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.
26 #include "libavutil/avstring.h"
27 #include "libavutil/common.h"
28 #include "libavutil/hwcontext.h"
29 #include "libavutil/hwcontext_cuda_internal.h"
30 #include "libavutil/cuda_check.h"
31 #include "libavutil/internal.h"
32 #include "libavutil/opt.h"
33 #include "libavutil/pixdesc.h"
41 static const enum AVPixelFormat supported_formats[] = {
49 #define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
50 #define ALIGN_UP(a, b) (((a) + (b) - 1) & ~((b) - 1))
55 #define CHECK_CU(x) FF_CUDA_CHECK_DL(ctx, s->hwctx->internal->cuda_dl, x)
57 typedef struct CUDAScaleContext {
60 AVCUDADeviceContext *hwctx;
62 enum AVPixelFormat in_fmt;
63 enum AVPixelFormat out_fmt;
68 } planes_in[3], planes_out[3];
70 AVBufferRef *frames_ctx;
77 * Output sw format. AV_PIX_FMT_NONE for no conversion.
79 enum AVPixelFormat format;
81 char *w_expr; ///< width expression string
82 char *h_expr; ///< height expression string
86 CUfunction cu_func_uchar;
87 CUfunction cu_func_uchar2;
88 CUfunction cu_func_uchar4;
89 CUfunction cu_func_ushort;
90 CUfunction cu_func_ushort2;
91 CUfunction cu_func_ushort4;
94 CUdeviceptr srcBuffer;
95 CUdeviceptr dstBuffer;
99 static av_cold int cudascale_init(AVFilterContext *ctx)
101 CUDAScaleContext *s = ctx->priv;
103 s->format = AV_PIX_FMT_NONE;
104 s->frame = av_frame_alloc();
106 return AVERROR(ENOMEM);
108 s->tmp_frame = av_frame_alloc();
110 return AVERROR(ENOMEM);
115 static av_cold void cudascale_uninit(AVFilterContext *ctx)
117 CUDAScaleContext *s = ctx->priv;
119 av_frame_free(&s->frame);
120 av_buffer_unref(&s->frames_ctx);
121 av_frame_free(&s->tmp_frame);
124 static int cudascale_query_formats(AVFilterContext *ctx)
126 static const enum AVPixelFormat pixel_formats[] = {
127 AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE,
129 AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats);
131 return ff_set_common_formats(ctx, pix_fmts);
134 static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
136 AVBufferRef *out_ref = NULL;
137 AVHWFramesContext *out_ctx;
138 int in_sw, in_sh, out_sw, out_sh;
141 av_pix_fmt_get_chroma_sub_sample(s->in_fmt, &in_sw, &in_sh);
142 av_pix_fmt_get_chroma_sub_sample(s->out_fmt, &out_sw, &out_sh);
143 if (!s->planes_out[0].width) {
144 s->planes_out[0].width = s->planes_in[0].width;
145 s->planes_out[0].height = s->planes_in[0].height;
148 for (i = 1; i < FF_ARRAY_ELEMS(s->planes_in); i++) {
149 s->planes_in[i].width = s->planes_in[0].width >> in_sw;
150 s->planes_in[i].height = s->planes_in[0].height >> in_sh;
151 s->planes_out[i].width = s->planes_out[0].width >> out_sw;
152 s->planes_out[i].height = s->planes_out[0].height >> out_sh;
155 out_ref = av_hwframe_ctx_alloc(device_ctx);
157 return AVERROR(ENOMEM);
158 out_ctx = (AVHWFramesContext*)out_ref->data;
160 out_ctx->format = AV_PIX_FMT_CUDA;
161 out_ctx->sw_format = s->out_fmt;
162 out_ctx->width = FFALIGN(s->planes_out[0].width, 32);
163 out_ctx->height = FFALIGN(s->planes_out[0].height, 32);
165 ret = av_hwframe_ctx_init(out_ref);
169 av_frame_unref(s->frame);
170 ret = av_hwframe_get_buffer(out_ref, s->frame, 0);
174 s->frame->width = s->planes_out[0].width;
175 s->frame->height = s->planes_out[0].height;
177 av_buffer_unref(&s->frames_ctx);
178 s->frames_ctx = out_ref;
182 av_buffer_unref(&out_ref);
186 static int format_is_supported(enum AVPixelFormat fmt)
190 for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
191 if (supported_formats[i] == fmt)
196 static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int in_height,
197 int out_width, int out_height)
199 CUDAScaleContext *s = ctx->priv;
201 AVHWFramesContext *in_frames_ctx;
203 enum AVPixelFormat in_format;
204 enum AVPixelFormat out_format;
207 /* check that we have a hw context */
208 if (!ctx->inputs[0]->hw_frames_ctx) {
209 av_log(ctx, AV_LOG_ERROR, "No hw context provided on input\n");
210 return AVERROR(EINVAL);
212 in_frames_ctx = (AVHWFramesContext*)ctx->inputs[0]->hw_frames_ctx->data;
213 in_format = in_frames_ctx->sw_format;
214 out_format = (s->format == AV_PIX_FMT_NONE) ? in_format : s->format;
216 if (!format_is_supported(in_format)) {
217 av_log(ctx, AV_LOG_ERROR, "Unsupported input format: %s\n",
218 av_get_pix_fmt_name(in_format));
219 return AVERROR(ENOSYS);
221 if (!format_is_supported(out_format)) {
222 av_log(ctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
223 av_get_pix_fmt_name(out_format));
224 return AVERROR(ENOSYS);
227 if (in_width == out_width && in_height == out_height)
230 s->in_fmt = in_format;
231 s->out_fmt = out_format;
233 s->planes_in[0].width = in_width;
234 s->planes_in[0].height = in_height;
235 s->planes_out[0].width = out_width;
236 s->planes_out[0].height = out_height;
238 ret = init_stage(s, in_frames_ctx->device_ref);
242 ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->frames_ctx);
243 if (!ctx->outputs[0]->hw_frames_ctx)
244 return AVERROR(ENOMEM);
249 static av_cold int cudascale_config_props(AVFilterLink *outlink)
251 AVFilterContext *ctx = outlink->src;
252 AVFilterLink *inlink = outlink->src->inputs[0];
253 CUDAScaleContext *s = ctx->priv;
254 AVHWFramesContext *frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
255 AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
256 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
257 CudaFunctions *cu = device_hwctx->internal->cuda_dl;
261 extern char vf_scale_cuda_ptx[];
263 s->hwctx = device_hwctx;
264 s->cu_stream = s->hwctx->stream;
266 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_ctx));
270 ret = CHECK_CU(cu->cuModuleLoadData(&s->cu_module, vf_scale_cuda_ptx));
274 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, "Subsample_Bilinear_uchar"));
278 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, "Subsample_Bilinear_uchar2"));
282 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar4, s->cu_module, "Subsample_Bilinear_uchar4"));
286 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, "Subsample_Bilinear_ushort"));
290 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, "Subsample_Bilinear_ushort2"));
294 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort4, s->cu_module, "Subsample_Bilinear_ushort4"));
299 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
301 if ((ret = ff_scale_eval_dimensions(s,
302 s->w_expr, s->h_expr,
307 if (((int64_t)h * inlink->w) > INT_MAX ||
308 ((int64_t)w * inlink->h) > INT_MAX)
309 av_log(ctx, AV_LOG_ERROR, "Rescaled value for width or height is too big.\n");
314 ret = init_processing_chain(ctx, inlink->w, inlink->h, w, h);
318 av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d\n",
319 inlink->w, inlink->h, outlink->w, outlink->h);
321 if (inlink->sample_aspect_ratio.num) {
322 outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w,
323 outlink->w*inlink->h},
324 inlink->sample_aspect_ratio);
326 outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
335 static int call_resize_kernel(AVFilterContext *ctx, CUfunction func, int channels,
336 uint8_t *src_dptr, int src_width, int src_height, int src_pitch,
337 uint8_t *dst_dptr, int dst_width, int dst_height, int dst_pitch,
340 CUDAScaleContext *s = ctx->priv;
341 CudaFunctions *cu = s->hwctx->internal->cuda_dl;
342 CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr;
344 void *args_uchar[] = { &tex, &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height };
347 CUDA_TEXTURE_DESC tex_desc = {
348 .filterMode = CU_TR_FILTER_MODE_LINEAR,
349 .flags = CU_TRSF_READ_AS_INTEGER,
352 CUDA_RESOURCE_DESC res_desc = {
353 .resType = CU_RESOURCE_TYPE_PITCH2D,
354 .res.pitch2D.format = pixel_size == 1 ?
355 CU_AD_FORMAT_UNSIGNED_INT8 :
356 CU_AD_FORMAT_UNSIGNED_INT16,
357 .res.pitch2D.numChannels = channels,
358 .res.pitch2D.width = src_width,
359 .res.pitch2D.height = src_height,
360 .res.pitch2D.pitchInBytes = src_pitch * pixel_size,
361 .res.pitch2D.devPtr = (CUdeviceptr)src_dptr,
364 ret = CHECK_CU(cu->cuTexObjectCreate(&tex, &res_desc, &tex_desc, NULL));
368 ret = CHECK_CU(cu->cuLaunchKernel(func,
369 DIV_UP(dst_width, BLOCKX), DIV_UP(dst_height, BLOCKY), 1,
370 BLOCKX, BLOCKY, 1, 0, s->cu_stream, args_uchar, NULL));
374 CHECK_CU(cu->cuTexObjectDestroy(tex));
379 static int scalecuda_resize(AVFilterContext *ctx,
380 AVFrame *out, AVFrame *in)
382 AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
383 CUDAScaleContext *s = ctx->priv;
385 switch (in_frames_ctx->sw_format) {
386 case AV_PIX_FMT_YUV420P:
387 call_resize_kernel(ctx, s->cu_func_uchar, 1,
388 in->data[0], in->width, in->height, in->linesize[0],
389 out->data[0], out->width, out->height, out->linesize[0],
391 call_resize_kernel(ctx, s->cu_func_uchar, 1,
392 in->data[0]+in->linesize[0]*in->height, in->width/2, in->height/2, in->linesize[0]/2,
393 out->data[0]+out->linesize[0]*out->height, out->width/2, out->height/2, out->linesize[0]/2,
395 call_resize_kernel(ctx, s->cu_func_uchar, 1,
396 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,
397 out->data[0]+(out->linesize[0]*out->height*5)/4, out->width/2, out->height/2, out->linesize[0]/2,
400 case AV_PIX_FMT_YUV444P:
401 call_resize_kernel(ctx, s->cu_func_uchar, 1,
402 in->data[0], in->width, in->height, in->linesize[0],
403 out->data[0], out->width, out->height, out->linesize[0],
405 call_resize_kernel(ctx, s->cu_func_uchar, 1,
406 in->data[0]+in->linesize[0]*in->height, in->width, in->height, in->linesize[0],
407 out->data[0]+out->linesize[0]*out->height, out->width, out->height, out->linesize[0],
409 call_resize_kernel(ctx, s->cu_func_uchar, 1,
410 in->data[0]+in->linesize[0]*in->height*2, in->width, in->height, in->linesize[0],
411 out->data[0]+out->linesize[0]*out->height*2, out->width, out->height, out->linesize[0],
414 case AV_PIX_FMT_NV12:
415 call_resize_kernel(ctx, s->cu_func_uchar, 1,
416 in->data[0], in->width, in->height, in->linesize[0],
417 out->data[0], out->width, out->height, out->linesize[0],
419 call_resize_kernel(ctx, s->cu_func_uchar2, 2,
420 in->data[1], in->width/2, in->height/2, in->linesize[1],
421 out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width/2, out->height/2, out->linesize[1]/2,
424 case AV_PIX_FMT_P010LE:
425 call_resize_kernel(ctx, s->cu_func_ushort, 1,
426 in->data[0], in->width, in->height, in->linesize[0]/2,
427 out->data[0], out->width, out->height, out->linesize[0]/2,
429 call_resize_kernel(ctx, s->cu_func_ushort2, 2,
430 in->data[1], in->width / 2, in->height / 2, in->linesize[1]/2,
431 out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
434 case AV_PIX_FMT_P016LE:
435 call_resize_kernel(ctx, s->cu_func_ushort, 1,
436 in->data[0], in->width, in->height, in->linesize[0] / 2,
437 out->data[0], out->width, out->height, out->linesize[0] / 2,
439 call_resize_kernel(ctx, s->cu_func_ushort2, 2,
440 in->data[1], in->width / 2, in->height / 2, in->linesize[1] / 2,
441 out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
451 static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
453 CUDAScaleContext *s = ctx->priv;
457 ret = scalecuda_resize(ctx, s->frame, src);
462 ret = av_hwframe_get_buffer(src->hw_frames_ctx, s->tmp_frame, 0);
466 av_frame_move_ref(out, s->frame);
467 av_frame_move_ref(s->frame, s->tmp_frame);
469 ret = av_frame_copy_props(out, in);
476 static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in)
478 AVFilterContext *ctx = link->dst;
479 CUDAScaleContext *s = ctx->priv;
480 AVFilterLink *outlink = ctx->outputs[0];
481 CudaFunctions *cu = s->hwctx->internal->cuda_dl;
487 out = av_frame_alloc();
489 ret = AVERROR(ENOMEM);
493 ret = CHECK_CU(cu->cuCtxPushCurrent(s->hwctx->cuda_ctx));
497 ret = cudascale_scale(ctx, out, in);
499 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
503 av_reduce(&out->sample_aspect_ratio.num, &out->sample_aspect_ratio.den,
504 (int64_t)in->sample_aspect_ratio.num * outlink->h * link->w,
505 (int64_t)in->sample_aspect_ratio.den * outlink->w * link->h,
509 return ff_filter_frame(outlink, out);
516 #define OFFSET(x) offsetof(CUDAScaleContext, x)
517 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM)
518 static const AVOption options[] = {
519 { "w", "Output video width", OFFSET(w_expr), AV_OPT_TYPE_STRING, { .str = "iw" }, .flags = FLAGS },
520 { "h", "Output video height", OFFSET(h_expr), AV_OPT_TYPE_STRING, { .str = "ih" }, .flags = FLAGS },
524 static const AVClass cudascale_class = {
525 .class_name = "cudascale",
526 .item_name = av_default_item_name,
528 .version = LIBAVUTIL_VERSION_INT,
531 static const AVFilterPad cudascale_inputs[] = {
534 .type = AVMEDIA_TYPE_VIDEO,
535 .filter_frame = cudascale_filter_frame,
540 static const AVFilterPad cudascale_outputs[] = {
543 .type = AVMEDIA_TYPE_VIDEO,
544 .config_props = cudascale_config_props,
549 AVFilter ff_vf_scale_cuda = {
550 .name = "scale_cuda",
551 .description = NULL_IF_CONFIG_SMALL("GPU accelerated video resizer"),
553 .init = cudascale_init,
554 .uninit = cudascale_uninit,
555 .query_formats = cudascale_query_formats,
557 .priv_size = sizeof(CUDAScaleContext),
558 .priv_class = &cudascale_class,
560 .inputs = cudascale_inputs,
561 .outputs = cudascale_outputs,
563 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,