1. ffmpeg
  2. Commit 921bd9a2be4c1f4a0524346a694e2da814488d7f
Browse code

avfilter/scale_cuda: add CUDA scale filter

Signed-off-by: Timo Rothenpieler <timo@rothenpieler.org>

Yogender Gupta authored on 2017/05/11 02:18:16
Showing 6 changed files
Changelog
History View file @ 921bd9a
... ... 
@@ -10,6 +10,7 @@ version <next>:
10 10 
 - config.log and other configuration files moved into ffbuild/ directory
11 11 
 - update cuvid/nvenc headers to Video Codec SDK 8.0.14
12 12 
 - afir audio filter
13 
+- scale_cuda CUDA based video scale filter
13 14
14 15 
 version 3.3:
15 16 
 - CrystalHD decoder moved to new decode API
libavfilter/Makefile
History View file @ 921bd9a
... ... 
@@ -267,6 +267,7 @@ OBJS-$(CONFIG_REVERSE_FILTER)                += f_reverse.o
267 267 
 OBJS-$(CONFIG_ROTATE_FILTER)                 += vf_rotate.o
268 268 
 OBJS-$(CONFIG_SAB_FILTER)                    += vf_sab.o
269 269 
 OBJS-$(CONFIG_SCALE_FILTER)                  += vf_scale.o scale.o
270 
+OBJS-$(CONFIG_SCALE_CUDA_FILTER)             += vf_scale_cuda.o vf_scale_cuda.ptx.o
270 271 
 OBJS-$(CONFIG_SCALE_NPP_FILTER)              += vf_scale_npp.o scale.o
271 272 
 OBJS-$(CONFIG_SCALE_QSV_FILTER)              += vf_scale_qsv.o
272 273 
 OBJS-$(CONFIG_SCALE_VAAPI_FILTER)            += vf_scale_vaapi.o scale.o
libavfilter/allfilters.c
History View file @ 921bd9a
... ... 
@@ -278,6 +278,7 @@ static void register_all(void)
278 278 
     REGISTER_FILTER(ROTATE,         rotate,         vf);
279 279 
     REGISTER_FILTER(SAB,            sab,            vf);
280 280 
     REGISTER_FILTER(SCALE,          scale,          vf);
281 
+    REGISTER_FILTER(SCALE_CUDA,     scale_cuda,     vf);
281 282 
     REGISTER_FILTER(SCALE_NPP,      scale_npp,      vf);
282 283 
     REGISTER_FILTER(SCALE_QSV,      scale_qsv,      vf);
283 284 
     REGISTER_FILTER(SCALE_VAAPI,    scale_vaapi,    vf);
libavfilter/version.h
History View file @ 921bd9a
... ... 
@@ -31,7 +31,7 @@
31 31
32 32 
 #define LIBAVFILTER_VERSION_MAJOR   6
33 33 
 #define LIBAVFILTER_VERSION_MINOR  89
34 
-#define LIBAVFILTER_VERSION_MICRO 100
34 
+#define LIBAVFILTER_VERSION_MICRO 101
35 35
36 36 
 #define LIBAVFILTER_VERSION_INT AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \
37 37 
                                                LIBAVFILTER_VERSION_MINOR, \
libavfilter/vf_scale_cuda.c
History View file @ 921bd9a
38 38 
new file mode 100644
... ... 
@@ -0,0 +1,556 @@
0 
+/*
1 
+* Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
2 
+*
3 
+* Permission is hereby granted, free of charge, to any person obtaining a
4 
+* copy of this software and associated documentation files (the "Software"),
5 
+* to deal in the Software without restriction, including without limitation
6 
+* the rights to use, copy, modify, merge, publish, distribute, sublicense,
7 
+* and/or sell copies of the Software, and to permit persons to whom the
8 
+* Software is furnished to do so, subject to the following conditions:
9 
+*
10 
+* The above copyright notice and this permission notice shall be included in
11 
+* all copies or substantial portions of the Software.
12 
+*
13 
+* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 
+* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 
+* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
16 
+* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 
+* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
18 
+* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
19 
+* DEALINGS IN THE SOFTWARE.
20 
+*/
21 
+
22 
+#include <cuda.h>
23 
+#include <stdio.h>
24 
+#include <string.h>
25 
+
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/internal.h"
31 
+#include "libavutil/opt.h"
32 
+#include "libavutil/pixdesc.h"
33 
+
34 
+#include "avfilter.h"
35 
+#include "formats.h"
36 
+#include "internal.h"
37 
+#include "scale.h"
38 
+#include "video.h"
39 
+
40 
+static const enum AVPixelFormat supported_formats[] = {
41 
+    AV_PIX_FMT_YUV420P,
42 
+    AV_PIX_FMT_NV12,
43 
+    AV_PIX_FMT_YUV444P,
44 
+    AV_PIX_FMT_P010,
45 
+    AV_PIX_FMT_P016
46 
+};
47 
+
48 
+#define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
49 
+#define ALIGN_UP(a, b) (((a) + (b) - 1) & ~((b) - 1))
50 
+#define NUM_BUFFERS 2
51 
+#define BLOCKX 32
52 
+#define BLOCKY 16
53 
+
54 
+typedef struct CUDAScaleContext {
55 
+    const AVClass *class;
56 
+    enum AVPixelFormat in_fmt;
57 
+    enum AVPixelFormat out_fmt;
58 
+
59 
+    struct {
60 
+        int width;
61 
+        int height;
62 
+    } planes_in[3], planes_out[3];
63 
+
64 
+    AVBufferRef *frames_ctx;
65 
+    AVFrame     *frame;
66 
+
67 
+    AVFrame *tmp_frame;
68 
+    int passthrough;
69 
+
70 
+    /**
71 
+     * Output sw format. AV_PIX_FMT_NONE for no conversion.
72 
+     */
73 
+    enum AVPixelFormat format;
74 
+
75 
+    char *w_expr;               ///< width  expression string
76 
+    char *h_expr;               ///< height expression string
77 
+
78 
+    CUcontext   cu_ctx;
79 
+    CUevent     cu_event;
80 
+    CUmodule    cu_module;
81 
+    CUfunction  cu_func_uchar;
82 
+    CUfunction  cu_func_uchar2;
83 
+    CUfunction  cu_func_uchar4;
84 
+    CUfunction  cu_func_ushort;
85 
+    CUfunction  cu_func_ushort2;
86 
+    CUfunction  cu_func_ushort4;
87 
+    CUtexref    cu_tex_uchar;
88 
+    CUtexref    cu_tex_uchar2;
89 
+    CUtexref    cu_tex_uchar4;
90 
+    CUtexref    cu_tex_ushort;
91 
+    CUtexref    cu_tex_ushort2;
92 
+    CUtexref    cu_tex_ushort4;
93 
+
94 
+    CUdeviceptr srcBuffer;
95 
+    CUdeviceptr dstBuffer;
96 
+    int         tex_alignment;
97 
+} CUDAScaleContext;
98 
+
99 
+static av_cold int cudascale_init(AVFilterContext *ctx)
100 
+{
101 
+    CUDAScaleContext *s = ctx->priv;
102 
+
103 
+    s->format = AV_PIX_FMT_NONE;
104 
+    s->frame = av_frame_alloc();
105 
+    if (!s->frame)
106 
+        return AVERROR(ENOMEM);
107 
+
108 
+    s->tmp_frame = av_frame_alloc();
109 
+    if (!s->tmp_frame)
110 
+        return AVERROR(ENOMEM);
111 
+
112 
+    return 0;
113 
+}
114 
+
115 
+static av_cold void cudascale_uninit(AVFilterContext *ctx)
116 
+{
117 
+    CUDAScaleContext *s = ctx->priv;
118 
+
119 
+    av_frame_free(&s->frame);
120 
+    av_buffer_unref(&s->frames_ctx);
121 
+    av_frame_free(&s->tmp_frame);
122 
+}
123 
+
124 
+static int cudascale_query_formats(AVFilterContext *ctx)
125 
+{
126 
+    static const enum AVPixelFormat pixel_formats[] = {
127 
+        AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE,
128 
+    };
129 
+    AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats);
130 
+
131 
+    return ff_set_common_formats(ctx, pix_fmts);
132 
+}
133 
+
134 
+static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
135 
+{
136 
+    AVBufferRef *out_ref = NULL;
137 
+    AVHWFramesContext *out_ctx;
138 
+    int in_sw, in_sh, out_sw, out_sh;
139 
+    int ret, i;
140 
+
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;
146 
+    }
147 
+
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;
153 
+    }
154 
+
155 
+    out_ref = av_hwframe_ctx_alloc(device_ctx);
156 
+    if (!out_ref)
157 
+        return AVERROR(ENOMEM);
158 
+    out_ctx = (AVHWFramesContext*)out_ref->data;
159 
+
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);
164 
+
165 
+    ret = av_hwframe_ctx_init(out_ref);
166 
+    if (ret < 0)
167 
+        goto fail;
168 
+
169 
+    av_frame_unref(s->frame);
170 
+    ret = av_hwframe_get_buffer(out_ref, s->frame, 0);
171 
+    if (ret < 0)
172 
+        goto fail;
173 
+
174 
+    s->frame->width  = s->planes_out[0].width;
175 
+    s->frame->height = s->planes_out[0].height;
176 
+
177 
+    av_buffer_unref(&s->frames_ctx);
178 
+    s->frames_ctx = out_ref;
179 
+
180 
+    return 0;
181 
+fail:
182 
+    av_buffer_unref(&out_ref);
183 
+    return ret;
184 
+}
185 
+
186 
+static int format_is_supported(enum AVPixelFormat fmt)
187 
+{
188 
+    int i;
189 
+
190 
+    for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
191 
+        if (supported_formats[i] == fmt)
192 
+            return 1;
193 
+    return 0;
194 
+}
195 
+
196 
+static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int in_height,
197 
+                                         int out_width, int out_height)
198 
+{
199 
+    CUDAScaleContext *s = ctx->priv;
200 
+
201 
+    AVHWFramesContext *in_frames_ctx;
202 
+
203 
+    enum AVPixelFormat in_format;
204 
+    enum AVPixelFormat out_format;
205 
+    int ret;
206 
+
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);
211 
+    }
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;
215 
+
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);
220 
+    }
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);
225 
+    }
226 
+
227 
+    if (in_width == out_width && in_height == out_height)
228 
+        s->passthrough = 1;
229 
+
230 
+    s->in_fmt = in_format;
231 
+    s->out_fmt = out_format;
232 
+
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;
237 
+
238 
+    ret = init_stage(s, in_frames_ctx->device_ref);
239 
+    if (ret < 0)
240 
+        return ret;
241 
+
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);
245 
+
246 
+    return 0;
247 
+}
248 
+
249 
+static av_cold int cudascale_config_props(AVFilterLink *outlink)
250 
+{
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 
+    CUresult err;
258 
+    int w, h;
259 
+    int ret;
260 
+
261 
+    extern char vf_scale_cuda_ptx[];
262 
+
263 
+    err = cuCtxPushCurrent(cuda_ctx);
264 
+    if (err != CUDA_SUCCESS) {
265 
+        av_log(ctx, AV_LOG_ERROR, "Error pushing cuda context\n");
266 
+        ret = AVERROR_UNKNOWN;
267 
+        goto fail;
268 
+    }
269 
+
270 
+    err = cuModuleLoadData(&s->cu_module, vf_scale_cuda_ptx);
271 
+    if (err != CUDA_SUCCESS) {
272 
+        av_log(ctx, AV_LOG_ERROR, "Error loading module data\n");
273 
+        ret = AVERROR_UNKNOWN;
274 
+        goto fail;
275 
+    }
276 
+
277 
+    cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, "Subsample_Bilinear_uchar");
278 
+    cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, "Subsample_Bilinear_uchar2");
279 
+    cuModuleGetFunction(&s->cu_func_uchar4, s->cu_module, "Subsample_Bilinear_uchar4");
280 
+    cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, "Subsample_Bilinear_ushort");
281 
+    cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, "Subsample_Bilinear_ushort2");
282 
+    cuModuleGetFunction(&s->cu_func_ushort4, s->cu_module, "Subsample_Bilinear_ushort4");
283 
+
284 
+    cuModuleGetTexRef(&s->cu_tex_uchar, s->cu_module, "uchar_tex");
285 
+    cuModuleGetTexRef(&s->cu_tex_uchar2, s->cu_module, "uchar2_tex");
286 
+    cuModuleGetTexRef(&s->cu_tex_uchar4, s->cu_module, "uchar4_tex");
287 
+    cuModuleGetTexRef(&s->cu_tex_ushort, s->cu_module, "ushort_tex");
288 
+    cuModuleGetTexRef(&s->cu_tex_ushort2, s->cu_module, "ushort2_tex");
289 
+    cuModuleGetTexRef(&s->cu_tex_ushort4, s->cu_module, "ushort4_tex");
290 
+
291 
+    cuTexRefSetFlags(s->cu_tex_uchar, CU_TRSF_READ_AS_INTEGER);
292 
+    cuTexRefSetFlags(s->cu_tex_uchar2, CU_TRSF_READ_AS_INTEGER);
293 
+    cuTexRefSetFlags(s->cu_tex_uchar4, CU_TRSF_READ_AS_INTEGER);
294 
+    cuTexRefSetFlags(s->cu_tex_ushort, CU_TRSF_READ_AS_INTEGER);
295 
+    cuTexRefSetFlags(s->cu_tex_ushort2, CU_TRSF_READ_AS_INTEGER);
296 
+    cuTexRefSetFlags(s->cu_tex_ushort4, CU_TRSF_READ_AS_INTEGER);
297 
+
298 
+    cuTexRefSetFilterMode(s->cu_tex_uchar, CU_TR_FILTER_MODE_LINEAR);
299 
+    cuTexRefSetFilterMode(s->cu_tex_uchar2, CU_TR_FILTER_MODE_LINEAR);
300 
+    cuTexRefSetFilterMode(s->cu_tex_uchar4, CU_TR_FILTER_MODE_LINEAR);
301 
+    cuTexRefSetFilterMode(s->cu_tex_ushort, CU_TR_FILTER_MODE_LINEAR);
302 
+    cuTexRefSetFilterMode(s->cu_tex_ushort2, CU_TR_FILTER_MODE_LINEAR);
303 
+    cuTexRefSetFilterMode(s->cu_tex_ushort4, CU_TR_FILTER_MODE_LINEAR);
304 
+
305 
+    cuCtxPopCurrent(&dummy);
306 
+
307 
+    if ((ret = ff_scale_eval_dimensions(s,
308 
+                                        s->w_expr, s->h_expr,
309 
+                                        inlink, outlink,
310 
+                                        &w, &h)) < 0)
311 
+        goto fail;
312 
+
313 
+    if (((int64_t)h * inlink->w) > INT_MAX  ||
314 
+        ((int64_t)w * inlink->h) > INT_MAX)
315 
+        av_log(ctx, AV_LOG_ERROR, "Rescaled value for width or height is too big.\n");
316 
+
317 
+    outlink->w = w;
318 
+    outlink->h = h;
319 
+
320 
+    ret = init_processing_chain(ctx, inlink->w, inlink->h, w, h);
321 
+    if (ret < 0)
322 
+        return ret;
323 
+
324 
+    av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d\n",
325 
+           inlink->w, inlink->h, outlink->w, outlink->h);
326 
+
327 
+    if (inlink->sample_aspect_ratio.num) {
328 
+        outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w,
329 
+                                                             outlink->w*inlink->h},
330 
+                                                inlink->sample_aspect_ratio);
331 
+    } else {
332 
+        outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
333 
+    }
334 
+
335 
+    return 0;
336 
+
337 
+fail:
338 
+    return ret;
339 
+}
340 
+
341 
+static int call_resize_kernel(CUDAScaleContext *s, CUfunction func, CUtexref tex, int channels,
342 
+                              uint8_t *src_dptr, int src_width, int src_height, int src_pitch,
343 
+                              uint8_t *dst_dptr, int dst_width, int dst_height, int dst_pitch,
344 
+                              int pixel_size)
345 
+{
346 
+    CUdeviceptr src_devptr = (CUdeviceptr)src_dptr;
347 
+    CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr;
348 
+    void *args_uchar[] = { &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height };
349 
+    CUDA_ARRAY_DESCRIPTOR desc;
350 
+
351 
+    desc.Width  = src_width;
352 
+    desc.Height = src_height;
353 
+    desc.NumChannels = channels;
354 
+    if (pixel_size == 1) {
355 
+        desc.Format = CU_AD_FORMAT_UNSIGNED_INT8;
356 
+    } else {
357 
+        desc.Format = CU_AD_FORMAT_UNSIGNED_INT16;
358 
+    }
359 
+
360 
+    cuTexRefSetAddress2D_v3(tex, &desc, src_devptr, src_pitch * pixel_size);
361 
+    cuLaunchKernel(func, DIV_UP(dst_width, BLOCKX), DIV_UP(dst_height, BLOCKY), 1, BLOCKX, BLOCKY, 1, 0, 0, args_uchar, NULL);
362 
+
363 
+    return 0;
364 
+}
365 
+
366 
+static int scalecuda_resize(AVFilterContext *ctx,
367 
+                            AVFrame *out, AVFrame *in)
368 
+{
369 
+    AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
370 
+    CUDAScaleContext *s = ctx->priv;
371 
+
372 
+    switch (in_frames_ctx->sw_format) {
373 
+    case AV_PIX_FMT_YUV420P:
374 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
375 
+                           in->data[0], in->width, in->height, in->linesize[0],
376 
+                           out->data[0], out->width, out->height, out->linesize[0],
377 
+                           1);
378 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
379 
+                           in->data[0]+in->linesize[0]*in->height, in->width/2, in->height/2, in->linesize[0]/2,
380 
+                           out->data[0]+out->linesize[0]*out->height, out->width/2, out->height/2, out->linesize[0]/2,
381 
+                           1);
382 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
383 
+                           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,
384 
+                           out->data[0]+(out->linesize[0]*out->height*5)/4, out->width/2, out->height/2, out->linesize[0]/2,
385 
+                           1);
386 
+        break;
387 
+    case AV_PIX_FMT_YUV444P:
388 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
389 
+                           in->data[0], in->width, in->height, in->linesize[0],
390 
+                           out->data[0], out->width, out->height, out->linesize[0],
391 
+                           1);
392 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
393 
+                           in->data[0]+in->linesize[0]*in->height, in->width, in->height, in->linesize[0],
394 
+                           out->data[0]+out->linesize[0]*out->height, out->width, out->height, out->linesize[0],
395 
+                           1);
396 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
397 
+                           in->data[0]+in->linesize[0]*in->height*2, in->width, in->height, in->linesize[0],
398 
+                           out->data[0]+out->linesize[0]*out->height*2, out->width, out->height, out->linesize[0],
399 
+                           1);
400 
+        break;
401 
+    case AV_PIX_FMT_NV12:
402 
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
403 
+                           in->data[0], in->width, in->height, in->linesize[0],
404 
+                           out->data[0], out->width, out->height, out->linesize[0],
405 
+                           1);
406 
+        call_resize_kernel(s, s->cu_func_uchar2, s->cu_tex_uchar2, 2,
407 
+                           in->data[1], in->width/2, in->height/2, in->linesize[1],
408 
+                           out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width/2, out->height/2, out->linesize[1]/2,
409 
+                           1);
410 
+        break;
411 
+    case AV_PIX_FMT_P010LE:
412 
+        call_resize_kernel(s, s->cu_func_ushort, s->cu_tex_ushort, 1,
413 
+                           in->data[0], in->width, in->height, in->linesize[0]/2,
414 
+                           out->data[0], out->width, out->height, out->linesize[0]/2,
415 
+                           2);
416 
+        call_resize_kernel(s, s->cu_func_ushort2, s->cu_tex_ushort2, 2,
417 
+                           in->data[1], in->width / 2, in->height / 2, in->linesize[1]/2,
418 
+                           out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
419 
+                           2);
420 
+        break;
421 
+    case AV_PIX_FMT_P016LE:
422 
+        call_resize_kernel(s, s->cu_func_ushort, s->cu_tex_ushort, 1,
423 
+                           in->data[0], in->width, in->height, in->linesize[0] / 2,
424 
+                           out->data[0], out->width, out->height, out->linesize[0] / 2,
425 
+                           2);
426 
+        call_resize_kernel(s, s->cu_func_ushort2, s->cu_tex_ushort2, 2,
427 
+                           in->data[1], in->width / 2, in->height / 2, in->linesize[1] / 2,
428 
+                           out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
429 
+                           2);
430 
+        break;
431 
+    default:
432 
+        return AVERROR_BUG;
433 
+    }
434 
+
435 
+    return 0;
436 
+}
437 
+
438 
+static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
439 
+{
440 
+    CUDAScaleContext *s = ctx->priv;
441 
+    AVFrame *src = in;
442 
+    int ret;
443 
+
444 
+    ret = scalecuda_resize(ctx, s->frame, src);
445 
+    if (ret < 0)
446 
+        return ret;
447 
+
448 
+    src = s->frame;
449 
+    ret = av_hwframe_get_buffer(src->hw_frames_ctx, s->tmp_frame, 0);
450 
+    if (ret < 0)
451 
+        return ret;
452 
+
453 
+    av_frame_move_ref(out, s->frame);
454 
+    av_frame_move_ref(s->frame, s->tmp_frame);
455 
+
456 
+    ret = av_frame_copy_props(out, in);
457 
+    if (ret < 0)
458 
+        return ret;
459 
+
460 
+    return 0;
461 
+}
462 
+
463 
+static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in)
464 
+{
465 
+    AVFilterContext              *ctx = link->dst;
466 
+    CUDAScaleContext               *s = ctx->priv;
467 
+    AVFilterLink             *outlink = ctx->outputs[0];
468 
+    AVHWFramesContext     *frames_ctx = (AVHWFramesContext*)s->frames_ctx->data;
469 
+    AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
470 
+
471 
+    AVFrame *out = NULL;
472 
+    CUresult err;
473 
+    CUcontext dummy;
474 
+    int ret = 0;
475 
+
476 
+    out = av_frame_alloc();
477 
+    if (!out) {
478 
+        ret = AVERROR(ENOMEM);
479 
+        goto fail;
480 
+    }
481 
+
482 
+    err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
483 
+    if (err != CUDA_SUCCESS) {
484 
+        ret = AVERROR_UNKNOWN;
485 
+        goto fail;
486 
+    }
487 
+
488 
+    ret = cudascale_scale(ctx, out, in);
489 
+
490 
+    cuCtxPopCurrent(&dummy);
491 
+    if (ret < 0)
492 
+        goto fail;
493 
+
494 
+    av_reduce(&out->sample_aspect_ratio.num, &out->sample_aspect_ratio.den,
495 
+              (int64_t)in->sample_aspect_ratio.num * outlink->h * link->w,
496 
+              (int64_t)in->sample_aspect_ratio.den * outlink->w * link->h,
497 
+              INT_MAX);
498 
+
499 
+    av_frame_free(&in);
500 
+    return ff_filter_frame(outlink, out);
501 
+fail:
502 
+    av_frame_free(&in);
503 
+    av_frame_free(&out);
504 
+    return ret;
505 
+}
506 
+
507 
+#define OFFSET(x) offsetof(CUDAScaleContext, x)
508 
+#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM)
509 
+static const AVOption options[] = {
510 
+    { "w",      "Output video width",  OFFSET(w_expr),     AV_OPT_TYPE_STRING, { .str = "iw"   }, .flags = FLAGS },
511 
+    { "h",      "Output video height", OFFSET(h_expr),     AV_OPT_TYPE_STRING, { .str = "ih"   }, .flags = FLAGS },
512 
+    { NULL },
513 
+};
514 
+
515 
+static const AVClass cudascale_class = {
516 
+    .class_name = "cudascale",
517 
+    .item_name  = av_default_item_name,
518 
+    .option     = options,
519 
+    .version    = LIBAVUTIL_VERSION_INT,
520 
+};
521 
+
522 
+static const AVFilterPad cudascale_inputs[] = {
523 
+    {
524 
+        .name        = "default",
525 
+        .type        = AVMEDIA_TYPE_VIDEO,
526 
+        .filter_frame = cudascale_filter_frame,
527 
+    },
528 
+    { NULL }
529 
+};
530 
+
531 
+static const AVFilterPad cudascale_outputs[] = {
532 
+    {
533 
+        .name         = "default",
534 
+        .type         = AVMEDIA_TYPE_VIDEO,
535 
+        .config_props = cudascale_config_props,
536 
+    },
537 
+    { NULL }
538 
+};
539 
+
540 
+AVFilter ff_vf_scale_cuda = {
541 
+    .name      = "scale_cuda",
542 
+    .description = NULL_IF_CONFIG_SMALL("GPU accelerated video resizer"),
543 
+
544 
+    .init          = cudascale_init,
545 
+    .uninit        = cudascale_uninit,
546 
+    .query_formats = cudascale_query_formats,
547 
+
548 
+    .priv_size = sizeof(CUDAScaleContext),
549 
+    .priv_class = &cudascale_class,
550 
+
551 
+    .inputs    = cudascale_inputs,
552 
+    .outputs   = cudascale_outputs,
553 
+
554 
+    .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
555 
+};
libavfilter/vf_scale_cuda.cu
History View file @ 921bd9a
0 556 
new file mode 100644
... ... 
@@ -0,0 +1,212 @@
0 
+/*
1 
+ * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
2 
+ *
3 
+ * Permission is hereby granted, free of charge, to any person obtaining a
4 
+ * copy of this software and associated documentation files (the "Software"),
5 
+ * to deal in the Software without restriction, including without limitation
6 
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
7 
+ * and/or sell copies of the Software, and to permit persons to whom the
8 
+ * Software is furnished to do so, subject to the following conditions:
9 
+ *
10 
+ * The above copyright notice and this permission notice shall be included in
11 
+ * all copies or substantial portions of the Software.
12 
+ *
13 
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
16 
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
18 
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
19 
+ * DEALINGS IN THE SOFTWARE.
20 
+ */
21 
+
22 
+extern "C" {
23 
+
24 
+texture<unsigned char, 2> uchar_tex;
25 
+texture<uchar2, 2>  uchar2_tex;
26 
+texture<uchar4, 2>  uchar4_tex;
27 
+texture<unsigned short, 2> ushort_tex;
28 
+texture<ushort2, 2>  ushort2_tex;
29 
+texture<ushort4, 2>  ushort4_tex;
30 
+
31 
+__global__ void Subsample_Bilinear_uchar(unsigned char *dst,
32 
+                                    int dst_width, int dst_height, int dst_pitch,
33 
+                                    int src_width, int src_height)
34 
+{
35 
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
36 
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
37 
+
38 
+    if (yo < dst_height && xo < dst_width)
39 
+    {
40 
+        float hscale = (float)src_width / (float)dst_width;
41 
+        float vscale = (float)src_height / (float)dst_height;
42 
+        float xi = (xo + 0.5f) * hscale;
43 
+        float yi = (yo + 0.5f) * vscale;
44 
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
45 
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
46 
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
47 
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
48 
+        float dx = wh / (0.5f + wh);
49 
+        float dy = wv / (0.5f + wv);
50 
+        int y0 = tex2D(uchar_tex, xi-dx, yi-dy);
51 
+        int y1 = tex2D(uchar_tex, xi+dx, yi-dy);
52 
+        int y2 = tex2D(uchar_tex, xi-dx, yi+dy);
53 
+        int y3 = tex2D(uchar_tex, xi+dx, yi+dy);
54 
+        dst[yo*dst_pitch+xo] = (unsigned char)((y0+y1+y2+y3+2) >> 2);
55 
+    }
56 
+}
57 
+
58 
+__global__ void Subsample_Bilinear_uchar2(uchar2 *dst,
59 
+                                    int dst_width, int dst_height, int dst_pitch2,
60 
+                                    int src_width, int src_height)
61 
+{
62 
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
63 
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
64 
+
65 
+    if (yo < dst_height && xo < dst_width)
66 
+    {
67 
+        float hscale = (float)src_width / (float)dst_width;
68 
+        float vscale = (float)src_height / (float)dst_height;
69 
+        float xi = (xo + 0.5f) * hscale;
70 
+        float yi = (yo + 0.5f) * vscale;
71 
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
72 
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
73 
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
74 
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
75 
+        float dx = wh / (0.5f + wh);
76 
+        float dy = wv / (0.5f + wv);
77 
+        uchar2 c0 = tex2D(uchar2_tex, xi-dx, yi-dy);
78 
+        uchar2 c1 = tex2D(uchar2_tex, xi+dx, yi-dy);
79 
+        uchar2 c2 = tex2D(uchar2_tex, xi-dx, yi+dy);
80 
+        uchar2 c3 = tex2D(uchar2_tex, xi+dx, yi+dy);
81 
+        int2 uv;
82 
+        uv.x = ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
83 
+        uv.y = ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
84 
+        dst[yo*dst_pitch2+xo] = make_uchar2((unsigned char)uv.x, (unsigned char)uv.y);
85 
+    }
86 
+}
87 
+
88 
+__global__ void Subsample_Bilinear_uchar4(uchar4 *dst,
89 
+                                    int dst_width, int dst_height, int dst_pitch,
90 
+                                    int src_width, int src_height)
91 
+{
92 
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
93 
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
94 
+
95 
+    if (yo < dst_height && xo < dst_width)
96 
+    {
97 
+        float hscale = (float)src_width / (float)dst_width;
98 
+        float vscale = (float)src_height / (float)dst_height;
99 
+        float xi = (xo + 0.5f) * hscale;
100 
+        float yi = (yo + 0.5f) * vscale;
101 
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
102 
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
103 
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
104 
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
105 
+        float dx = wh / (0.5f + wh);
106 
+        float dy = wv / (0.5f + wv);
107 
+        uchar4 c0 = tex2D(uchar4_tex, xi-dx, yi-dy);
108 
+        uchar4 c1 = tex2D(uchar4_tex, xi+dx, yi-dy);
109 
+        uchar4 c2 = tex2D(uchar4_tex, xi-dx, yi+dy);
110 
+        uchar4 c3 = tex2D(uchar4_tex, xi+dx, yi+dy);
111 
+        int4 res;
112 
+        res.x =  ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
113 
+        res.y =  ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
114 
+        res.z =  ((int)c0.z+(int)c1.z+(int)c2.z+(int)c3.z+2) >> 2;
115 
+        res.w =  ((int)c0.w+(int)c1.w+(int)c2.w+(int)c3.w+2) >> 2;
116 
+        dst[yo*dst_pitch+xo] = make_uchar4(
117 
+            (unsigned char)res.x, (unsigned char)res.y, (unsigned char)res.z, (unsigned char)res.w);
118 
+    }
119 
+}
120 
+
121 
+__global__ void Subsample_Bilinear_ushort(unsigned short *dst,
122 
+                                    int dst_width, int dst_height, int dst_pitch,
123 
+                                    int src_width, int src_height)
124 
+{
125 
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
126 
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
127 
+
128 
+    if (yo < dst_height && xo < dst_width)
129 
+    {
130 
+        float hscale = (float)src_width / (float)dst_width;
131 
+        float vscale = (float)src_height / (float)dst_height;
132 
+        float xi = (xo + 0.5f) * hscale;
133 
+        float yi = (yo + 0.5f) * vscale;
134 
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
135 
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
136 
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
137 
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
138 
+        float dx = wh / (0.5f + wh);
139 
+        float dy = wv / (0.5f + wv);
140 
+        int y0 = tex2D(ushort_tex, xi-dx, yi-dy);
141 
+        int y1 = tex2D(ushort_tex, xi+dx, yi-dy);
142 
+        int y2 = tex2D(ushort_tex, xi-dx, yi+dy);
143 
+        int y3 = tex2D(ushort_tex, xi+dx, yi+dy);
144 
+        dst[yo*dst_pitch+xo] = (unsigned short)((y0+y1+y2+y3+2) >> 2);
145 
+    }
146 
+}
147 
+
148 
+__global__ void Subsample_Bilinear_ushort2(ushort2 *dst,
149 
+                                    int dst_width, int dst_height, int dst_pitch2,
150 
+                                    int src_width, int src_height)
151 
+{
152 
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
153 
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
154 
+
155 
+    if (yo < dst_height && xo < dst_width)
156 
+    {
157 
+        float hscale = (float)src_width / (float)dst_width;
158 
+        float vscale = (float)src_height / (float)dst_height;
159 
+        float xi = (xo + 0.5f) * hscale;
160 
+        float yi = (yo + 0.5f) * vscale;
161 
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
162 
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
163 
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
164 
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
165 
+        float dx = wh / (0.5f + wh);
166 
+        float dy = wv / (0.5f + wv);
167 
+        ushort2 c0 = tex2D(ushort2_tex, xi-dx, yi-dy);
168 
+        ushort2 c1 = tex2D(ushort2_tex, xi+dx, yi-dy);
169 
+        ushort2 c2 = tex2D(ushort2_tex, xi-dx, yi+dy);
170 
+        ushort2 c3 = tex2D(ushort2_tex, xi+dx, yi+dy);
171 
+        int2 uv;
172 
+        uv.x = ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
173 
+        uv.y = ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
174 
+        dst[yo*dst_pitch2+xo] = make_ushort2((unsigned short)uv.x, (unsigned short)uv.y);
175 
+    }
176 
+}
177 
+
178 
+__global__ void Subsample_Bilinear_ushort4(ushort4 *dst,
179 
+                                    int dst_width, int dst_height, int dst_pitch,
180 
+                                    int src_width, int src_height)
181 
+{
182 
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
183 
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
184 
+
185 
+    if (yo < dst_height && xo < dst_width)
186 
+    {
187 
+        float hscale = (float)src_width / (float)dst_width;
188 
+        float vscale = (float)src_height / (float)dst_height;
189 
+        float xi = (xo + 0.5f) * hscale;
190 
+        float yi = (yo + 0.5f) * vscale;
191 
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
192 
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
193 
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
194 
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
195 
+        float dx = wh / (0.5f + wh);
196 
+        float dy = wv / (0.5f + wv);
197 
+        ushort4 c0 = tex2D(ushort4_tex, xi-dx, yi-dy);
198 
+        ushort4 c1 = tex2D(ushort4_tex, xi+dx, yi-dy);
199 
+        ushort4 c2 = tex2D(ushort4_tex, xi-dx, yi+dy);
200 
+        ushort4 c3 = tex2D(ushort4_tex, xi+dx, yi+dy);
201 
+        int4 res;
202 
+        res.x =  ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
203 
+        res.y =  ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
204 
+        res.z =  ((int)c0.z+(int)c1.z+(int)c2.z+(int)c3.z+2) >> 2;
205 
+        res.w =  ((int)c0.w+(int)c1.w+(int)c2.w+(int)c3.w+2) >> 2;
206 
+        dst[yo*dst_pitch+xo] = make_ushort4(
207 
+            (unsigned short)res.x, (unsigned short)res.y, (unsigned short)res.z, (unsigned short)res.w);
208 
+    }
209 
+}
210 
+
211 
+}