[FFmpeg-devel] [PATCH] avfilter: add OpenCL scale filter
Gabriel Machado
gabriel_machado at live.com
Tue Mar 27 07:48:22 EEST 2018
From: Gabriel Machado <gabriel_machado at live.com>
Some scaling filters implemented as OpenCL kernels. Can be used as:
scale_opencl=<width>:<height>:flags=<filter>
where <filter> can be `neighbor', `bilinear', `bicubic' or `fast_bicubic'
This is an initial draft, there's still a long way to go in terms of
completeness, configurability and performance.
---
configure | 1 +
libavfilter/Makefile | 1 +
libavfilter/allfilters.c | 1 +
libavfilter/opencl/scale.cl | 165 ++++++++++++++++++++++++
libavfilter/opencl_source.h | 1 +
libavfilter/vf_scale_opencl.c | 289 ++++++++++++++++++++++++++++++++++++++++++
6 files changed, 458 insertions(+)
create mode 100644 libavfilter/opencl/scale.cl
create mode 100644 libavfilter/vf_scale_opencl.c
diff --git a/configure b/configure
index 5ccf3ce..4007ee8 100755
--- a/configure
+++ b/configure
@@ -2821,6 +2821,7 @@ v4l2_m2m_deps_any="linux_videodev2_h"
hwupload_cuda_filter_deps="ffnvcodec"
scale_npp_filter_deps="ffnvcodec libnpp"
+scale_opencl_filter_deps="opencl"
scale_cuda_filter_deps="cuda_sdk"
thumbnail_cuda_filter_deps="cuda_sdk"
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index a90ca30..6303cbd 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -302,6 +302,7 @@ OBJS-$(CONFIG_SAB_FILTER) += vf_sab.o
OBJS-$(CONFIG_SCALE_FILTER) += vf_scale.o scale.o
OBJS-$(CONFIG_SCALE_CUDA_FILTER) += vf_scale_cuda.o vf_scale_cuda.ptx.o
OBJS-$(CONFIG_SCALE_NPP_FILTER) += vf_scale_npp.o scale.o
+OBJS-$(CONFIG_SCALE_OPENCL_FILTER) += vf_scale_opencl.o opencl.o opencl/scale.o
OBJS-$(CONFIG_SCALE_QSV_FILTER) += vf_scale_qsv.o
OBJS-$(CONFIG_SCALE_VAAPI_FILTER) += vf_scale_vaapi.o scale.o vaapi_vpp.o
OBJS-$(CONFIG_SCALE2REF_FILTER) += vf_scale.o scale.o
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index 1cf1340..3185b17 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -309,6 +309,7 @@ static void register_all(void)
REGISTER_FILTER(SCALE, scale, vf);
REGISTER_FILTER(SCALE_CUDA, scale_cuda, vf);
REGISTER_FILTER(SCALE_NPP, scale_npp, vf);
+ REGISTER_FILTER(SCALE_OPENCL, scale_opencl, vf);
REGISTER_FILTER(SCALE_QSV, scale_qsv, vf);
REGISTER_FILTER(SCALE_VAAPI, scale_vaapi, vf);
REGISTER_FILTER(SCALE2REF, scale2ref, vf);
diff --git a/libavfilter/opencl/scale.cl b/libavfilter/opencl/scale.cl
new file mode 100644
index 0000000..b0e6cb2
--- /dev/null
+++ b/libavfilter/opencl/scale.cl
@@ -0,0 +1,165 @@
+/*
+ * Copyright (c) 2018 Gabriel Machado
+ *
+ * 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 neighbor(__write_only image2d_t dst,
+ __read_only image2d_t src)
+{
+ const sampler_t sampler = (CLK_NORMALIZED_COORDS_TRUE |
+ CLK_ADDRESS_CLAMP_TO_EDGE |
+ CLK_FILTER_NEAREST);
+
+ int2 coord = {get_global_id(0), get_global_id(1)};
+ int2 size = {get_global_size(0), get_global_size(1)};
+
+ float2 pos = (convert_float2(coord) + 0.5) / convert_float2(size);
+
+ float4 c = read_imagef(src, sampler, pos);
+ write_imagef(dst, coord, c);
+}
+
+__kernel void bilinear(__write_only image2d_t dst,
+ __read_only image2d_t src)
+{
+ const sampler_t sampler = (CLK_NORMALIZED_COORDS_TRUE |
+ CLK_ADDRESS_CLAMP_TO_EDGE |
+ CLK_FILTER_LINEAR);
+
+ int2 coord = {get_global_id(0), get_global_id(1)};
+ int2 size = {get_global_size(0), get_global_size(1)};
+
+ float2 pos = (convert_float2(coord) + 0.5) / convert_float2(size);
+
+ float4 c = read_imagef(src, sampler, pos);
+ write_imagef(dst, coord, c);
+}
+
+// https://developer.nvidia.com/gpugems/GPUGems/gpugems_ch24.html
+float MitchellNetravali(float x, float B, float C)
+{
+ float t = fabs(x);
+ float tt = t*t;
+ float ttt = tt*t;
+
+ if (t < 1) {
+ return ((12 - 9 * B - 6 * C) * ttt +
+ (-18 + 12 * B + 6 * C) * tt + (6 - 2 * B)) / 6;
+ } else if ((t >= 1) && (t < 2)) {
+ return ((-B - 6 * C) * ttt +
+ (6 * B + 30 * C) * tt + (-12 * B - 48 * C) *
+ t + (8 * B + 24 * C)) / 6;
+ } else {
+ return 0;
+ }
+}
+
+float4 cubic(float4 c0, float4 c1, float4 c2, float4 c3, float t)
+{
+ float B = 0, C = 0.6; // libswscale default
+ float a = MitchellNetravali(t + 1, B, C);
+ float b = MitchellNetravali(t, B, C);
+ float c = MitchellNetravali(1 - t, B, C);
+ float d = MitchellNetravali(2 - t, B, C);
+ return a*c0 + b*c1 + c*c2 + d*c3;
+}
+
+__kernel void bicubic(__write_only image2d_t dst,
+ __read_only image2d_t src)
+{
+ const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
+ CLK_ADDRESS_CLAMP_TO_EDGE |
+ CLK_FILTER_NEAREST);
+
+ int2 dst_coord = {get_global_id(0), get_global_id(1)};
+
+ float2 dst_size = {get_global_size(0), get_global_size(1)};
+ float2 src_size = convert_float2(get_image_dim(src));
+
+ float2 uv = convert_float2(dst_coord) / dst_size;
+
+ float2 src_pos = uv * convert_float2(src_size) - 0.5;
+
+ float2 src_coordf;
+ float2 t = fract(src_pos, &src_coordf);
+ int2 src_coord = convert_int2(src_coordf);
+
+#define TEX(x,y) read_imagef(src, sampler, src_coord + (int2){x,y})
+ float4 col = cubic(cubic(TEX(-1,-1), TEX(0,-1), TEX(1,-1), TEX(2,-1), t.x),
+ cubic(TEX(-1, 0), TEX(0, 0), TEX(1, 0), TEX(2, 0), t.x),
+ cubic(TEX(-1, 1), TEX(0, 1), TEX(1, 1), TEX(2, 1), t.x),
+ cubic(TEX(-1, 2), TEX(0, 2), TEX(1, 2), TEX(2, 2), t.x),
+ t.y);
+#undef TEX
+
+ write_imagef(dst, dst_coord, col);
+}
+
+// https://www.shadertoy.com/view/4df3Dn
+// 4x4 bicubic filter using 4 bilinear texture lookups
+// cubic B-spline basis functions
+float w0(float a) { return (1.0/6.0)*(a*(a*(-a + 3.0) - 3.0) + 1.0); }
+float w1(float a) { return (1.0/6.0)*(a*a*(3.0*a - 6.0) + 4.0); }
+float w2(float a) { return (1.0/6.0)*(a*(a*(-3.0*a + 3.0) + 3.0) + 1.0); }
+float w3(float a) { return (1.0/6.0)*(a*a*a); }
+
+float g0(float a) { return w0(a) + w1(a); }
+float g1(float a) { return w2(a) + w3(a); }
+
+float h0(float a) { return -1.0 + w1(a) / g0(a); }
+float h1(float a) { return 1.0 + w3(a) / g1(a); }
+
+__kernel void fast_bicubic(__write_only image2d_t dst,
+ __read_only image2d_t src)
+{
+ const sampler_t sampler = (CLK_NORMALIZED_COORDS_TRUE |
+ CLK_ADDRESS_CLAMP_TO_EDGE |
+ CLK_FILTER_LINEAR);
+
+ int2 dst_coord = {get_global_id(0), get_global_id(1)};
+
+ float2 dst_size = {get_global_size(0), get_global_size(1)};
+ float2 src_size = convert_float2(get_image_dim(src));
+
+ float2 uv = convert_float2(dst_coord) / dst_size;
+
+ float2 src_pos = uv * src_size + 0.5;
+
+ float2 src_coordf;
+ float2 frac = fract(src_pos, &src_coordf);
+ int2 src_coord = convert_int2(src_coordf);
+
+ float g0x = g0(frac.x);
+ float g1x = g1(frac.x);
+ float h0x = h0(frac.x);
+ float h1x = h1(frac.x);
+ float h0y = h0(frac.y);
+ float h1y = h1(frac.y);
+
+ float2 p0 = ((float2){src_coord.x + h0x, src_coord.y + h0y} - 0.5) / src_size;
+ float2 p1 = ((float2){src_coord.x + h1x, src_coord.y + h0y} - 0.5) / src_size;
+ float2 p2 = ((float2){src_coord.x + h0x, src_coord.y + h1y} - 0.5) / src_size;
+ float2 p3 = ((float2){src_coord.x + h1x, src_coord.y + h1y} - 0.5) / src_size;
+
+ float4 c = g0(frac.y) * (g0x * read_imagef(src, sampler, p0) +
+ g1x * read_imagef(src, sampler, p1)) +
+ g1(frac.y) * (g0x * read_imagef(src, sampler, p2) +
+ g1x * read_imagef(src, sampler, p3));
+
+ write_imagef(dst, dst_coord, c);
+}
diff --git a/libavfilter/opencl_source.h b/libavfilter/opencl_source.h
index 4bb9969..e3bb887 100644
--- a/libavfilter/opencl_source.h
+++ b/libavfilter/opencl_source.h
@@ -22,6 +22,7 @@
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_scale;
extern const char *ff_opencl_source_unsharp;
#endif /* AVFILTER_OPENCL_SOURCE_H */
diff --git a/libavfilter/vf_scale_opencl.c b/libavfilter/vf_scale_opencl.c
new file mode 100644
index 0000000..14479c7
--- /dev/null
+++ b/libavfilter/vf_scale_opencl.c
@@ -0,0 +1,289 @@
+/*
+ * Copyright (c) 2018 Gabriel Machado
+ *
+ * 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 "avfilter.h"
+#include "internal.h"
+#include "opencl.h"
+#include "opencl_source.h"
+#include "scale.h"
+#include "video.h"
+
+#define F_NEIGHBOR 0
+#define F_BILINEAR 1
+#define F_BICUBIC 2
+#define F_FAST_BICUBIC 3
+
+typedef struct ScaleOpenCLContext {
+ OpenCLFilterContext ocf;
+
+ int initialised;
+ cl_kernel kernel;
+ cl_command_queue command_queue;
+
+ char *w_expr;
+ char *h_expr;
+ int algorithm;
+} ScaleOpenCLContext;
+
+static const char *kernel_name[] = {
+ "neighbor",
+ "bilinear",
+ "bicubic",
+ "fast_bicubic"
+};
+
+static int scale_opencl_init(AVFilterContext *avctx)
+{
+ ScaleOpenCLContext *ctx = avctx->priv;
+ cl_int cle;
+ int err;
+
+ err = ff_opencl_filter_load_program(avctx, &ff_opencl_source_scale, 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, kernel_name[ctx->algorithm], &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 config_input_props(AVFilterLink *inlink)
+{
+ AVFilterContext *avctx = inlink->dst;
+ AVFilterLink *outlink = avctx->outputs[0];
+ ScaleOpenCLContext *ctx = avctx->priv;
+ int w, h;
+ int ret;
+
+ if ((ret = ff_scale_eval_dimensions(ctx,
+ ctx->w_expr, ctx->h_expr,
+ inlink, outlink,
+ &w, &h)) < 0)
+ return ret;
+
+ if (((int64_t)h * inlink->w) > INT_MAX ||
+ ((int64_t)w * inlink->h) > INT_MAX)
+ av_log(ctx, AV_LOG_ERROR, "Rescaled value for width or height is too big.\n");
+
+ ctx->ocf.output_width = w;
+ ctx->ocf.output_height = h;
+
+ return ff_opencl_filter_config_input(inlink);
+}
+
+static int scale_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
+{
+ AVFilterContext *avctx = inlink->dst;
+ AVFilterLink *outlink = avctx->outputs[0];
+ ScaleOpenCLContext *ctx = avctx->priv;
+ AVHWFramesContext *main_fc = (AVHWFramesContext *) inlink->hw_frames_ctx->data;
+ const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(main_fc->sw_format);
+ AVFrame *output = NULL;
+
+ cl_int cle;
+ size_t global_work[2];
+ cl_mem src, dst;
+ int err, p;
+ int x_subsample = 1 << pix_desc->log2_chroma_w;
+ int y_subsample = 1 << pix_desc->log2_chroma_h;
+
+ 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 = scale_opencl_init(avctx);
+ if (err < 0)
+ goto fail;
+ }
+
+ output = ff_get_video_buffer(outlink, outlink->w, outlink->h);
+ if (!output) {
+ err = AVERROR(ENOMEM);
+ goto fail;
+ }
+
+ err = av_frame_copy_props(output, input);
+ if (err < 0)
+ goto fail;
+ output->width = outlink->w;
+ output->height = outlink->h;
+
+ 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;
+ }
+
+ global_work[0] = output->width / (p ? x_subsample : 1);
+ global_work[1] = output->height / (p ? y_subsample : 1);
+
+ 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;
+ }
+
+ 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 scale_opencl_uninit(AVFilterContext *avctx)
+{
+ ScaleOpenCLContext *ctx = avctx->priv;
+ cl_int cle;
+
+ 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(ScaleOpenCLContext, x)
+#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
+static const AVOption scale_opencl_options[] = {
+ { "w", "Output video width", OFFSET(w_expr), AV_OPT_TYPE_STRING, { .str = "iw" }, .flags = FLAGS },
+ { "h", "Output video height", OFFSET(h_expr), AV_OPT_TYPE_STRING, { .str = "ih" }, .flags = FLAGS },
+ { "algorithm", "Scaling algorithm", OFFSET(algorithm), AV_OPT_TYPE_INT, { .i64 = F_BICUBIC }, INT_MIN, INT_MAX, FLAGS, "algorithm" },
+ { "bilinear", "bilinear", 0, AV_OPT_TYPE_CONST, { .i64 = F_BILINEAR }, 0, 0, FLAGS, "algorithm" },
+ { "bicubic", "bicubic", 0, AV_OPT_TYPE_CONST, { .i64 = F_BICUBIC }, 0, 0, FLAGS, "algorithm" },
+ { "fast_bicubic", "fast bicubic", 0, AV_OPT_TYPE_CONST, { .i64 = F_FAST_BICUBIC }, 0, 0, FLAGS, "algorithm" },
+ { "neighbor", "nearest neighbor", 0, AV_OPT_TYPE_CONST, { .i64 = F_NEIGHBOR }, 0, 0, FLAGS, "algorithm" },
+ { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(scale_opencl);
+
+static const AVFilterPad scale_opencl_inputs[] = {
+ {
+ .name = "default",
+ .type = AVMEDIA_TYPE_VIDEO,
+ .filter_frame = &scale_opencl_filter_frame,
+ .config_props = &config_input_props,
+ },
+ { NULL }
+};
+
+static const AVFilterPad scale_opencl_outputs[] = {
+ {
+ .name = "default",
+ .type = AVMEDIA_TYPE_VIDEO,
+ .config_props = &ff_opencl_filter_config_output,
+ },
+ { NULL }
+};
+
+AVFilter ff_vf_scale_opencl = {
+ .name = "scale_opencl",
+ .description = NULL_IF_CONFIG_SMALL("Scale the input video size."),
+ .priv_size = sizeof(ScaleOpenCLContext),
+ .priv_class = &scale_opencl_class,
+ .init = &ff_opencl_filter_init,
+ .uninit = &scale_opencl_uninit,
+ .query_formats = &ff_opencl_filter_query_formats,
+ .inputs = scale_opencl_inputs,
+ .outputs = scale_opencl_outputs,
+ .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
+};
--
2.7.4
More information about the ffmpeg-devel
mailing list