a/libavfilter/allfilters.c
> +++ b/libavfilter/allfilters.c
> @@ -166,6 +166,7 @@ static void register_all(void)
> REGISTER_FILTER(COLORMATRIX, colormatrix, vf);
> REGISTER_FILTER(COLORSPACE, colorspace, vf);
> REGISTER_FILTER(CONVOLUTION, convolution, vf);
> + REGISTER_FILTER(CONVOLUTION_OPENCL, convolution_opencl, vf);
> REGISTER_FILTER(CONVOLVE, convolve, vf);
> REGISTER_FILTER(COPY, copy, vf);
> REGISTER_FILTER(COREIMAGE, coreimage, vf);
> diff --git a/libavfilter/opencl/convolution.cl b/libavfilter/opencl/convolution.cl
> new file mode 100644
> index 0000000..aa1db97
> --- /dev/null
> +++ b/libavfilter/opencl/convolution.cl
> @@ -0,0 +1,40 @@
> +/*
> + * 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
> + */
> +
> +__kernel void convolution_global(__write_only image2d_t dst,
> + __read_only image2d_t src,
> + int coef_matrix_dim,
> + __constant float *coef_matrix,
> + float div,
> + float bias)
> +{
> + const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST);
> +
> + const int half_matrix_dim = (coef_matrix_dim / 2);
> + int2 loc = (int2)(get_global_id(0), get_global_id(1));
> + float4 convPix = (float4)(0.0f, 0.0f, 0.0f, 0.0f);
> +
> + for (int conv_i = -half_matrix_dim; conv_i <= half_matrix_dim; conv_i++) {
> + for (int conv_j = -half_matrix_dim; conv_j <= half_matrix_dim; conv_j++) {
> + float4 px = read_imagef(src, sampler, loc + (int2)(conv_j, conv_i));
> + convPix += px * coef_matrix[(conv_i+half_matrix_dim)*coef_matrix_dim+(conv_j+half_matrix_dim)];
> + }
> + }
> + float4 dstPix = convPix * div + bias;
> + write_imagef(dst, loc, dstPix);
> +}
> diff --git a/libavfilter/opencl_source.h b/libavfilter/opencl_source.h
> index 23cdfc6..158fb3e 100644
> --- a/libavfilter/opencl_source.h
> +++ b/libavfilter/opencl_source.h
> @@ -19,6 +19,7 @@
> #ifndef AVFILTER_OPENCL_SOURCE_H
> #define AVFILTER_OPENCL_SOURCE_H
>
> +extern const char *ff_opencl_source_convolution;
> extern const char *ff_opencl_source_overlay;
> extern const char *ff_opencl_source_unsharp;
>
> diff --git a/libavfilter/vf_convolution_opencl.c b/libavfilter/vf_convolution_opencl.c
> new file mode 100644
> index 0000000..b788033
> --- /dev/null
> +++ b/libavfilter/vf_convolution_opencl.c
> @@ -0,0 +1,365 @@
> +/*
> + * 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
> + */
> +
> +#include "libavutil/common.h"
> +#include "libavutil/imgutils.h"
> +#include "libavutil/mem.h"
> +#include "libavutil/opt.h"
> +#include "libavutil/pixdesc.h"
> +#include "libavutil/avstring.h"
> +
> +
> +#include "avfilter.h"
> +#include "internal.h"
> +#include "opencl.h"
> +#include "opencl_source.h"
> +#include "video.h"
> +
> +typedef struct ConvolutionOpenCLContext {
> + OpenCLFilterContext ocf;
> +
> + int initialised;
> + cl_kernel kernel;
> + cl_command_queue command_queue;
> +
> + char *matrix_str[4];
> +
> + cl_mem matrix[4];
> + cl_int matrix_sizes[4];
> + cl_int dims[4];
> + cl_float rdivs[4];
> + cl_float biases[4];
> +
> +} ConvolutionOpenCLContext;
> +
> +
> +static int convolution_opencl_init(AVFilterContext *avctx)
> +{
> + ConvolutionOpenCLContext *ctx = avctx->priv;
> + cl_int cle;
> + int err;
> +
> + err = ff_opencl_filter_load_program(avctx, &ff_opencl_source_convolution, 1);
> + if (err < 0)
> + goto fail;
> +
> + ctx->command_queue = clCreateCommandQueue(ctx->ocf.hwctx->context,
> + ctx->ocf.hwctx->device_id,
> + 0, &cle);
> + if (!ctx->command_queue) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to create OpenCL "
> + "command queue: %d.\n", cle);
> + err = AVERROR(EIO);
> + goto fail;
> + }
> +
> + ctx->kernel = clCreateKernel(ctx->ocf.program, "convolution_global", &cle);
> + if (!ctx->kernel) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to create kernel: %d.\n", cle);
> + err = AVERROR(EIO);
> + goto fail;
> + }
> +
> + ctx->initialised = 1;
> + return 0;
> +
> +fail:
> + if (ctx->command_queue)
> + clReleaseCommandQueue(ctx->command_queue);
> + if (ctx->kernel)
> + clReleaseKernel(ctx->kernel);
> + return err;
> +}
> +
> +
> +
> +static int convolution_opencl_make_filter_params(AVFilterContext *avctx)
> +{
> + ConvolutionOpenCLContext *ctx = avctx->priv;
> + float *matrix;
> + size_t matrix_bytes;
> + cl_mem buffer;
> + cl_int cle;
> + int i, j;
> + int err, sscanf_err;
> + char *p, *arg, *saveptr = NULL;
> + float input_matrix[4][49];
> +
> + for (i = 0; i < 4; i++) {
> + ctx->biases[i] = ctx->biases[i] / 255;
I think prefer writing 255.0 to make it obvious that this is floating-point division.
> + }
> +
> + for (i = 0; i < 4; i++) {
> + p = ctx->matrix_str[i];
> + while (ctx->matrix_sizes[i] < 49) {
> + arg = av_strtok(p, " ", &saveptr);
> + if (!arg) {
> + break;
> + }
> + p = NULL;
> + sscanf_err = sscanf(arg, "%f", &input_matrix[i][ctx->matrix_sizes[i]]);
> + if (sscanf_err != 1) {
> + av_log(ctx, AV_LOG_ERROR, "Matrix is sequence of 9, 25 or 49 signed numbers\n");
> + return AVERROR(EINVAL);
> + }
> + ctx->matrix_sizes[i]++;
> + }
> + if (ctx->matrix_sizes[i] == 9) {
> + ctx->dims[i] = 3;
> + } else if (ctx->matrix_sizes[i] == 25) {
> + ctx->dims[i] = 5;
> + } else if (ctx->matrix_sizes[i] == 49) {
> + ctx->dims[i] = 7;
> + } else {
> + av_log(ctx, AV_LOG_ERROR, "Invalid matrix size:%d\n", ctx->matrix_sizes[i]);
> + return AVERROR(EINVAL);
> + }
> +
> + }
> +
> + for (j = 0; j < 4; j++) {
> + matrix_bytes = sizeof(float)*ctx->matrix_sizes[j];
> + matrix = av_malloc(matrix_bytes);
This gets leaked for every plane except the last.
> + if (!matrix) {
> + err = AVERROR(ENOMEM);
> + goto fail;
> + }
> +
> + for (i = 0; i < ctx->matrix_sizes[j]; i++)
> + matrix[i] = input_matrix[j][i];
> +
> + buffer = clCreateBuffer(ctx->ocf.hwctx->context,
> + CL_MEM_READ_ONLY |
> + CL_MEM_COPY_HOST_PTR |
> + CL_MEM_HOST_NO_ACCESS,
> + matrix_bytes, matrix, &cle);
> + if (!buffer) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to create matrix buffer: "
> + "%d.\n", cle);
> + err = AVERROR(EIO);
> + goto fail;
> + }
> + ctx->matrix[j] = buffer;
> + }
> +
> + err = 0;
> +fail:
> + av_freep(&matrix);
> + return err;
> +}
> +
> +static int convolution_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
> +{
> + AVFilterContext *avctx = inlink->dst;
> + AVFilterLink *outlink = avctx->outputs[0];
> + ConvolutionOpenCLContext *ctx = avctx->priv;
> + AVFrame *output = NULL;
> + cl_int cle;
> + size_t global_work[2];
> + cl_mem src, dst;
> + int err, p;
> +
> + av_log(ctx, AV_LOG_DEBUG, "Filter input: %s, %ux%u (%"PRId64").\n",
> + av_get_pix_fmt_name(input->format),
> + input->width, input->height, input->pts);
> +
> + if (!input->hw_frames_ctx)
> + return AVERROR(EINVAL);
> +
> + if (!ctx->initialised) {
> + err = convolution_opencl_init(avctx);
> + if (err < 0)
> + goto fail;
> +
> + err = convolution_opencl_make_filter_params(avctx);
> + if (err < 0)
> + goto fail;
> + }
> +
> + output = ff_get_video_buffer(outlink, outlink->w, outlink->h);
> + if (!output) {
> + err = AVERROR(ENOMEM);
> + goto fail;
> + }
> +
> + for (p = 0; p < FF_ARRAY_ELEMS(output->data); p++) {
> + src = (cl_mem) input->data[p];
> + dst = (cl_mem)output->data[p];
> +
> + if (!dst)
> + break;
> +
> + cle = clSetKernelArg(ctx->kernel, 0, sizeof(cl_mem), &dst);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
> + "destination image argument: %d.\n", cle);
> + goto fail;
> + }
> + cle = clSetKernelArg(ctx->kernel, 1, sizeof(cl_mem), &src);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
> + "source image argument: %d.\n", cle);
> + goto fail;
> + }
> + cle = clSetKernelArg(ctx->kernel, 2, sizeof(cl_int), &ctx->dims[p]);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
> + "matrix size argument: %d.\n", cle);
> + goto fail;
> + }
> + cle = clSetKernelArg(ctx->kernel, 3, sizeof(cl_mem), &ctx->matrix[p]);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
> + "matrix argument: %d.\n", cle);
> + goto fail;
> + }
> + cle = clSetKernelArg(ctx->kernel, 4, sizeof(cl_float), &ctx->rdivs[p]);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
> + "div argument: %d.\n", cle);
> + goto fail;
> + }
> + cle = clSetKernelArg(ctx->kernel, 5, sizeof(cl_float), &ctx->biases[p]);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
> + "bias argument: %d.\n", cle);
> + goto fail;
> + }
> +
> + global_work[0] = output->width;
> + global_work[1] = output->height;
These aren't the right sizes for later planes in subsampled formats. I cleaned this up recently, have a look at ff_opencl_filter_work_size_from_image().
> +
> + av_log(avctx, AV_LOG_DEBUG, "Run kernel on plane %d "
> + "(%"SIZE_SPECIFIER"x%"SIZE_SPECIFIER").\n",
> + p, global_work[0], global_work[1]);
> +
> + cle = clEnqueueNDRangeKernel(ctx->command_queue, ctx->kernel, 2, NULL,
> + global_work, NULL,
> + 0, NULL, NULL);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to enqueue kernel: %d.\n",
> + cle);
> + err = AVERROR(EIO);
> + goto fail;
> + }
> + }
> +
> + cle = clFinish(ctx->command_queue);
> + if (cle != CL_SUCCESS) {
> + av_log(avctx, AV_LOG_ERROR, "Failed to finish command queue: %d.\n",
> + cle);
> + err = AVERROR(EIO);
> + goto fail;
> + }
> +
> + err = av_frame_copy_props(output, input);
> + if (err < 0)
> + goto fail;
> +
> + av_frame_free(&input);
> +
> + av_log(ctx, AV_LOG_DEBUG, "Filter output: %s, %ux%u (%"PRId64").\n",
> + av_get_pix_fmt_name(output->format),
> + output->width, output->height, output->pts);
> +
> + return ff_filter_frame(outlink, output);
> +
> +fail:
> + clFinish(ctx->command_queue);
> + av_frame_free(&input);
> + av_frame_free(&output);
> + return err;
> +}
> +
> +static av_cold void convolution_opencl_uninit(AVFilterContext *avctx)
> +{
> + ConvolutionOpenCLContext *ctx = avctx->priv;
> + cl_int cle;
> +
> + if (ctx->matrix) {
> + clReleaseMemObject(ctx->matrix);
This isn't a single memory object, it's an array. (For me, the filter always segfaults at the end of a file because of this.)
> + }
> +
> + if (ctx->kernel) {
> + cle = clReleaseKernel(ctx->kernel);
> + if (cle != CL_SUCCESS)
> + av_log(avctx, AV_LOG_ERROR, "Failed to release "
> + "kernel: %d.\n", cle);
> + }
> +
> + if (ctx->command_queue) {
> + cle = clReleaseCommandQueue(ctx->command_queue);
> + if (cle != CL_SUCCESS)
> + av_log(avctx, AV_LOG_ERROR, "Failed to release "
> + "command queue: %d.\n", cle);
> + }
> +
> + ff_opencl_filter_uninit(avctx);
> +}
> +
> +#define OFFSET(x) offsetof(ConvolutionOpenCLContext, x)
> +#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
> +static const AVOption convolution_opencl_options[] = {
> + { "0m", "set matrix for 2nd plane", OFFSET(matrix_str[0]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
> + { "1m", "set matrix for 2nd plane", OFFSET(matrix_str[1]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
> + { "2m", "set matrix for 3rd plane", OFFSET(matrix_str[2]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
> + { "3m", "set matrix for 4th plane", OFFSET(matrix_str[3]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
> + { "0rdiv", "set rdiv for 1nd plane", OFFSET(rdivs[0]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
> + { "1rdiv", "set rdiv for 2nd plane", OFFSET(rdivs[1]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
> + { "2rdiv", "set rdiv for 3rd plane", OFFSET(rdivs[2]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
> + { "3rdiv", "set rdiv for 4th plane", OFFSET(rdivs[3]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
> + { "0bias", "set bias for 1st plane", OFFSET(biases[0]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
> + { "1bias", "set bias for 2nd plane", OFFSET(biases[1]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
> + { "2bias", "set bias for 3rd plane", OFFSET(biases[2]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
> + { "3bias", "set bias for 4th plane", OFFSET(biases[3]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
> + { NULL }
> +};
> +
> +AVFILTER_DEFINE_CLASS(convolution_opencl);
> +
> +static const AVFilterPad convolution_opencl_inputs[] = {
> + {
> + .name = "default",
> + .type = AVMEDIA_TYPE_VIDEO,
> + .filter_frame = &convolution_opencl_filter_frame,
> + .config_props = &ff_opencl_filter_config_input,
> + },
> + { NULL }
> +};
> +
> +static const AVFilterPad convolution_opencl_outputs[] = {
> + {
> + .name = "default",
> + .type = AVMEDIA_TYPE_VIDEO,
> + .config_props = &ff_opencl_filter_config_output,
> + },
> + { NULL }
> +};
> +
> +AVFilter ff_vf_convolution_opencl = {
> + .name = "convolution_opencl",
> + .description = NULL_IF_CONFIG_SMALL("Apply convolution mask to input video"),
> + .priv_size = sizeof(ConvolutionOpenCLContext),
> + .priv_class = &convolution_opencl_class,
> + .init = &ff_opencl_filter_init,
> + .uninit = &convolution_opencl_uninit,
> + .query_formats = &ff_opencl_filter_query_formats,
> + .inputs = convolution_opencl_inputs,
> + .outputs = convolution_opencl_outputs,
> + .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
> +};
>
More information about the ffmpeg-devel
mailing list