X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavfilter%2Fvf_scale_cuda.c;h=99fef6b22b8fef2f43b1b79322b65541508460db;hb=9298e8eb61f6c796aaf7c6e14e59f345318d2753;hp=b287bd8c126a150274d2d41ed4049f029e30f81d;hpb=f1d0f83712470c0fef13b8215cccbdb77ba7f3bf;p=ffmpeg diff --git a/libavfilter/vf_scale_cuda.c b/libavfilter/vf_scale_cuda.c index b287bd8c126..99fef6b22b8 100644 --- a/libavfilter/vf_scale_cuda.c +++ b/libavfilter/vf_scale_cuda.c @@ -20,6 +20,7 @@ * DEALINGS IN THE SOFTWARE. */ +#include #include #include @@ -38,6 +39,8 @@ #include "scale_eval.h" #include "video.h" +#include "vf_scale_cuda.h" + static const enum AVPixelFormat supported_formats[] = { AV_PIX_FMT_YUV420P, AV_PIX_FMT_NV12, @@ -45,11 +48,11 @@ static const enum AVPixelFormat supported_formats[] = { AV_PIX_FMT_P010, AV_PIX_FMT_P016, AV_PIX_FMT_YUV444P16, + AV_PIX_FMT_0RGB32, + AV_PIX_FMT_0BGR32, }; #define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) ) -#define ALIGN_UP(a, b) (((a) + (b) - 1) & ~((b) - 1)) -#define NUM_BUFFERS 2 #define BLOCKX 32 #define BLOCKY 16 @@ -58,8 +61,10 @@ static const enum AVPixelFormat supported_formats[] = { enum { INTERP_ALGO_DEFAULT, + INTERP_ALGO_NEAREST, INTERP_ALGO_BILINEAR, INTERP_ALGO_BICUBIC, + INTERP_ALGO_LANCZOS, INTERP_ALGO_COUNT }; @@ -72,11 +77,6 @@ typedef struct CUDAScaleContext { enum AVPixelFormat in_fmt; enum AVPixelFormat out_fmt; - struct { - int width; - int height; - } planes_in[3], planes_out[3]; - AVBufferRef *frames_ctx; AVFrame *frame; @@ -110,6 +110,9 @@ typedef struct CUDAScaleContext { int interp_algo; int interp_use_linear; + int interp_as_integer; + + float param; } CUDAScaleContext; static av_cold int cudascale_init(AVFilterContext *ctx) @@ -153,30 +156,17 @@ static int cudascale_query_formats(AVFilterContext *ctx) AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE, }; AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats); + if (!pix_fmts) + return AVERROR(ENOMEM); return ff_set_common_formats(ctx, pix_fmts); } -static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx) +static av_cold int init_hwframe_ctx(CUDAScaleContext *s, AVBufferRef *device_ctx, int width, int height) { AVBufferRef *out_ref = NULL; AVHWFramesContext *out_ctx; - int in_sw, in_sh, out_sw, out_sh; - int ret, i; - - av_pix_fmt_get_chroma_sub_sample(s->in_fmt, &in_sw, &in_sh); - av_pix_fmt_get_chroma_sub_sample(s->out_fmt, &out_sw, &out_sh); - if (!s->planes_out[0].width) { - s->planes_out[0].width = s->planes_in[0].width; - s->planes_out[0].height = s->planes_in[0].height; - } - - for (i = 1; i < FF_ARRAY_ELEMS(s->planes_in); i++) { - s->planes_in[i].width = s->planes_in[0].width >> in_sw; - s->planes_in[i].height = s->planes_in[0].height >> in_sh; - s->planes_out[i].width = s->planes_out[0].width >> out_sw; - s->planes_out[i].height = s->planes_out[0].height >> out_sh; - } + int ret; out_ref = av_hwframe_ctx_alloc(device_ctx); if (!out_ref) @@ -185,8 +175,8 @@ static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx) out_ctx->format = AV_PIX_FMT_CUDA; out_ctx->sw_format = s->out_fmt; - out_ctx->width = FFALIGN(s->planes_out[0].width, 32); - out_ctx->height = FFALIGN(s->planes_out[0].height, 32); + out_ctx->width = FFALIGN(width, 32); + out_ctx->height = FFALIGN(height, 32); ret = av_hwframe_ctx_init(out_ref); if (ret < 0) @@ -197,8 +187,8 @@ static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx) if (ret < 0) goto fail; - s->frame->width = s->planes_out[0].width; - s->frame->height = s->planes_out[0].height; + s->frame->width = width; + s->frame->height = height; av_buffer_unref(&s->frames_ctx); s->frames_ctx = out_ref; @@ -250,20 +240,20 @@ static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int return AVERROR(ENOSYS); } - if (in_width == out_width && in_height == out_height) - s->passthrough = 1; - s->in_fmt = in_format; s->out_fmt = out_format; - s->planes_in[0].width = in_width; - s->planes_in[0].height = in_height; - s->planes_out[0].width = out_width; - s->planes_out[0].height = out_height; + if (s->passthrough && in_width == out_width && in_height == out_height && in_format == out_format) { + s->frames_ctx = av_buffer_ref(ctx->inputs[0]->hw_frames_ctx); + if (!s->frames_ctx) + return AVERROR(ENOMEM); + } else { + s->passthrough = 0; - ret = init_stage(s, in_frames_ctx->device_ref); - if (ret < 0) - return ret; + ret = init_hwframe_ctx(s, in_frames_ctx->device_ref, out_width, out_height); + if (ret < 0) + return ret; + } ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->frames_ctx); if (!ctx->outputs[0]->hw_frames_ctx) @@ -292,16 +282,30 @@ static av_cold int cudascale_config_props(AVFilterLink *outlink) extern char vf_scale_cuda_bicubic_ptx[]; switch(s->interp_algo) { + case INTERP_ALGO_NEAREST: + scaler_ptx = vf_scale_cuda_ptx; + function_infix = "_Nearest"; + s->interp_use_linear = 0; + s->interp_as_integer = 1; + break; case INTERP_ALGO_BILINEAR: scaler_ptx = vf_scale_cuda_ptx; function_infix = "_Bilinear"; s->interp_use_linear = 1; + s->interp_as_integer = 1; break; case INTERP_ALGO_DEFAULT: case INTERP_ALGO_BICUBIC: scaler_ptx = vf_scale_cuda_bicubic_ptx; function_infix = "_Bicubic"; s->interp_use_linear = 0; + s->interp_as_integer = 0; + break; + case INTERP_ALGO_LANCZOS: + scaler_ptx = vf_scale_cuda_bicubic_ptx; + function_infix = "_Lanczos"; + s->interp_use_linear = 0; + s->interp_as_integer = 0; break; default: av_log(ctx, AV_LOG_ERROR, "Unknown interpolation algorithm\n"); @@ -372,8 +376,8 @@ static av_cold int cudascale_config_props(AVFilterLink *outlink) if (ret < 0) return ret; - av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d\n", - inlink->w, inlink->h, outlink->w, outlink->h); + av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d%s\n", + inlink->w, inlink->h, outlink->w, outlink->h, s->passthrough ? " (passthrough)" : ""); if (inlink->sample_aspect_ratio.num) { outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w, @@ -398,14 +402,15 @@ static int call_resize_kernel(AVFilterContext *ctx, CUfunction func, int channel CudaFunctions *cu = s->hwctx->internal->cuda_dl; CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr; CUtexObject tex = 0; - void *args_uchar[] = { &tex, &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height, &src_pitch, &bit_depth }; + void *args_uchar[] = { &tex, &dst_devptr, &dst_width, &dst_height, &dst_pitch, + &src_width, &src_height, &bit_depth, &s->param }; int ret; CUDA_TEXTURE_DESC tex_desc = { .filterMode = s->interp_use_linear ? CU_TR_FILTER_MODE_LINEAR : CU_TR_FILTER_MODE_POINT, - .flags = CU_TRSF_READ_AS_INTEGER, + .flags = s->interp_as_integer ? CU_TRSF_READ_AS_INTEGER : 0, }; CUDA_RESOURCE_DESC res_desc = { @@ -416,10 +421,14 @@ static int call_resize_kernel(AVFilterContext *ctx, CUfunction func, int channel .res.pitch2D.numChannels = channels, .res.pitch2D.width = src_width, .res.pitch2D.height = src_height, - .res.pitch2D.pitchInBytes = src_pitch * pixel_size, + .res.pitch2D.pitchInBytes = src_pitch, .res.pitch2D.devPtr = (CUdeviceptr)src_dptr, }; + // Handling of channels is done via vector-types in cuda, so their size is implicitly part of the pitch + // Same for pixel_size, which is represented via datatypes on the cuda side of things. + dst_pitch /= channels * pixel_size; + ret = CHECK_CU(cu->cuTexObjectCreate(&tex, &res_desc, &tex_desc, NULL)); if (ret < 0) goto exit; @@ -448,12 +457,12 @@ static int scalecuda_resize(AVFilterContext *ctx, out->data[0], out->width, out->height, out->linesize[0], 1, 8); call_resize_kernel(ctx, s->cu_func_uchar, 1, - in->data[1], in->width / 2, in->height / 2, in->linesize[0] / 2, - out->data[1], out->width / 2, out->height / 2, out->linesize[0] / 2, + in->data[1], in->width / 2, in->height / 2, in->linesize[1], + out->data[1], out->width / 2, out->height / 2, out->linesize[1], 1, 8); call_resize_kernel(ctx, s->cu_func_uchar, 1, - in->data[2], in->width / 2, in->height / 2, in->linesize[0] / 2, - out->data[2], out->width / 2, out->height / 2, out->linesize[0] / 2, + in->data[2], in->width / 2, in->height / 2, in->linesize[2], + out->data[2], out->width / 2, out->height / 2, out->linesize[2], 1, 8); break; case AV_PIX_FMT_YUV444P: @@ -462,26 +471,26 @@ static int scalecuda_resize(AVFilterContext *ctx, out->data[0], out->width, out->height, out->linesize[0], 1, 8); call_resize_kernel(ctx, s->cu_func_uchar, 1, - in->data[1], in->width, in->height, in->linesize[0], - out->data[1], out->width, out->height, out->linesize[0], + in->data[1], in->width, in->height, in->linesize[1], + out->data[1], out->width, out->height, out->linesize[1], 1, 8); call_resize_kernel(ctx, s->cu_func_uchar, 1, - in->data[2], in->width, in->height, in->linesize[0], - out->data[2], out->width, out->height, out->linesize[0], + in->data[2], in->width, in->height, in->linesize[2], + out->data[2], out->width, out->height, out->linesize[2], 1, 8); break; case AV_PIX_FMT_YUV444P16: call_resize_kernel(ctx, s->cu_func_ushort, 1, - in->data[0], in->width, in->height, in->linesize[0] / 2, - out->data[0], out->width, out->height, out->linesize[0] / 2, + in->data[0], in->width, in->height, in->linesize[0], + out->data[0], out->width, out->height, out->linesize[0], 2, 16); call_resize_kernel(ctx, s->cu_func_ushort, 1, - in->data[1], in->width, in->height, in->linesize[1] / 2, - out->data[1], out->width, out->height, out->linesize[1] / 2, + in->data[1], in->width, in->height, in->linesize[1], + out->data[1], out->width, out->height, out->linesize[1], 2, 16); call_resize_kernel(ctx, s->cu_func_ushort, 1, - in->data[2], in->width, in->height, in->linesize[2] / 2, - out->data[2], out->width, out->height, out->linesize[2] / 2, + in->data[2], in->width, in->height, in->linesize[2], + out->data[2], out->width, out->height, out->linesize[2], 2, 16); break; case AV_PIX_FMT_NV12: @@ -491,29 +500,36 @@ static int scalecuda_resize(AVFilterContext *ctx, 1, 8); call_resize_kernel(ctx, s->cu_func_uchar2, 2, in->data[1], in->width / 2, in->height / 2, in->linesize[1], - out->data[1], out->width / 2, out->height / 2, out->linesize[1] / 2, + out->data[1], out->width / 2, out->height / 2, out->linesize[1], 1, 8); break; case AV_PIX_FMT_P010LE: call_resize_kernel(ctx, s->cu_func_ushort, 1, - in->data[0], in->width, in->height, in->linesize[0] / 2, - out->data[0], out->width, out->height, out->linesize[0] / 2, + in->data[0], in->width, in->height, in->linesize[0], + out->data[0], out->width, out->height, out->linesize[0], 2, 10); call_resize_kernel(ctx, s->cu_func_ushort2, 2, - in->data[1], in->width / 2, in->height / 2, in->linesize[1] / 2, - out->data[1], out->width / 2, out->height / 2, out->linesize[1] / 4, + in->data[1], in->width / 2, in->height / 2, in->linesize[1], + out->data[1], out->width / 2, out->height / 2, out->linesize[1], 2, 10); break; case AV_PIX_FMT_P016LE: call_resize_kernel(ctx, s->cu_func_ushort, 1, - in->data[0], in->width, in->height, in->linesize[0] / 2, - out->data[0], out->width, out->height, out->linesize[0] / 2, + in->data[0], in->width, in->height, in->linesize[0], + out->data[0], out->width, out->height, out->linesize[0], 2, 16); call_resize_kernel(ctx, s->cu_func_ushort2, 2, - in->data[1], in->width / 2, in->height / 2, in->linesize[1] / 2, - out->data[1], out->width / 2, out->height / 2, out->linesize[1] / 4, + in->data[1], in->width / 2, in->height / 2, in->linesize[1], + out->data[1], out->width / 2, out->height / 2, out->linesize[1], 2, 16); break; + case AV_PIX_FMT_0RGB32: + case AV_PIX_FMT_0BGR32: + call_resize_kernel(ctx, s->cu_func_uchar4, 4, + in->data[0], in->width, in->height, in->linesize[0], + out->data[0], out->width, out->height, out->linesize[0], + 1, 8); + break; default: return AVERROR_BUG; } @@ -524,6 +540,7 @@ static int scalecuda_resize(AVFilterContext *ctx, static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in) { CUDAScaleContext *s = ctx->priv; + AVFilterLink *outlink = ctx->outputs[0]; AVFrame *src = in; int ret; @@ -539,8 +556,8 @@ static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in) av_frame_move_ref(out, s->frame); av_frame_move_ref(s->frame, s->tmp_frame); - s->frame->width = s->planes_out[0].width; - s->frame->height = s->planes_out[0].height; + s->frame->width = outlink->w; + s->frame->height = outlink->h; ret = av_frame_copy_props(out, in); if (ret < 0) @@ -560,6 +577,9 @@ static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in) CUcontext dummy; int ret = 0; + if (s->passthrough) + return ff_filter_frame(outlink, in); + out = av_frame_alloc(); if (!out) { ret = AVERROR(ENOMEM); @@ -589,19 +609,32 @@ fail: return ret; } +static AVFrame *cudascale_get_video_buffer(AVFilterLink *inlink, int w, int h) +{ + CUDAScaleContext *s = inlink->dst->priv; + + return s->passthrough ? + ff_null_get_video_buffer (inlink, w, h) : + ff_default_get_video_buffer(inlink, w, h); +} + #define OFFSET(x) offsetof(CUDAScaleContext, x) #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM) static const AVOption 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 }, + { "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 }, { "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" }, - { "bilinear", "bilinear", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BILINEAR }, 0, 0, FLAGS, "interp_algo" }, - { "bicubic", "bicubic", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BICUBIC }, 0, 0, FLAGS, "interp_algo" }, - { "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" }, - { "disable", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 0 }, 0, 0, FLAGS, "force_oar" }, - { "decrease", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 1 }, 0, 0, FLAGS, "force_oar" }, - { "increase", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 2 }, 0, 0, FLAGS, "force_oar" }, - { "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 }, + { "nearest", "nearest neighbour", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_NEAREST }, 0, 0, FLAGS, "interp_algo" }, + { "bilinear", "bilinear", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BILINEAR }, 0, 0, FLAGS, "interp_algo" }, + { "bicubic", "bicubic", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BICUBIC }, 0, 0, FLAGS, "interp_algo" }, + { "lanczos", "lanczos", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_LANCZOS }, 0, 0, FLAGS, "interp_algo" }, + { "passthrough", "Do not process frames at all if parameters match", OFFSET(passthrough), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, FLAGS }, + { "param", "Algorithm-Specific parameter", OFFSET(param), AV_OPT_TYPE_FLOAT, { .dbl = SCALE_CUDA_PARAM_DEFAULT }, -FLT_MAX, FLT_MAX, FLAGS }, + { "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" }, + { "disable", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 0 }, 0, 0, FLAGS, "force_oar" }, + { "decrease", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 1 }, 0, 0, FLAGS, "force_oar" }, + { "increase", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 2 }, 0, 0, FLAGS, "force_oar" }, + { "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 }, { NULL }, }; @@ -617,6 +650,7 @@ static const AVFilterPad cudascale_inputs[] = { .name = "default", .type = AVMEDIA_TYPE_VIDEO, .filter_frame = cudascale_filter_frame, + .get_video_buffer = cudascale_get_video_buffer, }, { NULL } };