2 * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
27 #include "libavutil/avstring.h"
28 #include "libavutil/common.h"
29 #include "libavutil/hwcontext.h"
30 #include "libavutil/hwcontext_cuda.h"
31 #include "libavutil/cuda_check.h"
32 #include "libavutil/internal.h"
33 #include "libavutil/opt.h"
34 #include "libavutil/pixdesc.h"
42 static const enum AVPixelFormat supported_formats[] = {
50 #define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
51 #define ALIGN_UP(a, b) (((a) + (b) - 1) & ~((b) - 1))
56 #define CHECK_CU(x) FF_CUDA_CHECK(ctx, x)
58 typedef struct CUDAScaleContext {
60 enum AVPixelFormat in_fmt;
61 enum AVPixelFormat out_fmt;
66 } planes_in[3], planes_out[3];
68 AVBufferRef *frames_ctx;
75 * Output sw format. AV_PIX_FMT_NONE for no conversion.
77 enum AVPixelFormat format;
79 char *w_expr; ///< width expression string
80 char *h_expr; ///< height expression string
85 CUfunction cu_func_uchar;
86 CUfunction cu_func_uchar2;
87 CUfunction cu_func_uchar4;
88 CUfunction cu_func_ushort;
89 CUfunction cu_func_ushort2;
90 CUfunction cu_func_ushort4;
91 CUtexref cu_tex_uchar;
92 CUtexref cu_tex_uchar2;
93 CUtexref cu_tex_uchar4;
94 CUtexref cu_tex_ushort;
95 CUtexref cu_tex_ushort2;
96 CUtexref cu_tex_ushort4;
98 CUdeviceptr srcBuffer;
99 CUdeviceptr dstBuffer;
103 static av_cold int cudascale_init(AVFilterContext *ctx)
105 CUDAScaleContext *s = ctx->priv;
107 s->format = AV_PIX_FMT_NONE;
108 s->frame = av_frame_alloc();
110 return AVERROR(ENOMEM);
112 s->tmp_frame = av_frame_alloc();
114 return AVERROR(ENOMEM);
119 static av_cold void cudascale_uninit(AVFilterContext *ctx)
121 CUDAScaleContext *s = ctx->priv;
123 av_frame_free(&s->frame);
124 av_buffer_unref(&s->frames_ctx);
125 av_frame_free(&s->tmp_frame);
128 static int cudascale_query_formats(AVFilterContext *ctx)
130 static const enum AVPixelFormat pixel_formats[] = {
131 AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE,
133 AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats);
135 return ff_set_common_formats(ctx, pix_fmts);
138 static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
140 AVBufferRef *out_ref = NULL;
141 AVHWFramesContext *out_ctx;
142 int in_sw, in_sh, out_sw, out_sh;
145 av_pix_fmt_get_chroma_sub_sample(s->in_fmt, &in_sw, &in_sh);
146 av_pix_fmt_get_chroma_sub_sample(s->out_fmt, &out_sw, &out_sh);
147 if (!s->planes_out[0].width) {
148 s->planes_out[0].width = s->planes_in[0].width;
149 s->planes_out[0].height = s->planes_in[0].height;
152 for (i = 1; i < FF_ARRAY_ELEMS(s->planes_in); i++) {
153 s->planes_in[i].width = s->planes_in[0].width >> in_sw;
154 s->planes_in[i].height = s->planes_in[0].height >> in_sh;
155 s->planes_out[i].width = s->planes_out[0].width >> out_sw;
156 s->planes_out[i].height = s->planes_out[0].height >> out_sh;
159 out_ref = av_hwframe_ctx_alloc(device_ctx);
161 return AVERROR(ENOMEM);
162 out_ctx = (AVHWFramesContext*)out_ref->data;
164 out_ctx->format = AV_PIX_FMT_CUDA;
165 out_ctx->sw_format = s->out_fmt;
166 out_ctx->width = FFALIGN(s->planes_out[0].width, 32);
167 out_ctx->height = FFALIGN(s->planes_out[0].height, 32);
169 ret = av_hwframe_ctx_init(out_ref);
173 av_frame_unref(s->frame);
174 ret = av_hwframe_get_buffer(out_ref, s->frame, 0);
178 s->frame->width = s->planes_out[0].width;
179 s->frame->height = s->planes_out[0].height;
181 av_buffer_unref(&s->frames_ctx);
182 s->frames_ctx = out_ref;
186 av_buffer_unref(&out_ref);
190 static int format_is_supported(enum AVPixelFormat fmt)
194 for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
195 if (supported_formats[i] == fmt)
200 static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int in_height,
201 int out_width, int out_height)
203 CUDAScaleContext *s = ctx->priv;
205 AVHWFramesContext *in_frames_ctx;
207 enum AVPixelFormat in_format;
208 enum AVPixelFormat out_format;
211 /* check that we have a hw context */
212 if (!ctx->inputs[0]->hw_frames_ctx) {
213 av_log(ctx, AV_LOG_ERROR, "No hw context provided on input\n");
214 return AVERROR(EINVAL);
216 in_frames_ctx = (AVHWFramesContext*)ctx->inputs[0]->hw_frames_ctx->data;
217 in_format = in_frames_ctx->sw_format;
218 out_format = (s->format == AV_PIX_FMT_NONE) ? in_format : s->format;
220 if (!format_is_supported(in_format)) {
221 av_log(ctx, AV_LOG_ERROR, "Unsupported input format: %s\n",
222 av_get_pix_fmt_name(in_format));
223 return AVERROR(ENOSYS);
225 if (!format_is_supported(out_format)) {
226 av_log(ctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
227 av_get_pix_fmt_name(out_format));
228 return AVERROR(ENOSYS);
231 if (in_width == out_width && in_height == out_height)
234 s->in_fmt = in_format;
235 s->out_fmt = out_format;
237 s->planes_in[0].width = in_width;
238 s->planes_in[0].height = in_height;
239 s->planes_out[0].width = out_width;
240 s->planes_out[0].height = out_height;
242 ret = init_stage(s, in_frames_ctx->device_ref);
246 ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->frames_ctx);
247 if (!ctx->outputs[0]->hw_frames_ctx)
248 return AVERROR(ENOMEM);
253 static av_cold int cudascale_config_props(AVFilterLink *outlink)
255 AVFilterContext *ctx = outlink->src;
256 AVFilterLink *inlink = outlink->src->inputs[0];
257 CUDAScaleContext *s = ctx->priv;
258 AVHWFramesContext *frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
259 AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
260 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
264 extern char vf_scale_cuda_ptx[];
266 ret = CHECK_CU(cuCtxPushCurrent(cuda_ctx));
270 ret = CHECK_CU(cuModuleLoadData(&s->cu_module, vf_scale_cuda_ptx));
274 CHECK_CU(cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, "Subsample_Bilinear_uchar"));
275 CHECK_CU(cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, "Subsample_Bilinear_uchar2"));
276 CHECK_CU(cuModuleGetFunction(&s->cu_func_uchar4, s->cu_module, "Subsample_Bilinear_uchar4"));
277 CHECK_CU(cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, "Subsample_Bilinear_ushort"));
278 CHECK_CU(cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, "Subsample_Bilinear_ushort2"));
279 CHECK_CU(cuModuleGetFunction(&s->cu_func_ushort4, s->cu_module, "Subsample_Bilinear_ushort4"));
281 CHECK_CU(cuModuleGetTexRef(&s->cu_tex_uchar, s->cu_module, "uchar_tex"));
282 CHECK_CU(cuModuleGetTexRef(&s->cu_tex_uchar2, s->cu_module, "uchar2_tex"));
283 CHECK_CU(cuModuleGetTexRef(&s->cu_tex_uchar4, s->cu_module, "uchar4_tex"));
284 CHECK_CU(cuModuleGetTexRef(&s->cu_tex_ushort, s->cu_module, "ushort_tex"));
285 CHECK_CU(cuModuleGetTexRef(&s->cu_tex_ushort2, s->cu_module, "ushort2_tex"));
286 CHECK_CU(cuModuleGetTexRef(&s->cu_tex_ushort4, s->cu_module, "ushort4_tex"));
288 CHECK_CU(cuTexRefSetFlags(s->cu_tex_uchar, CU_TRSF_READ_AS_INTEGER));
289 CHECK_CU(cuTexRefSetFlags(s->cu_tex_uchar2, CU_TRSF_READ_AS_INTEGER));
290 CHECK_CU(cuTexRefSetFlags(s->cu_tex_uchar4, CU_TRSF_READ_AS_INTEGER));
291 CHECK_CU(cuTexRefSetFlags(s->cu_tex_ushort, CU_TRSF_READ_AS_INTEGER));
292 CHECK_CU(cuTexRefSetFlags(s->cu_tex_ushort2, CU_TRSF_READ_AS_INTEGER));
293 CHECK_CU(cuTexRefSetFlags(s->cu_tex_ushort4, CU_TRSF_READ_AS_INTEGER));
295 CHECK_CU(cuTexRefSetFilterMode(s->cu_tex_uchar, CU_TR_FILTER_MODE_LINEAR));
296 CHECK_CU(cuTexRefSetFilterMode(s->cu_tex_uchar2, CU_TR_FILTER_MODE_LINEAR));
297 CHECK_CU(cuTexRefSetFilterMode(s->cu_tex_uchar4, CU_TR_FILTER_MODE_LINEAR));
298 CHECK_CU(cuTexRefSetFilterMode(s->cu_tex_ushort, CU_TR_FILTER_MODE_LINEAR));
299 CHECK_CU(cuTexRefSetFilterMode(s->cu_tex_ushort2, CU_TR_FILTER_MODE_LINEAR));
300 CHECK_CU(cuTexRefSetFilterMode(s->cu_tex_ushort4, CU_TR_FILTER_MODE_LINEAR));
302 CHECK_CU(cuCtxPopCurrent(&dummy));
304 if ((ret = ff_scale_eval_dimensions(s,
305 s->w_expr, s->h_expr,
310 if (((int64_t)h * inlink->w) > INT_MAX ||
311 ((int64_t)w * inlink->h) > INT_MAX)
312 av_log(ctx, AV_LOG_ERROR, "Rescaled value for width or height is too big.\n");
317 ret = init_processing_chain(ctx, inlink->w, inlink->h, w, h);
321 av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d\n",
322 inlink->w, inlink->h, outlink->w, outlink->h);
324 if (inlink->sample_aspect_ratio.num) {
325 outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w,
326 outlink->w*inlink->h},
327 inlink->sample_aspect_ratio);
329 outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
338 static int call_resize_kernel(CUDAScaleContext *ctx, CUfunction func, CUtexref tex, int channels,
339 uint8_t *src_dptr, int src_width, int src_height, int src_pitch,
340 uint8_t *dst_dptr, int dst_width, int dst_height, int dst_pitch,
343 CUdeviceptr src_devptr = (CUdeviceptr)src_dptr;
344 CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr;
345 void *args_uchar[] = { &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height };
346 CUDA_ARRAY_DESCRIPTOR desc;
348 desc.Width = src_width;
349 desc.Height = src_height;
350 desc.NumChannels = channels;
351 if (pixel_size == 1) {
352 desc.Format = CU_AD_FORMAT_UNSIGNED_INT8;
354 desc.Format = CU_AD_FORMAT_UNSIGNED_INT16;
357 CHECK_CU(cuTexRefSetAddress2D_v3(tex, &desc, src_devptr, src_pitch * pixel_size));
358 CHECK_CU(cuLaunchKernel(func, DIV_UP(dst_width, BLOCKX), DIV_UP(dst_height, BLOCKY), 1,
359 BLOCKX, BLOCKY, 1, 0, 0, args_uchar, NULL));
364 static int scalecuda_resize(AVFilterContext *ctx,
365 AVFrame *out, AVFrame *in)
367 AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
368 CUDAScaleContext *s = ctx->priv;
370 switch (in_frames_ctx->sw_format) {
371 case AV_PIX_FMT_YUV420P:
372 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
373 in->data[0], in->width, in->height, in->linesize[0],
374 out->data[0], out->width, out->height, out->linesize[0],
376 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
377 in->data[0]+in->linesize[0]*in->height, in->width/2, in->height/2, in->linesize[0]/2,
378 out->data[0]+out->linesize[0]*out->height, out->width/2, out->height/2, out->linesize[0]/2,
380 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
381 in->data[0]+ ALIGN_UP((in->linesize[0]*in->height*5)/4, s->tex_alignment), in->width/2, in->height/2, in->linesize[0]/2,
382 out->data[0]+(out->linesize[0]*out->height*5)/4, out->width/2, out->height/2, out->linesize[0]/2,
385 case AV_PIX_FMT_YUV444P:
386 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
387 in->data[0], in->width, in->height, in->linesize[0],
388 out->data[0], out->width, out->height, out->linesize[0],
390 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
391 in->data[0]+in->linesize[0]*in->height, in->width, in->height, in->linesize[0],
392 out->data[0]+out->linesize[0]*out->height, out->width, out->height, out->linesize[0],
394 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
395 in->data[0]+in->linesize[0]*in->height*2, in->width, in->height, in->linesize[0],
396 out->data[0]+out->linesize[0]*out->height*2, out->width, out->height, out->linesize[0],
399 case AV_PIX_FMT_NV12:
400 call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
401 in->data[0], in->width, in->height, in->linesize[0],
402 out->data[0], out->width, out->height, out->linesize[0],
404 call_resize_kernel(s, s->cu_func_uchar2, s->cu_tex_uchar2, 2,
405 in->data[1], in->width/2, in->height/2, in->linesize[1],
406 out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width/2, out->height/2, out->linesize[1]/2,
409 case AV_PIX_FMT_P010LE:
410 call_resize_kernel(s, s->cu_func_ushort, s->cu_tex_ushort, 1,
411 in->data[0], in->width, in->height, in->linesize[0]/2,
412 out->data[0], out->width, out->height, out->linesize[0]/2,
414 call_resize_kernel(s, s->cu_func_ushort2, s->cu_tex_ushort2, 2,
415 in->data[1], in->width / 2, in->height / 2, in->linesize[1]/2,
416 out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
419 case AV_PIX_FMT_P016LE:
420 call_resize_kernel(s, s->cu_func_ushort, s->cu_tex_ushort, 1,
421 in->data[0], in->width, in->height, in->linesize[0] / 2,
422 out->data[0], out->width, out->height, out->linesize[0] / 2,
424 call_resize_kernel(s, s->cu_func_ushort2, s->cu_tex_ushort2, 2,
425 in->data[1], in->width / 2, in->height / 2, in->linesize[1] / 2,
426 out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
436 static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
438 CUDAScaleContext *s = ctx->priv;
442 ret = scalecuda_resize(ctx, s->frame, src);
447 ret = av_hwframe_get_buffer(src->hw_frames_ctx, s->tmp_frame, 0);
451 av_frame_move_ref(out, s->frame);
452 av_frame_move_ref(s->frame, s->tmp_frame);
454 ret = av_frame_copy_props(out, in);
461 static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in)
463 AVFilterContext *ctx = link->dst;
464 CUDAScaleContext *s = ctx->priv;
465 AVFilterLink *outlink = ctx->outputs[0];
466 AVHWFramesContext *frames_ctx = (AVHWFramesContext*)s->frames_ctx->data;
467 AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
473 out = av_frame_alloc();
475 ret = AVERROR(ENOMEM);
479 ret = CHECK_CU(cuCtxPushCurrent(device_hwctx->cuda_ctx));
483 ret = cudascale_scale(ctx, out, in);
485 CHECK_CU(cuCtxPopCurrent(&dummy));
489 av_reduce(&out->sample_aspect_ratio.num, &out->sample_aspect_ratio.den,
490 (int64_t)in->sample_aspect_ratio.num * outlink->h * link->w,
491 (int64_t)in->sample_aspect_ratio.den * outlink->w * link->h,
495 return ff_filter_frame(outlink, out);
502 #define OFFSET(x) offsetof(CUDAScaleContext, x)
503 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM)
504 static const AVOption options[] = {
505 { "w", "Output video width", OFFSET(w_expr), AV_OPT_TYPE_STRING, { .str = "iw" }, .flags = FLAGS },
506 { "h", "Output video height", OFFSET(h_expr), AV_OPT_TYPE_STRING, { .str = "ih" }, .flags = FLAGS },
510 static const AVClass cudascale_class = {
511 .class_name = "cudascale",
512 .item_name = av_default_item_name,
514 .version = LIBAVUTIL_VERSION_INT,
517 static const AVFilterPad cudascale_inputs[] = {
520 .type = AVMEDIA_TYPE_VIDEO,
521 .filter_frame = cudascale_filter_frame,
526 static const AVFilterPad cudascale_outputs[] = {
529 .type = AVMEDIA_TYPE_VIDEO,
530 .config_props = cudascale_config_props,
535 AVFilter ff_vf_scale_cuda = {
536 .name = "scale_cuda",
537 .description = NULL_IF_CONFIG_SMALL("GPU accelerated video resizer"),
539 .init = cudascale_init,
540 .uninit = cudascale_uninit,
541 .query_formats = cudascale_query_formats,
543 .priv_size = sizeof(CUDAScaleContext),
544 .priv_class = &cudascale_class,
546 .inputs = cudascale_inputs,
547 .outputs = cudascale_outputs,
549 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,