Commit 581bafa8 authored by Danil Iashchenko's avatar Danil Iashchenko Committed by Mark Thompson

lavfi: add sobel, prewitt, roberts filters

Add opencl version of sobel, prewitt, roberts filters.
parent 54b425a7
......@@ -3372,12 +3372,14 @@ perspective_filter_deps="gpl"
phase_filter_deps="gpl"
pp7_filter_deps="gpl"
pp_filter_deps="gpl postproc"
prewitt_opencl_filter_deps="opencl"
procamp_vaapi_filter_deps="vaapi VAProcPipelineParameterBuffer"
program_opencl_filter_deps="opencl"
pullup_filter_deps="gpl"
removelogo_filter_deps="avcodec avformat swscale"
repeatfields_filter_deps="gpl"
resample_filter_deps="avresample"
roberts_opencl_filter_deps="opencl"
rubberband_filter_deps="librubberband"
sab_filter_deps="gpl swscale"
scale2ref_filter_deps="swscale"
......@@ -3396,6 +3398,7 @@ showspectrumpic_filter_deps="avcodec"
showspectrumpic_filter_select="fft"
signature_filter_deps="gpl avcodec avformat"
smartblur_filter_deps="gpl swscale"
sobel_opencl_filter_deps="opencl"
sofalizer_filter_deps="libmysofa avcodec"
sofalizer_filter_select="fft"
spectrumsynth_filter_deps="avcodec"
......
......@@ -171,7 +171,7 @@ OBJS-$(CONFIG_COLORMATRIX_FILTER) += vf_colormatrix.o
OBJS-$(CONFIG_COLORSPACE_FILTER) += vf_colorspace.o 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
opencl/convolution.o
OBJS-$(CONFIG_CONVOLVE_FILTER) += vf_convolve.o framesync.o
OBJS-$(CONFIG_COPY_FILTER) += vf_copy.o
OBJS-$(CONFIG_COREIMAGE_FILTER) += vf_coreimage.o
......@@ -294,6 +294,8 @@ OBJS-$(CONFIG_PP_FILTER) += vf_pp.o
OBJS-$(CONFIG_PP7_FILTER) += vf_pp7.o
OBJS-$(CONFIG_PREMULTIPLY_FILTER) += vf_premultiply.o framesync.o
OBJS-$(CONFIG_PREWITT_FILTER) += vf_convolution.o
OBJS-$(CONFIG_PREWITT_OPENCL_FILTER) += vf_convolution_opencl.o opencl.o \
opencl/convolution.o
OBJS-$(CONFIG_PROCAMP_VAAPI_FILTER) += vf_procamp_vaapi.o vaapi_vpp.o
OBJS-$(CONFIG_PROGRAM_OPENCL_FILTER) += vf_program_opencl.o opencl.o framesync.o
OBJS-$(CONFIG_PSEUDOCOLOR_FILTER) += vf_pseudocolor.o
......@@ -310,6 +312,8 @@ OBJS-$(CONFIG_REMOVELOGO_FILTER) += bbox.o lswsutils.o lavfutils.o v
OBJS-$(CONFIG_REPEATFIELDS_FILTER) += vf_repeatfields.o
OBJS-$(CONFIG_REVERSE_FILTER) += f_reverse.o
OBJS-$(CONFIG_ROBERTS_FILTER) += vf_convolution.o
OBJS-$(CONFIG_ROBERTS_OPENCL_FILTER) += vf_convolution_opencl.o opencl.o \
opencl/convolution.o
OBJS-$(CONFIG_ROTATE_FILTER) += vf_rotate.o
OBJS-$(CONFIG_SAB_FILTER) += vf_sab.o
OBJS-$(CONFIG_SCALE_FILTER) += vf_scale.o scale.o
......@@ -338,6 +342,8 @@ OBJS-$(CONFIG_SIGNALSTATS_FILTER) += vf_signalstats.o
OBJS-$(CONFIG_SIGNATURE_FILTER) += vf_signature.o
OBJS-$(CONFIG_SMARTBLUR_FILTER) += vf_smartblur.o
OBJS-$(CONFIG_SOBEL_FILTER) += vf_convolution.o
OBJS-$(CONFIG_SOBEL_OPENCL_FILTER) += vf_convolution_opencl.o opencl.o \
opencl/convolution.o
OBJS-$(CONFIG_SPLIT_FILTER) += split.o
OBJS-$(CONFIG_SPP_FILTER) += vf_spp.o
OBJS-$(CONFIG_SR_FILTER) += vf_sr.o
......
......@@ -282,6 +282,7 @@ extern AVFilter ff_vf_pp;
extern AVFilter ff_vf_pp7;
extern AVFilter ff_vf_premultiply;
extern AVFilter ff_vf_prewitt;
extern AVFilter ff_vf_prewitt_opencl;
extern AVFilter ff_vf_procamp_vaapi;
extern AVFilter ff_vf_program_opencl;
extern AVFilter ff_vf_pseudocolor;
......@@ -298,6 +299,7 @@ extern AVFilter ff_vf_removelogo;
extern AVFilter ff_vf_repeatfields;
extern AVFilter ff_vf_reverse;
extern AVFilter ff_vf_roberts;
extern AVFilter ff_vf_roberts_opencl;
extern AVFilter ff_vf_rotate;
extern AVFilter ff_vf_sab;
extern AVFilter ff_vf_scale;
......@@ -326,6 +328,7 @@ extern AVFilter ff_vf_signalstats;
extern AVFilter ff_vf_signature;
extern AVFilter ff_vf_smartblur;
extern AVFilter ff_vf_sobel;
extern AVFilter ff_vf_sobel_opencl;
extern AVFilter ff_vf_split;
extern AVFilter ff_vf_spp;
extern AVFilter ff_vf_sr;
......
......@@ -43,3 +43,85 @@ __kernel void convolution_global(__write_only image2d_t dst,
float4 dstPix = convPix * div + bias;
write_imagef(dst, loc, dstPix);
}
__kernel void sobel_global(__write_only image2d_t dst,
__read_only image2d_t src,
float div,
float bias)
{
const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
CLK_ADDRESS_CLAMP_TO_EDGE |
CLK_FILTER_NEAREST);
int2 loc = (int2)(get_global_id(0), get_global_id(1));
float4 sum1 = read_imagef(src, sampler, loc + (int2)(-1,-1)) * -1 +
read_imagef(src, sampler, loc + (int2)( 0,-1)) * -2 +
read_imagef(src, sampler, loc + (int2)( 1,-1)) * -1 +
read_imagef(src, sampler, loc + (int2)(-1, 1)) * 1 +
read_imagef(src, sampler, loc + (int2)( 0, 1)) * 2 +
read_imagef(src, sampler, loc + (int2)( 1, 1)) * 1;
float4 sum2 = read_imagef(src, sampler, loc + (int2)(-1,-1)) * -1 +
read_imagef(src, sampler, loc + (int2)(-1, 0)) * -2 +
read_imagef(src, sampler, loc + (int2)(-1, 1)) * -1 +
read_imagef(src, sampler, loc + (int2)( 1,-1)) * 1 +
read_imagef(src, sampler, loc + (int2)( 1, 0)) * 2 +
read_imagef(src, sampler, loc + (int2)( 1, 1)) * 1;
float4 dstPix = hypot(sum1, sum2) * div + bias;
write_imagef(dst, loc, dstPix);
}
__kernel void prewitt_global(__write_only image2d_t dst,
__read_only image2d_t src,
float div,
float bias)
{
const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
CLK_ADDRESS_CLAMP_TO_EDGE |
CLK_FILTER_NEAREST);
int2 loc = (int2)(get_global_id(0), get_global_id(1));
float4 sum1 = read_imagef(src, sampler, loc + (int2)(-1,-1)) * 1 +
read_imagef(src, sampler, loc + (int2)( 0,-1)) * 1 +
read_imagef(src, sampler, loc + (int2)( 1,-1)) * 1 +
read_imagef(src, sampler, loc + (int2)(-1, 1)) * -1 +
read_imagef(src, sampler, loc + (int2)( 0, 1)) * -1 +
read_imagef(src, sampler, loc + (int2)( 1, 1)) * -1;
float4 sum2 = read_imagef(src, sampler, loc + (int2)(-1,-1)) * 1 +
read_imagef(src, sampler, loc + (int2)(-1, 0)) * 1 +
read_imagef(src, sampler, loc + (int2)(-1, 1)) * 1 +
read_imagef(src, sampler, loc + (int2)( 1,-1)) * -1 +
read_imagef(src, sampler, loc + (int2)( 1, 0)) * -1 +
read_imagef(src, sampler, loc + (int2)( 1, 1)) * -1;
float4 dstPix = hypot(sum1, sum2) * div + bias;
write_imagef(dst, loc, dstPix);
}
__kernel void roberts_global(__write_only image2d_t dst,
__read_only image2d_t src,
float div,
float bias)
{
const sampler_t sampler = (CLK_NORMALIZED_COORDS_FALSE |
CLK_ADDRESS_CLAMP_TO_EDGE |
CLK_FILTER_NEAREST);
int2 loc = (int2)(get_global_id(0), get_global_id(1));
float4 sum1 = read_imagef(src, sampler, loc + (int2)(-1,-1)) * 1 +
read_imagef(src, sampler, loc + (int2)( 0,-1)) * -1;
float4 sum2 = read_imagef(src, sampler, loc + (int2)(-1, 0)) * -1 +
read_imagef(src, sampler, loc + (int2)( 0, 0)) * 1;
float4 dstPix = hypot(sum1, sum2) * div + bias;
write_imagef(dst, loc, dstPix);
}
......@@ -47,12 +47,16 @@ typedef struct ConvolutionOpenCLContext {
cl_float rdivs[4];
cl_float biases[4];
} ConvolutionOpenCLContext;
cl_int planes;
cl_float scale;
cl_float delta;
} ConvolutionOpenCLContext;
static int convolution_opencl_init(AVFilterContext *avctx)
{
ConvolutionOpenCLContext *ctx = avctx->priv;
const char *kernel_name;
cl_int cle;
int err;
......@@ -70,7 +74,16 @@ static int convolution_opencl_init(AVFilterContext *avctx)
goto fail;
}
ctx->kernel = clCreateKernel(ctx->ocf.program, "convolution_global", &cle);
if (!strcmp(avctx->filter->name, "convolution_opencl")) {
kernel_name = "convolution_global";
} else if (!strcmp(avctx->filter->name, "sobel_opencl")) {
kernel_name = "sobel_global";
} else if (!strcmp(avctx->filter->name, "prewitt_opencl")){
kernel_name = "prewitt_global";
} else if (!strcmp(avctx->filter->name, "roberts_opencl")){
kernel_name = "roberts_global";
}
ctx->kernel = clCreateKernel(ctx->ocf.program, kernel_name, &cle);
if (!ctx->kernel) {
av_log(avctx, AV_LOG_ERROR, "Failed to create kernel: %d.\n", cle);
err = AVERROR(EIO);
......@@ -173,6 +186,8 @@ static int convolution_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
size_t global_work[2];
cl_mem src, dst;
int err, p;
size_t origin[3] = {0, 0, 0};
size_t region[3] = {0, 0, 1};
av_log(ctx, AV_LOG_DEBUG, "Filter input: %s, %ux%u (%"PRId64").\n",
av_get_pix_fmt_name(input->format),
......@@ -186,9 +201,14 @@ static int convolution_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
if (err < 0)
goto fail;
err = convolution_opencl_make_filter_params(avctx);
if (err < 0)
goto fail;
if (!strcmp(avctx->filter->name, "convolution_opencl")) {
err = convolution_opencl_make_filter_params(avctx);
if (err < 0)
goto fail;
} else {
ctx->delta /= 255.0;
}
}
output = ff_get_video_buffer(outlink, outlink->w, outlink->h);
......@@ -204,29 +224,69 @@ static int convolution_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
if (!dst)
break;
CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_int, &ctx->dims[p]);
CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_mem, &ctx->matrix[p]);
CL_SET_KERNEL_ARG(ctx->kernel, 4, cl_float, &ctx->rdivs[p]);
CL_SET_KERNEL_ARG(ctx->kernel, 5, cl_float, &ctx->biases[p]);
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;
if (!strcmp(avctx->filter->name, "convolution_opencl")) {
CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_int, &ctx->dims[p]);
CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_mem, &ctx->matrix[p]);
CL_SET_KERNEL_ARG(ctx->kernel, 4, cl_float, &ctx->rdivs[p]);
CL_SET_KERNEL_ARG(ctx->kernel, 5, cl_float, &ctx->biases[p]);
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;
}
} else {
if (!(ctx->planes & (1 << p))) {
err = ff_opencl_filter_work_size_from_image(avctx, region, output, p, 0);
if (err < 0)
goto fail;
cle = clEnqueueCopyImage(ctx->command_queue, src, dst,
origin, origin, region, 0, NULL, NULL);
if (cle != CL_SUCCESS) {
av_log(avctx, AV_LOG_ERROR, "Failed to copy plane %d: %d.\n",
p, cle);
err = AVERROR(EIO);
goto fail;
}
} else {
CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_float, &ctx->scale);
CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_float, &ctx->delta);
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;
}
}
}
}
......@@ -284,8 +344,30 @@ static av_cold void convolution_opencl_uninit(AVFilterContext *avctx)
ff_opencl_filter_uninit(avctx);
}
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 }
};
#define OFFSET(x) offsetof(ConvolutionOpenCLContext, x)
#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
#if CONFIG_CONVOLUTION_OPENCL_FILTER
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 },
......@@ -304,30 +386,89 @@ static const AVOption convolution_opencl_options[] = {
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,
},
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,
};
#endif /* CONFIG_CONVOLUTION_OPENCL_FILTER */
#if CONFIG_SOBEL_OPENCL_FILTER
static const AVOption sobel_opencl_options[] = {
{ "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15}, 0, 15, FLAGS},
{ "scale", "set scale", OFFSET(scale), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, 65535, FLAGS},
{ "delta", "set delta", OFFSET(delta), AV_OPT_TYPE_FLOAT, {.dbl=0}, -65535, 65535, FLAGS},
{ NULL }
};
static const AVFilterPad convolution_opencl_outputs[] = {
{
.name = "default",
.type = AVMEDIA_TYPE_VIDEO,
.config_props = &ff_opencl_filter_config_output,
},
AVFILTER_DEFINE_CLASS(sobel_opencl);
AVFilter ff_vf_sobel_opencl = {
.name = "sobel_opencl",
.description = NULL_IF_CONFIG_SMALL("Apply sobel operator"),
.priv_size = sizeof(ConvolutionOpenCLContext),
.priv_class = &sobel_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,
};
#endif /* CONFIG_SOBEL_OPENCL_FILTER */
#if CONFIG_PREWITT_OPENCL_FILTER
static const AVOption prewitt_opencl_options[] = {
{ "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15}, 0, 15, FLAGS},
{ "scale", "set scale", OFFSET(scale), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, 65535, FLAGS},
{ "delta", "set delta", OFFSET(delta), AV_OPT_TYPE_FLOAT, {.dbl=0}, -65535, 65535, FLAGS},
{ NULL }
};
AVFilter ff_vf_convolution_opencl = {
.name = "convolution_opencl",
.description = NULL_IF_CONFIG_SMALL("Apply convolution mask to input video"),
AVFILTER_DEFINE_CLASS(prewitt_opencl);
AVFilter ff_vf_prewitt_opencl = {
.name = "prewitt_opencl",
.description = NULL_IF_CONFIG_SMALL("Apply prewitt operator"),
.priv_size = sizeof(ConvolutionOpenCLContext),
.priv_class = &convolution_opencl_class,
.priv_class = &prewitt_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,
};
#endif /* CONFIG_PREWITT_OPENCL_FILTER */
#if CONFIG_ROBERTS_OPENCL_FILTER
static const AVOption roberts_opencl_options[] = {
{ "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15}, 0, 15, FLAGS},
{ "scale", "set scale", OFFSET(scale), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, 65535, FLAGS},
{ "delta", "set delta", OFFSET(delta), AV_OPT_TYPE_FLOAT, {.dbl=0}, -65535, 65535, FLAGS},
{ NULL }
};
AVFILTER_DEFINE_CLASS(roberts_opencl);
AVFilter ff_vf_roberts_opencl = {
.name = "roberts_opencl",
.description = NULL_IF_CONFIG_SMALL("Apply roberts operator"),
.priv_size = sizeof(ConvolutionOpenCLContext),
.priv_class = &roberts_opencl_class,
.init = &ff_opencl_filter_init,
.uninit = &convolution_opencl_uninit,
.query_formats = &ff_opencl_filter_query_formats,
......@@ -335,3 +476,5 @@ AVFilter ff_vf_convolution_opencl = {
.outputs = convolution_opencl_outputs,
.flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
};
#endif /* CONFIG_ROBERTS_OPENCL_FILTER */
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment