2 * This file is part of FFmpeg.
4 * FFmpeg 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.
9 * FFmpeg 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.
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with FFmpeg; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "hwcontext.h"
22 #include "hwcontext_internal.h"
23 #include "hwcontext_cuda_internal.h"
24 #include "cuda_check.h"
30 #define CUDA_FRAME_ALIGNMENT 256
32 typedef struct CUDAFramesContext {
33 int shift_width, shift_height;
36 static const enum AVPixelFormat supported_formats[] = {
47 #define CHECK_CU(x) FF_CUDA_CHECK_DL(device_ctx, cu, x)
49 static int cuda_frames_get_constraints(AVHWDeviceContext *ctx,
51 AVHWFramesConstraints *constraints)
55 constraints->valid_sw_formats = av_malloc_array(FF_ARRAY_ELEMS(supported_formats) + 1,
56 sizeof(*constraints->valid_sw_formats));
57 if (!constraints->valid_sw_formats)
58 return AVERROR(ENOMEM);
60 for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
61 constraints->valid_sw_formats[i] = supported_formats[i];
62 constraints->valid_sw_formats[FF_ARRAY_ELEMS(supported_formats)] = AV_PIX_FMT_NONE;
64 constraints->valid_hw_formats = av_malloc_array(2, sizeof(*constraints->valid_hw_formats));
65 if (!constraints->valid_hw_formats)
66 return AVERROR(ENOMEM);
68 constraints->valid_hw_formats[0] = AV_PIX_FMT_CUDA;
69 constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
74 static void cuda_buffer_free(void *opaque, uint8_t *data)
76 AVHWFramesContext *ctx = opaque;
77 AVHWDeviceContext *device_ctx = ctx->device_ctx;
78 AVCUDADeviceContext *hwctx = device_ctx->hwctx;
79 CudaFunctions *cu = hwctx->internal->cuda_dl;
83 CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
85 CHECK_CU(cu->cuMemFree((CUdeviceptr)data));
87 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
90 static AVBufferRef *cuda_pool_alloc(void *opaque, int size)
92 AVHWFramesContext *ctx = opaque;
93 AVHWDeviceContext *device_ctx = ctx->device_ctx;
94 AVCUDADeviceContext *hwctx = device_ctx->hwctx;
95 CudaFunctions *cu = hwctx->internal->cuda_dl;
97 AVBufferRef *ret = NULL;
98 CUcontext dummy = NULL;
102 err = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
106 err = CHECK_CU(cu->cuMemAlloc(&data, size));
110 ret = av_buffer_create((uint8_t*)data, size, cuda_buffer_free, ctx, 0);
112 CHECK_CU(cu->cuMemFree(data));
117 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
121 static int cuda_frames_init(AVHWFramesContext *ctx)
123 CUDAFramesContext *priv = ctx->internal->priv;
126 for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++) {
127 if (ctx->sw_format == supported_formats[i])
130 if (i == FF_ARRAY_ELEMS(supported_formats)) {
131 av_log(ctx, AV_LOG_ERROR, "Pixel format '%s' is not supported\n",
132 av_get_pix_fmt_name(ctx->sw_format));
133 return AVERROR(ENOSYS);
136 av_pix_fmt_get_chroma_sub_sample(ctx->sw_format, &priv->shift_width, &priv->shift_height);
139 int size = av_image_get_buffer_size(ctx->sw_format, ctx->width, ctx->height, CUDA_FRAME_ALIGNMENT);
143 ctx->internal->pool_internal = av_buffer_pool_init2(size, ctx, cuda_pool_alloc, NULL);
144 if (!ctx->internal->pool_internal)
145 return AVERROR(ENOMEM);
151 static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
155 frame->buf[0] = av_buffer_pool_get(ctx->pool);
157 return AVERROR(ENOMEM);
159 res = av_image_fill_arrays(frame->data, frame->linesize, frame->buf[0]->data,
160 ctx->sw_format, ctx->width, ctx->height, CUDA_FRAME_ALIGNMENT);
164 // YUV420P is a special case.
165 // Nvenc expects the U/V planes in swapped order from how ffmpeg expects them, also chroma is half-aligned
166 if (ctx->sw_format == AV_PIX_FMT_YUV420P) {
167 frame->linesize[1] = frame->linesize[2] = frame->linesize[0] / 2;
168 frame->data[2] = frame->data[1];
169 frame->data[1] = frame->data[2] + frame->linesize[2] * ctx->height / 2;
172 frame->format = AV_PIX_FMT_CUDA;
173 frame->width = ctx->width;
174 frame->height = ctx->height;
179 static int cuda_transfer_get_formats(AVHWFramesContext *ctx,
180 enum AVHWFrameTransferDirection dir,
181 enum AVPixelFormat **formats)
183 enum AVPixelFormat *fmts;
185 fmts = av_malloc_array(2, sizeof(*fmts));
187 return AVERROR(ENOMEM);
189 fmts[0] = ctx->sw_format;
190 fmts[1] = AV_PIX_FMT_NONE;
197 static int cuda_transfer_data_from(AVHWFramesContext *ctx, AVFrame *dst,
200 CUDAFramesContext *priv = ctx->internal->priv;
201 AVHWDeviceContext *device_ctx = ctx->device_ctx;
202 AVCUDADeviceContext *hwctx = device_ctx->hwctx;
203 CudaFunctions *cu = hwctx->internal->cuda_dl;
208 ret = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
212 for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
213 CUDA_MEMCPY2D cpy = {
214 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
215 .dstMemoryType = CU_MEMORYTYPE_HOST,
216 .srcDevice = (CUdeviceptr)src->data[i],
217 .dstHost = dst->data[i],
218 .srcPitch = src->linesize[i],
219 .dstPitch = dst->linesize[i],
220 .WidthInBytes = FFMIN(src->linesize[i], dst->linesize[i]),
221 .Height = src->height >> (i ? priv->shift_height : 0),
224 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, hwctx->stream));
229 ret = CHECK_CU(cu->cuStreamSynchronize(hwctx->stream));
234 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
239 static int cuda_transfer_data_to(AVHWFramesContext *ctx, AVFrame *dst,
242 CUDAFramesContext *priv = ctx->internal->priv;
243 AVHWDeviceContext *device_ctx = ctx->device_ctx;
244 AVCUDADeviceContext *hwctx = device_ctx->hwctx;
245 CudaFunctions *cu = hwctx->internal->cuda_dl;
250 ret = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
254 for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
255 CUDA_MEMCPY2D cpy = {
256 .srcMemoryType = CU_MEMORYTYPE_HOST,
257 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
258 .srcHost = src->data[i],
259 .dstDevice = (CUdeviceptr)dst->data[i],
260 .srcPitch = src->linesize[i],
261 .dstPitch = dst->linesize[i],
262 .WidthInBytes = FFMIN(src->linesize[i], dst->linesize[i]),
263 .Height = src->height >> (i ? priv->shift_height : 0),
266 ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, hwctx->stream));
272 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
277 static void cuda_device_uninit(AVHWDeviceContext *device_ctx)
279 AVCUDADeviceContext *hwctx = device_ctx->hwctx;
281 if (hwctx->internal) {
282 CudaFunctions *cu = hwctx->internal->cuda_dl;
283 if (hwctx->internal->is_allocated && hwctx->cuda_ctx) {
284 CHECK_CU(cu->cuCtxDestroy(hwctx->cuda_ctx));
285 hwctx->cuda_ctx = NULL;
287 cuda_free_functions(&hwctx->internal->cuda_dl);
290 av_freep(&hwctx->internal);
293 static int cuda_device_init(AVHWDeviceContext *ctx)
295 AVCUDADeviceContext *hwctx = ctx->hwctx;
298 if (!hwctx->internal) {
299 hwctx->internal = av_mallocz(sizeof(*hwctx->internal));
300 if (!hwctx->internal)
301 return AVERROR(ENOMEM);
304 if (!hwctx->internal->cuda_dl) {
305 ret = cuda_load_functions(&hwctx->internal->cuda_dl, ctx);
307 av_log(ctx, AV_LOG_ERROR, "Could not dynamically load CUDA\n");
315 cuda_device_uninit(ctx);
319 static int cuda_device_create(AVHWDeviceContext *device_ctx,
321 AVDictionary *opts, int flags)
323 AVCUDADeviceContext *hwctx = device_ctx->hwctx;
327 int ret, device_idx = 0;
330 device_idx = strtol(device, NULL, 0);
332 if (cuda_device_init(device_ctx) < 0)
335 cu = hwctx->internal->cuda_dl;
337 ret = CHECK_CU(cu->cuInit(0));
341 ret = CHECK_CU(cu->cuDeviceGet(&cu_device, device_idx));
345 ret = CHECK_CU(cu->cuCtxCreate(&hwctx->cuda_ctx, CU_CTX_SCHED_BLOCKING_SYNC, cu_device));
349 // Setting stream to NULL will make functions automatically use the default CUstream
350 hwctx->stream = NULL;
352 CHECK_CU(cu->cuCtxPopCurrent(&dummy));
354 hwctx->internal->is_allocated = 1;
359 cuda_device_uninit(device_ctx);
360 return AVERROR_UNKNOWN;
363 const HWContextType ff_hwcontext_type_cuda = {
364 .type = AV_HWDEVICE_TYPE_CUDA,
367 .device_hwctx_size = sizeof(AVCUDADeviceContext),
368 .frames_priv_size = sizeof(CUDAFramesContext),
370 .device_create = cuda_device_create,
371 .device_init = cuda_device_init,
372 .device_uninit = cuda_device_uninit,
373 .frames_get_constraints = cuda_frames_get_constraints,
374 .frames_init = cuda_frames_init,
375 .frames_get_buffer = cuda_get_buffer,
376 .transfer_get_formats = cuda_transfer_get_formats,
377 .transfer_data_to = cuda_transfer_data_to,
378 .transfer_data_from = cuda_transfer_data_from,
380 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE },