libavfilter: Add OpenCL convolution filter
authorDanil Iashchenko <danyaschenko@gmail.com>
Sat, 24 Mar 2018 15:57:43 +0000 (18:57 +0300)
committerMark Thompson <sw@jkqxz.net>
Mon, 26 Mar 2018 21:26:35 +0000 (22:26 +0100)
Behaves like existing convolution filter.

Signed-off-by: Mark Thompson <sw@jkqxz.net>
configure
libavfilter/Makefile
libavfilter/allfilters.c
libavfilter/opencl/convolution.cl [new file with mode: 0644]
libavfilter/opencl_source.h
libavfilter/vf_convolution_opencl.c [new file with mode: 0644]

index cc3edeb..5ccf3ce 100755 (executable)
--- a/configure
+++ b/configure
@@ -3235,6 +3235,7 @@ 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"
index 1043b41..a90ca30 100644 (file)
@@ -159,6 +159,8 @@ 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
index 3f67e32..1cf1340 100644 (file)
@@ -168,6 +168,7 @@ 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 (file)
index 0000000..03ef4ef
--- /dev/null
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 2018 Danil Iashchenko
+ *
+ * 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);
+}
index 02bc172..4bb9969 100644 (file)
@@ -20,6 +20,7 @@
 #define AVFILTER_OPENCL_SOURCE_H
 
 extern const char *ff_opencl_source_avgblur;
+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 (file)
index 0000000..2df51e0
--- /dev/null
@@ -0,0 +1,368 @@
+/*
+ * Copyright (c) 2018 Danil Iashchenko
+ *
+ * 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 = NULL;
+    size_t matrix_bytes;
+    cl_mem buffer;
+    cl_int cle;
+    int i, j;
+    int sscanf_err;
+    char *p, *arg, *saveptr = NULL;
+    float input_matrix[4][49];
+
+    for (i = 0; i < 4; i++) {
+        ctx->biases[i] = ctx->biases[i] / 255.0;
+    }
+
+    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);
+        if (!matrix) {
+            av_freep(&matrix);
+            return AVERROR(ENOMEM);
+        }
+
+        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);
+            av_freep(&matrix);
+            return AVERROR(EIO);
+        }
+        ctx->matrix[j] = buffer;
+        av_freep(&matrix);
+    }
+
+    return 0;
+}
+
+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 "
+                   "rdiv 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;
+        }
+
+
+        err = ff_opencl_filter_work_size_from_image(avctx, global_work, output, p, 0);
+        if (err < 0)
+            goto fail;
+
+        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;
+    int i;
+
+    for (i = 0; i < 4; i++) {
+        clReleaseMemObject(ctx->matrix[i]);
+    }
+
+    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,
+};