History

hwcontext: add a CUDA implementation

Anton Khirnov committed 10 years ago

Commit ad884d11 parent 7bc780c7 files · +325 −0Browse files

Unified · Split

7 files
  1. doc/APIchanges +2 −0
  2. libavutil/Makefile +2 −0
  3. libavutil/hwcontext.c +3 −0
  4. libavutil/hwcontext.h +1 −0
  5. libavutil/hwcontext_cuda.c +270 −0
  6. libavutil/hwcontext_cuda.h +46 −0
  7. libavutil/hwcontext_internal.h +1 −0
doc/APIchanges+2−0
@@ -17,6 +17,8 @@ API changes, most recent first: · 16 unmodified lines
1717 xxxxxxx buffer.h - Add av_buffer_pool_init2().
1818 xxxxxxx hwcontext.h - Add a new installed header hwcontext.h with a new API
1919 for handling hwaccel frames.
20+ xxxxxxx hwcontext_cuda.h - Add a new installed header hwcontext_cuda.h with
21+ CUDA-specific hwcontext definitions.
2022 xxxxxxx hwcontext_vdpau.h - Add a new installed header hwcontext_vdpau.h with
2123 VDPAU-specific hwcontext definitions.
2224 xxxxxxx pixfmt.h - Add AV_PIX_FMT_CUDA.
libavutil/Makefile+2−0
@@ -24,6 +24,7 @@ HEADERS = adler32.h \ · 23 unmodified lines
2424 frame.h \
2525 hmac.h \
2626 hwcontext.h \
27+ hwcontext_cuda.h \
2728 hwcontext_vdpau.h \
2829 imgutils.h \
2930 intfloat.h \
@@ -106,6 +107,7 @@ OBJS = adler32.o \ · 76 unmodified lines
106107 xtea.o \
107108
108109 OBJS-$(CONFIG_LZO) += lzo.o
110+OBJS-$(CONFIG_CUDA) += hwcontext_cuda.o
109111 OBJS-$(CONFIG_VDPAU) += hwcontext_vdpau.o
110112
111113 OBJS += $(COMPAT_OBJS:%=../compat/%)
libavutil/hwcontext.c+3−0
@@ -29,6 +29,9 @@ · 28 unmodified lines
2929 #include "pixfmt.h"
3030
3131 static const HWContextType *hw_table[] = {
32+#if CONFIG_CUDA
33+ &ff_hwcontext_type_cuda,
34+#endif
3235 #if CONFIG_VDPAU
3336 &ff_hwcontext_type_vdpau,
3437 #endif
libavutil/hwcontext.h+1−0
@@ -26,6 +26,7 @@ · 25 unmodified lines
2626
2727 enum AVHWDeviceType {
2828 AV_HWDEVICE_TYPE_VDPAU,
29+ AV_HWDEVICE_TYPE_CUDA,
2930 };
3031
3132 typedef struct AVHWDeviceInternal AVHWDeviceInternal;
libavutil/hwcontext_cuda.c+270−0
@@ -0,0 +1,270 @@
1+/*
2+ * This file is part of Libav.
3+ *
4+ * Libav is free software; you can redistribute it and/or
5+ * modify it under the terms of the GNU Lesser General Public
6+ * License as published by the Free Software Foundation; either
7+ * version 2.1 of the License, or (at your option) any later version.
8+ *
9+ * Libav is distributed in the hope that it will be useful,
10+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
11+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12+ * Lesser General Public License for more details.
13+ *
14+ * You should have received a copy of the GNU Lesser General Public
15+ * License along with Libav; if not, write to the Free Software
16+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17+ */
18+
19+#include "buffer.h"
20+#include "common.h"
21+#include "hwcontext.h"
22+#include "hwcontext_internal.h"
23+#include "hwcontext_cuda.h"
24+#include "mem.h"
25+#include "pixdesc.h"
26+#include "pixfmt.h"
27+
28+typedef struct CUDAFramesContext {
29+ int shift_width, shift_height;
30+} CUDAFramesContext;
31+
32+static const enum AVPixelFormat supported_formats[] = {
33+ AV_PIX_FMT_NV12,
34+ AV_PIX_FMT_YUV420P,
35+ AV_PIX_FMT_YUV444P,
36+};
37+
38+static void cuda_buffer_free(void *opaque, uint8_t *data)
39+{
40+ AVHWFramesContext *ctx = opaque;
41+ AVCUDADeviceContext *hwctx = ctx->device_ctx->hwctx;
42+
43+ CUcontext dummy;
44+
45+ cuCtxPushCurrent(hwctx->cuda_ctx);
46+
47+ cuMemFree((CUdeviceptr)data);
48+
49+ cuCtxPopCurrent(&dummy);
50+}
51+
52+static AVBufferRef *cuda_pool_alloc(void *opaque, int size)
53+{
54+ AVHWFramesContext *ctx = opaque;
55+ AVCUDADeviceContext *hwctx = ctx->device_ctx->hwctx;
56+
57+ AVBufferRef *ret = NULL;
58+ CUcontext dummy = NULL;
59+ CUdeviceptr data;
60+ CUresult err;
61+
62+ err = cuCtxPushCurrent(hwctx->cuda_ctx);
63+ if (err != CUDA_SUCCESS) {
64+ av_log(ctx, AV_LOG_ERROR, "Error setting current CUDA context\n");
65+ return NULL;
66+ }
67+
68+ err = cuMemAlloc(&data, size);
69+ if (err != CUDA_SUCCESS)
70+ goto fail;
71+
72+ ret = av_buffer_create((uint8_t*)data, size, cuda_buffer_free, ctx, 0);
73+ if (!ret) {
74+ cuMemFree(data);
75+ goto fail;
76+ }
77+
78+fail:
79+ cuCtxPopCurrent(&dummy);
80+ return ret;
81+}
82+
83+static int cuda_frames_init(AVHWFramesContext *ctx)
84+{
85+ CUDAFramesContext *priv = ctx->internal->priv;
86+ int i;
87+
88+ for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++) {
89+ if (ctx->sw_format == supported_formats[i])
90+ break;
91+ }
92+ if (i == FF_ARRAY_ELEMS(supported_formats)) {
93+ av_log(ctx, AV_LOG_ERROR, "Pixel format '%s' is not supported\n",
94+ av_get_pix_fmt_name(ctx->sw_format));
95+ return AVERROR(ENOSYS);
96+ }
97+
98+ av_pix_fmt_get_chroma_sub_sample(ctx->sw_format, &priv->shift_width, &priv->shift_height);
99+
100+ if (!ctx->pool) {
101+ int size;
102+
103+ switch (ctx->sw_format) {
104+ case AV_PIX_FMT_NV12:
105+ case AV_PIX_FMT_YUV420P:
106+ size = ctx->width * ctx->height * 3 / 2;
107+ break;
108+ case AV_PIX_FMT_YUV444P:
109+ size = ctx->width * ctx->height * 3;
110+ break;
111+ }
112+
113+ ctx->internal->pool_internal = av_buffer_pool_init2(size, ctx, cuda_pool_alloc, NULL);
114+ if (!ctx->internal->pool_internal)
115+ return AVERROR(ENOMEM);
116+ }
117+
118+ return 0;
119+}
120+
121+static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
122+{
123+ frame->buf[0] = av_buffer_pool_get(ctx->pool);
124+ if (!frame->buf[0])
125+ return AVERROR(ENOMEM);
126+
127+ switch (ctx->sw_format) {
128+ case AV_PIX_FMT_NV12:
129+ frame->data[0] = frame->buf[0]->data;
130+ frame->data[1] = frame->data[0] + ctx->width * ctx->height;
131+ frame->linesize[0] = ctx->width;
132+ frame->linesize[1] = ctx->width;
133+ break;
134+ case AV_PIX_FMT_YUV420P:
135+ frame->data[0] = frame->buf[0]->data;
136+ frame->data[2] = frame->data[0] + ctx->width * ctx->height;
137+ frame->data[1] = frame->data[2] + ctx->width * ctx->height / 4;
138+ frame->linesize[0] = ctx->width;
139+ frame->linesize[1] = ctx->width / 2;
140+ frame->linesize[2] = ctx->width / 2;
141+ break;
142+ case AV_PIX_FMT_YUV444P:
143+ frame->data[0] = frame->buf[0]->data;
144+ frame->data[1] = frame->data[0] + ctx->width * ctx->height;
145+ frame->data[2] = frame->data[1] + ctx->width * ctx->height;
146+ frame->linesize[0] = ctx->width;
147+ frame->linesize[1] = ctx->width;
148+ frame->linesize[2] = ctx->width;
149+ break;
150+ default:
151+ av_frame_unref(frame);
152+ return AVERROR_BUG;
153+ }
154+
155+ frame->format = AV_PIX_FMT_CUDA;
156+ frame->width = ctx->width;
157+ frame->height = ctx->height;
158+
159+ return 0;
160+}
161+
162+static int cuda_transfer_get_formats(AVHWFramesContext *ctx,
163+ enum AVHWFrameTransferDirection dir,
164+ enum AVPixelFormat **formats)
165+{
166+ enum AVPixelFormat *fmts;
167+
168+ fmts = av_malloc_array(2, sizeof(*fmts));
169+ if (!fmts)
170+ return AVERROR(ENOMEM);
171+
172+ fmts[0] = ctx->sw_format;
173+ fmts[1] = AV_PIX_FMT_NONE;
174+
175+ *formats = fmts;
176+
177+ return 0;
178+}
179+
180+static int cuda_transfer_data_from(AVHWFramesContext *ctx, AVFrame *dst,
181+ const AVFrame *src)
182+{
183+ CUDAFramesContext *priv = ctx->internal->priv;
184+ AVCUDADeviceContext *device_hwctx = ctx->device_ctx->hwctx;
185+
186+ CUcontext dummy;
187+ CUresult err;
188+ int i;
189+
190+ err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
191+ if (err != CUDA_SUCCESS)
192+ return AVERROR_UNKNOWN;
193+
194+ for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
195+ CUDA_MEMCPY2D cpy = {
196+ .srcMemoryType = CU_MEMORYTYPE_DEVICE,
197+ .dstMemoryType = CU_MEMORYTYPE_HOST,
198+ .srcDevice = (CUdeviceptr)src->data[i],
199+ .dstHost = dst->data[i],
200+ .srcPitch = src->linesize[i],
201+ .dstPitch = dst->linesize[i],
202+ .WidthInBytes = FFMIN(src->linesize[i], dst->linesize[i]),
203+ .Height = src->height >> (i ? priv->shift_height : 0),
204+ };
205+
206+ err = cuMemcpy2D(&cpy);
207+ if (err != CUDA_SUCCESS) {
208+ av_log(ctx, AV_LOG_ERROR, "Error transferring the data from the CUDA frame\n");
209+ return AVERROR_UNKNOWN;
210+ }
211+ }
212+
213+ cuCtxPopCurrent(&dummy);
214+
215+ return 0;
216+}
217+
218+static int cuda_transfer_data_to(AVHWFramesContext *ctx, AVFrame *dst,
219+ const AVFrame *src)
220+{
221+ CUDAFramesContext *priv = ctx->internal->priv;
222+ AVCUDADeviceContext *device_hwctx = ctx->device_ctx->hwctx;
223+
224+ CUcontext dummy;
225+ CUresult err;
226+ int i;
227+
228+ err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
229+ if (err != CUDA_SUCCESS)
230+ return AVERROR_UNKNOWN;
231+
232+ for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
233+ CUDA_MEMCPY2D cpy = {
234+ .srcMemoryType = CU_MEMORYTYPE_HOST,
235+ .dstMemoryType = CU_MEMORYTYPE_DEVICE,
236+ .srcHost = src->data[i],
237+ .dstDevice = (CUdeviceptr)dst->data[i],
238+ .srcPitch = src->linesize[i],
239+ .dstPitch = dst->linesize[i],
240+ .WidthInBytes = FFMIN(src->linesize[i], dst->linesize[i]),
241+ .Height = src->height >> (i ? priv->shift_height : 0),
242+ };
243+
244+ err = cuMemcpy2D(&cpy);
245+ if (err != CUDA_SUCCESS) {
246+ av_log(ctx, AV_LOG_ERROR, "Error transferring the data from the CUDA frame\n");
247+ return AVERROR_UNKNOWN;
248+ }
249+ }
250+
251+ cuCtxPopCurrent(&dummy);
252+
253+ return 0;
254+}
255+
256+const HWContextType ff_hwcontext_type_cuda = {
257+ .type = AV_HWDEVICE_TYPE_CUDA,
258+ .name = "CUDA",
259+
260+ .device_hwctx_size = sizeof(AVCUDADeviceContext),
261+ .frames_priv_size = sizeof(CUDAFramesContext),
262+
263+ .frames_init = cuda_frames_init,
264+ .frames_get_buffer = cuda_get_buffer,
265+ .transfer_get_formats = cuda_transfer_get_formats,
266+ .transfer_data_to = cuda_transfer_data_to,
267+ .transfer_data_from = cuda_transfer_data_from,
268+
269+ .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE },
270+};
libavutil/hwcontext_cuda.h+46−0
@@ -0,0 +1,46 @@
1+/*
2+ * This file is part of Libav.
3+ *
4+ * Libav is free software; you can redistribute it and/or
5+ * modify it under the terms of the GNU Lesser General Public
6+ * License as published by the Free Software Foundation; either
7+ * version 2.1 of the License, or (at your option) any later version.
8+ *
9+ * Libav is distributed in the hope that it will be useful,
10+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
11+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12+ * Lesser General Public License for more details.
13+ *
14+ * You should have received a copy of the GNU Lesser General Public
15+ * License along with Libav; if not, write to the Free Software
16+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17+ */
18+
19+
20+#ifndef AVUTIL_HWCONTEXT_CUDA_H
21+#define AVUTIL_HWCONTEXT_CUDA_H
22+
23+#include <cuda.h>
24+
25+#include "pixfmt.h"
26+
27+/**
28+ * @file
29+ * An API-specific header for AV_HWDEVICE_TYPE_CUDA.
30+ *
31+ * This API supports dynamic frame pools. AVHWFramesContext.pool must return
32+ * AVBufferRefs whose data pointer is a CUdeviceptr.
33+ */
34+
35+/**
36+ * This struct is allocated as AVHWDeviceContext.hwctx
37+ */
38+typedef struct AVCUDADeviceContext {
39+ CUcontext cuda_ctx;
40+} AVCUDADeviceContext;
41+
42+/**
43+ * AVHWFramesContext.hwctx is currently not used
44+ */
45+
46+#endif /* AVUTIL_HWCONTEXT_CUDA_H */
libavutil/hwcontext_internal.h+1−0
@@ -86,6 +86,7 @@ struct AVHWFramesInternal { · 85 unmodified lines
8686 AVBufferPool *pool_internal;
8787 };
8888
89+extern const HWContextType ff_hwcontext_type_cuda;
8990 extern const HWContextType ff_hwcontext_type_vdpau;
9091
9192 #endif /* AVUTIL_HWCONTEXT_INTERNAL_H */

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t