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_internal.h"
31 #include "libavutil/cuda_check.h"
32 #include "libavutil/internal.h"
33 #include "libavutil/opt.h"
34 #include "libavutil/pixdesc.h"
39 #include "scale_eval.h"
42 #include "vf_scale_cuda.h"
44 static const enum AVPixelFormat supported_formats[] = {
55 #define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
59 #define CHECK_CU(x) FF_CUDA_CHECK_DL(ctx, s->hwctx->internal->cuda_dl, x)
72 typedef struct CUDAScaleContext {
75 AVCUDADeviceContext *hwctx;
77 enum AVPixelFormat in_fmt;
78 enum AVPixelFormat out_fmt;
80 AVBufferRef *frames_ctx;
87 * Output sw format. AV_PIX_FMT_NONE for no conversion.
89 enum AVPixelFormat format;
91 char *w_expr; ///< width expression string
92 char *h_expr; ///< height expression string
94 int force_original_aspect_ratio;
95 int force_divisible_by;
99 CUfunction cu_func_uchar;
100 CUfunction cu_func_uchar2;
101 CUfunction cu_func_uchar4;
102 CUfunction cu_func_ushort;
103 CUfunction cu_func_ushort2;
104 CUfunction cu_func_ushort4;
107 CUdeviceptr srcBuffer;
108 CUdeviceptr dstBuffer;
112 int interp_use_linear;
113 int interp_as_integer;
118 static av_cold int cudascale_init(AVFilterContext *ctx)
120 CUDAScaleContext *s = ctx->priv;
122 s->format = AV_PIX_FMT_NONE;
123 s->frame = av_frame_alloc();
125 return AVERROR(ENOMEM);
127 s->tmp_frame = av_frame_alloc();
129 return AVERROR(ENOMEM);
134 static av_cold void cudascale_uninit(AVFilterContext *ctx)
136 CUDAScaleContext *s = ctx->priv;
138 if (s->hwctx && s->cu_module) {
139 CudaFunctions *cu = s->hwctx->internal->cuda_dl;
142 CHECK_CU(cu->cuCtxPushCurrent(s->hwctx->cuda_ctx));
143 CHECK_CU(cu->cuModuleUnload(s->cu_module));
145 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
148 av_frame_free(&s->frame);
149 av_buffer_unref(&s->frames_ctx);
150 av_frame_free(&s->tmp_frame);
153 static int cudascale_query_formats(AVFilterContext *ctx)
155 static const enum AVPixelFormat pixel_formats[] = {
156 AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE,
158 AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats);
160 return AVERROR(ENOMEM);
162 return ff_set_common_formats(ctx, pix_fmts);
165 static av_cold int init_hwframe_ctx(CUDAScaleContext *s, AVBufferRef *device_ctx, int width, int height)
167 AVBufferRef *out_ref = NULL;
168 AVHWFramesContext *out_ctx;
171 out_ref = av_hwframe_ctx_alloc(device_ctx);
173 return AVERROR(ENOMEM);
174 out_ctx = (AVHWFramesContext*)out_ref->data;
176 out_ctx->format = AV_PIX_FMT_CUDA;
177 out_ctx->sw_format = s->out_fmt;
178 out_ctx->width = FFALIGN(width, 32);
179 out_ctx->height = FFALIGN(height, 32);
181 ret = av_hwframe_ctx_init(out_ref);
185 av_frame_unref(s->frame);
186 ret = av_hwframe_get_buffer(out_ref, s->frame, 0);
190 s->frame->width = width;
191 s->frame->height = height;
193 av_buffer_unref(&s->frames_ctx);
194 s->frames_ctx = out_ref;
198 av_buffer_unref(&out_ref);
202 static int format_is_supported(enum AVPixelFormat fmt)
206 for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
207 if (supported_formats[i] == fmt)
212 static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int in_height,
213 int out_width, int out_height)
215 CUDAScaleContext *s = ctx->priv;
217 AVHWFramesContext *in_frames_ctx;
219 enum AVPixelFormat in_format;
220 enum AVPixelFormat out_format;
223 /* check that we have a hw context */
224 if (!ctx->inputs[0]->hw_frames_ctx) {
225 av_log(ctx, AV_LOG_ERROR, "No hw context provided on input\n");
226 return AVERROR(EINVAL);
228 in_frames_ctx = (AVHWFramesContext*)ctx->inputs[0]->hw_frames_ctx->data;
229 in_format = in_frames_ctx->sw_format;
230 out_format = (s->format == AV_PIX_FMT_NONE) ? in_format : s->format;
232 if (!format_is_supported(in_format)) {
233 av_log(ctx, AV_LOG_ERROR, "Unsupported input format: %s\n",
234 av_get_pix_fmt_name(in_format));
235 return AVERROR(ENOSYS);
237 if (!format_is_supported(out_format)) {
238 av_log(ctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
239 av_get_pix_fmt_name(out_format));
240 return AVERROR(ENOSYS);
243 s->in_fmt = in_format;
244 s->out_fmt = out_format;
246 if (s->passthrough && in_width == out_width && in_height == out_height && in_format == out_format) {
247 s->frames_ctx = av_buffer_ref(ctx->inputs[0]->hw_frames_ctx);
249 return AVERROR(ENOMEM);
253 ret = init_hwframe_ctx(s, in_frames_ctx->device_ref, out_width, out_height);
258 ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->frames_ctx);
259 if (!ctx->outputs[0]->hw_frames_ctx)
260 return AVERROR(ENOMEM);
265 static av_cold int cudascale_config_props(AVFilterLink *outlink)
267 AVFilterContext *ctx = outlink->src;
268 AVFilterLink *inlink = outlink->src->inputs[0];
269 CUDAScaleContext *s = ctx->priv;
270 AVHWFramesContext *frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
271 AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
272 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
273 CudaFunctions *cu = device_hwctx->internal->cuda_dl;
279 const char *function_infix = "";
281 extern char vf_scale_cuda_ptx[];
282 extern char vf_scale_cuda_bicubic_ptx[];
284 switch(s->interp_algo) {
285 case INTERP_ALGO_NEAREST:
286 scaler_ptx = vf_scale_cuda_ptx;
287 function_infix = "_Nearest";
288 s->interp_use_linear = 0;
289 s->interp_as_integer = 1;
291 case INTERP_ALGO_BILINEAR:
292 scaler_ptx = vf_scale_cuda_ptx;
293 function_infix = "_Bilinear";
294 s->interp_use_linear = 1;
295 s->interp_as_integer = 1;
297 case INTERP_ALGO_DEFAULT:
298 case INTERP_ALGO_BICUBIC:
299 scaler_ptx = vf_scale_cuda_bicubic_ptx;
300 function_infix = "_Bicubic";
301 s->interp_use_linear = 0;
302 s->interp_as_integer = 0;
304 case INTERP_ALGO_LANCZOS:
305 scaler_ptx = vf_scale_cuda_bicubic_ptx;
306 function_infix = "_Lanczos";
307 s->interp_use_linear = 0;
308 s->interp_as_integer = 0;
311 av_log(ctx, AV_LOG_ERROR, "Unknown interpolation algorithm\n");
315 s->hwctx = device_hwctx;
316 s->cu_stream = s->hwctx->stream;
318 ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_ctx));
322 ret = CHECK_CU(cu->cuModuleLoadData(&s->cu_module, scaler_ptx));
326 snprintf(buf, sizeof(buf), "Subsample%s_uchar", function_infix);
327 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, buf));
331 snprintf(buf, sizeof(buf), "Subsample%s_uchar2", function_infix);
332 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, buf));
336 snprintf(buf, sizeof(buf), "Subsample%s_uchar4", function_infix);
337 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar4, s->cu_module, buf));
341 snprintf(buf, sizeof(buf), "Subsample%s_ushort", function_infix);
342 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, buf));
346 snprintf(buf, sizeof(buf), "Subsample%s_ushort2", function_infix);
347 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, buf));
351 snprintf(buf, sizeof(buf), "Subsample%s_ushort4", function_infix);
352 CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort4, s->cu_module, buf));
357 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
359 if ((ret = ff_scale_eval_dimensions(s,
360 s->w_expr, s->h_expr,
365 ff_scale_adjust_dimensions(inlink, &w, &h,
366 s->force_original_aspect_ratio, s->force_divisible_by);
368 if (((int64_t)h * inlink->w) > INT_MAX ||
369 ((int64_t)w * inlink->h) > INT_MAX)
370 av_log(ctx, AV_LOG_ERROR, "Rescaled value for width or height is too big.\n");
375 ret = init_processing_chain(ctx, inlink->w, inlink->h, w, h);
379 av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d%s\n",
380 inlink->w, inlink->h, outlink->w, outlink->h, s->passthrough ? " (passthrough)" : "");
382 if (inlink->sample_aspect_ratio.num) {
383 outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w,
384 outlink->w*inlink->h},
385 inlink->sample_aspect_ratio);
387 outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
396 static int call_resize_kernel(AVFilterContext *ctx, CUfunction func, int channels,
397 uint8_t *src_dptr, int src_width, int src_height, int src_pitch,
398 uint8_t *dst_dptr, int dst_width, int dst_height, int dst_pitch,
399 int pixel_size, int bit_depth)
401 CUDAScaleContext *s = ctx->priv;
402 CudaFunctions *cu = s->hwctx->internal->cuda_dl;
403 CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr;
405 void *args_uchar[] = { &tex, &dst_devptr, &dst_width, &dst_height, &dst_pitch,
406 &src_width, &src_height, &bit_depth, &s->param };
409 CUDA_TEXTURE_DESC tex_desc = {
410 .filterMode = s->interp_use_linear ?
411 CU_TR_FILTER_MODE_LINEAR :
412 CU_TR_FILTER_MODE_POINT,
413 .flags = s->interp_as_integer ? CU_TRSF_READ_AS_INTEGER : 0,
416 CUDA_RESOURCE_DESC res_desc = {
417 .resType = CU_RESOURCE_TYPE_PITCH2D,
418 .res.pitch2D.format = pixel_size == 1 ?
419 CU_AD_FORMAT_UNSIGNED_INT8 :
420 CU_AD_FORMAT_UNSIGNED_INT16,
421 .res.pitch2D.numChannels = channels,
422 .res.pitch2D.width = src_width,
423 .res.pitch2D.height = src_height,
424 .res.pitch2D.pitchInBytes = src_pitch,
425 .res.pitch2D.devPtr = (CUdeviceptr)src_dptr,
428 // Handling of channels is done via vector-types in cuda, so their size is implicitly part of the pitch
429 // Same for pixel_size, which is represented via datatypes on the cuda side of things.
430 dst_pitch /= channels * pixel_size;
432 ret = CHECK_CU(cu->cuTexObjectCreate(&tex, &res_desc, &tex_desc, NULL));
436 ret = CHECK_CU(cu->cuLaunchKernel(func,
437 DIV_UP(dst_width, BLOCKX), DIV_UP(dst_height, BLOCKY), 1,
438 BLOCKX, BLOCKY, 1, 0, s->cu_stream, args_uchar, NULL));
442 CHECK_CU(cu->cuTexObjectDestroy(tex));
447 static int scalecuda_resize(AVFilterContext *ctx,
448 AVFrame *out, AVFrame *in)
450 AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
451 CUDAScaleContext *s = ctx->priv;
453 switch (in_frames_ctx->sw_format) {
454 case AV_PIX_FMT_YUV420P:
455 call_resize_kernel(ctx, s->cu_func_uchar, 1,
456 in->data[0], in->width, in->height, in->linesize[0],
457 out->data[0], out->width, out->height, out->linesize[0],
459 call_resize_kernel(ctx, s->cu_func_uchar, 1,
460 in->data[1], in->width / 2, in->height / 2, in->linesize[1],
461 out->data[1], out->width / 2, out->height / 2, out->linesize[1],
463 call_resize_kernel(ctx, s->cu_func_uchar, 1,
464 in->data[2], in->width / 2, in->height / 2, in->linesize[2],
465 out->data[2], out->width / 2, out->height / 2, out->linesize[2],
468 case AV_PIX_FMT_YUV444P:
469 call_resize_kernel(ctx, s->cu_func_uchar, 1,
470 in->data[0], in->width, in->height, in->linesize[0],
471 out->data[0], out->width, out->height, out->linesize[0],
473 call_resize_kernel(ctx, s->cu_func_uchar, 1,
474 in->data[1], in->width, in->height, in->linesize[1],
475 out->data[1], out->width, out->height, out->linesize[1],
477 call_resize_kernel(ctx, s->cu_func_uchar, 1,
478 in->data[2], in->width, in->height, in->linesize[2],
479 out->data[2], out->width, out->height, out->linesize[2],
482 case AV_PIX_FMT_YUV444P16:
483 call_resize_kernel(ctx, s->cu_func_ushort, 1,
484 in->data[0], in->width, in->height, in->linesize[0],
485 out->data[0], out->width, out->height, out->linesize[0],
487 call_resize_kernel(ctx, s->cu_func_ushort, 1,
488 in->data[1], in->width, in->height, in->linesize[1],
489 out->data[1], out->width, out->height, out->linesize[1],
491 call_resize_kernel(ctx, s->cu_func_ushort, 1,
492 in->data[2], in->width, in->height, in->linesize[2],
493 out->data[2], out->width, out->height, out->linesize[2],
496 case AV_PIX_FMT_NV12:
497 call_resize_kernel(ctx, s->cu_func_uchar, 1,
498 in->data[0], in->width, in->height, in->linesize[0],
499 out->data[0], out->width, out->height, out->linesize[0],
501 call_resize_kernel(ctx, s->cu_func_uchar2, 2,
502 in->data[1], in->width / 2, in->height / 2, in->linesize[1],
503 out->data[1], out->width / 2, out->height / 2, out->linesize[1],
506 case AV_PIX_FMT_P010LE:
507 call_resize_kernel(ctx, s->cu_func_ushort, 1,
508 in->data[0], in->width, in->height, in->linesize[0],
509 out->data[0], out->width, out->height, out->linesize[0],
511 call_resize_kernel(ctx, s->cu_func_ushort2, 2,
512 in->data[1], in->width / 2, in->height / 2, in->linesize[1],
513 out->data[1], out->width / 2, out->height / 2, out->linesize[1],
516 case AV_PIX_FMT_P016LE:
517 call_resize_kernel(ctx, s->cu_func_ushort, 1,
518 in->data[0], in->width, in->height, in->linesize[0],
519 out->data[0], out->width, out->height, out->linesize[0],
521 call_resize_kernel(ctx, s->cu_func_ushort2, 2,
522 in->data[1], in->width / 2, in->height / 2, in->linesize[1],
523 out->data[1], out->width / 2, out->height / 2, out->linesize[1],
526 case AV_PIX_FMT_0RGB32:
527 case AV_PIX_FMT_0BGR32:
528 call_resize_kernel(ctx, s->cu_func_uchar4, 4,
529 in->data[0], in->width, in->height, in->linesize[0],
530 out->data[0], out->width, out->height, out->linesize[0],
540 static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
542 CUDAScaleContext *s = ctx->priv;
543 AVFilterLink *outlink = ctx->outputs[0];
547 ret = scalecuda_resize(ctx, s->frame, src);
552 ret = av_hwframe_get_buffer(src->hw_frames_ctx, s->tmp_frame, 0);
556 av_frame_move_ref(out, s->frame);
557 av_frame_move_ref(s->frame, s->tmp_frame);
559 s->frame->width = outlink->w;
560 s->frame->height = outlink->h;
562 ret = av_frame_copy_props(out, in);
569 static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in)
571 AVFilterContext *ctx = link->dst;
572 CUDAScaleContext *s = ctx->priv;
573 AVFilterLink *outlink = ctx->outputs[0];
574 CudaFunctions *cu = s->hwctx->internal->cuda_dl;
581 return ff_filter_frame(outlink, in);
583 out = av_frame_alloc();
585 ret = AVERROR(ENOMEM);
589 ret = CHECK_CU(cu->cuCtxPushCurrent(s->hwctx->cuda_ctx));
593 ret = cudascale_scale(ctx, out, in);
595 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
599 av_reduce(&out->sample_aspect_ratio.num, &out->sample_aspect_ratio.den,
600 (int64_t)in->sample_aspect_ratio.num * outlink->h * link->w,
601 (int64_t)in->sample_aspect_ratio.den * outlink->w * link->h,
605 return ff_filter_frame(outlink, out);
612 static AVFrame *cudascale_get_video_buffer(AVFilterLink *inlink, int w, int h)
614 CUDAScaleContext *s = inlink->dst->priv;
616 return s->passthrough ?
617 ff_null_get_video_buffer (inlink, w, h) :
618 ff_default_get_video_buffer(inlink, w, h);
621 #define OFFSET(x) offsetof(CUDAScaleContext, x)
622 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM)
623 static const AVOption options[] = {
624 { "w", "Output video width", OFFSET(w_expr), AV_OPT_TYPE_STRING, { .str = "iw" }, .flags = FLAGS },
625 { "h", "Output video height", OFFSET(h_expr), AV_OPT_TYPE_STRING, { .str = "ih" }, .flags = FLAGS },
626 { "interp_algo", "Interpolation algorithm used for resizing", OFFSET(interp_algo), AV_OPT_TYPE_INT, { .i64 = INTERP_ALGO_DEFAULT }, 0, INTERP_ALGO_COUNT - 1, FLAGS, "interp_algo" },
627 { "nearest", "nearest neighbour", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_NEAREST }, 0, 0, FLAGS, "interp_algo" },
628 { "bilinear", "bilinear", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BILINEAR }, 0, 0, FLAGS, "interp_algo" },
629 { "bicubic", "bicubic", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BICUBIC }, 0, 0, FLAGS, "interp_algo" },
630 { "lanczos", "lanczos", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_LANCZOS }, 0, 0, FLAGS, "interp_algo" },
631 { "passthrough", "Do not process frames at all if parameters match", OFFSET(passthrough), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, FLAGS },
632 { "param", "Algorithm-Specific parameter", OFFSET(param), AV_OPT_TYPE_FLOAT, { .dbl = SCALE_CUDA_PARAM_DEFAULT }, -FLT_MAX, FLT_MAX, FLAGS },
633 { "force_original_aspect_ratio", "decrease or increase w/h if necessary to keep the original AR", OFFSET(force_original_aspect_ratio), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 2, FLAGS, "force_oar" },
634 { "disable", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 0 }, 0, 0, FLAGS, "force_oar" },
635 { "decrease", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 1 }, 0, 0, FLAGS, "force_oar" },
636 { "increase", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 2 }, 0, 0, FLAGS, "force_oar" },
637 { "force_divisible_by", "enforce that the output resolution is divisible by a defined integer when force_original_aspect_ratio is used", OFFSET(force_divisible_by), AV_OPT_TYPE_INT, { .i64 = 1 }, 1, 256, FLAGS },
641 static const AVClass cudascale_class = {
642 .class_name = "cudascale",
643 .item_name = av_default_item_name,
645 .version = LIBAVUTIL_VERSION_INT,
648 static const AVFilterPad cudascale_inputs[] = {
651 .type = AVMEDIA_TYPE_VIDEO,
652 .filter_frame = cudascale_filter_frame,
653 .get_video_buffer = cudascale_get_video_buffer,
658 static const AVFilterPad cudascale_outputs[] = {
661 .type = AVMEDIA_TYPE_VIDEO,
662 .config_props = cudascale_config_props,
667 AVFilter ff_vf_scale_cuda = {
668 .name = "scale_cuda",
669 .description = NULL_IF_CONFIG_SMALL("GPU accelerated video resizer"),
671 .init = cudascale_init,
672 .uninit = cudascale_uninit,
673 .query_formats = cudascale_query_formats,
675 .priv_size = sizeof(CUDAScaleContext),
676 .priv_class = &cudascale_class,
678 .inputs = cudascale_inputs,
679 .outputs = cudascale_outputs,
681 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,