libavutil/hwcontext_cuda.c

   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_P010,
  36     AV_PIX_FMT_YUV444P,
  37     AV_PIX_FMT_YUV444P16,
  38 };
  39
  40 static int cuda_frames_get_constraints(AVHWDeviceContext *ctx,
  41                                        const void *hwconfig,
  42                                        AVHWFramesConstraints *constraints)
  43 {
  44     int i;
  45
  46     constraints->valid_sw_formats = av_malloc_array(FF_ARRAY_ELEMS(supported_formats) + 1,
  47                                                     sizeof(*constraints->valid_sw_formats));
  48     if (!constraints->valid_sw_formats)
  49         return AVERROR(ENOMEM);
  50
  51     for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
  52         constraints->valid_sw_formats[i] = supported_formats[i];
  53     constraints->valid_sw_formats[FF_ARRAY_ELEMS(supported_formats)] = AV_PIX_FMT_NONE;
  54
  55     constraints->valid_hw_formats = av_malloc_array(2, sizeof(*constraints->valid_hw_formats));
  56     if (!constraints->valid_hw_formats)
  57         return AVERROR(ENOMEM);
  58
  59     constraints->valid_hw_formats[0] = AV_PIX_FMT_CUDA;
  60     constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
  61
  62     return 0;
  63 }
  64
  65 static void cuda_buffer_free(void *opaque, uint8_t *data)
  66 {
  67     AVHWFramesContext *ctx = opaque;
  68     AVCUDADeviceContext *hwctx = ctx->device_ctx->hwctx;
  69
  70     CUcontext dummy;
  71
  72     cuCtxPushCurrent(hwctx->cuda_ctx);
  73
  74     cuMemFree((CUdeviceptr)data);
  75
  76     cuCtxPopCurrent(&dummy);
  77 }
  78
  79 static AVBufferRef *cuda_pool_alloc(void *opaque, int size)
  80 {
  81     AVHWFramesContext     *ctx = opaque;
  82     AVCUDADeviceContext *hwctx = ctx->device_ctx->hwctx;
  83
  84     AVBufferRef *ret = NULL;
  85     CUcontext dummy = NULL;
  86     CUdeviceptr data;
  87     CUresult err;
  88
  89     err = cuCtxPushCurrent(hwctx->cuda_ctx);
  90     if (err != CUDA_SUCCESS) {
  91         av_log(ctx, AV_LOG_ERROR, "Error setting current CUDA context\n");
  92         return NULL;
  93     }
  94
  95     err = cuMemAlloc(&data, size);
  96     if (err != CUDA_SUCCESS)
  97         goto fail;
  98
  99     ret = av_buffer_create((uint8_t*)data, size, cuda_buffer_free, ctx, 0);
 100     if (!ret) {
 101         cuMemFree(data);
 102         goto fail;
 103     }
 104
 105 fail:
 106     cuCtxPopCurrent(&dummy);
 107     return ret;
 108 }
 109
 110 static int cuda_frames_init(AVHWFramesContext *ctx)
 111 {
 112     CUDAFramesContext *priv = ctx->internal->priv;
 113     int i;
 114
 115     for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++) {
 116         if (ctx->sw_format == supported_formats[i])
 117             break;
 118     }
 119     if (i == FF_ARRAY_ELEMS(supported_formats)) {
 120         av_log(ctx, AV_LOG_ERROR, "Pixel format '%s' is not supported\n",
 121                av_get_pix_fmt_name(ctx->sw_format));
 122         return AVERROR(ENOSYS);
 123     }
 124
 125     av_pix_fmt_get_chroma_sub_sample(ctx->sw_format, &priv->shift_width, &priv->shift_height);
 126
 127     if (!ctx->pool) {
 128         int size;
 129
 130         switch (ctx->sw_format) {
 131         case AV_PIX_FMT_NV12:
 132         case AV_PIX_FMT_YUV420P:
 133             size = ctx->width * ctx->height * 3 / 2;
 134             break;
 135         case AV_PIX_FMT_P010:
 136             size = ctx->width * ctx->height * 3;
 137             break;
 138         case AV_PIX_FMT_YUV444P:
 139             size = ctx->width * ctx->height * 3;
 140             break;
 141         case AV_PIX_FMT_YUV444P16:
 142             size = ctx->width * ctx->height * 6;
 143             break;
 144         }
 145
 146         ctx->internal->pool_internal = av_buffer_pool_init2(size, ctx, cuda_pool_alloc, NULL);
 147         if (!ctx->internal->pool_internal)
 148             return AVERROR(ENOMEM);
 149     }
 150
 151     return 0;
 152 }
 153
 154 static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
 155 {
 156     frame->buf[0] = av_buffer_pool_get(ctx->pool);
 157     if (!frame->buf[0])
 158         return AVERROR(ENOMEM);
 159
 160     switch (ctx->sw_format) {
 161     case AV_PIX_FMT_NV12:
 162         frame->data[0]     = frame->buf[0]->data;
 163         frame->data[1]     = frame->data[0] + ctx->width * ctx->height;
 164         frame->linesize[0] = ctx->width;
 165         frame->linesize[1] = ctx->width;
 166         break;
 167     case AV_PIX_FMT_YUV420P:
 168         frame->data[0]     = frame->buf[0]->data;
 169         frame->data[2]     = frame->data[0] + ctx->width * ctx->height;
 170         frame->data[1]     = frame->data[2] + ctx->width * ctx->height / 4;
 171         frame->linesize[0] = ctx->width;
 172         frame->linesize[1] = ctx->width / 2;
 173         frame->linesize[2] = ctx->width / 2;
 174         break;
 175     case AV_PIX_FMT_P010:
 176         frame->data[0]     = frame->buf[0]->data;
 177         frame->data[1]     = frame->data[0] + 2 * ctx->width * ctx->height;
 178         frame->linesize[0] = 2 * ctx->width;
 179         frame->linesize[1] = 2 * ctx->width;
 180         break;
 181     case AV_PIX_FMT_YUV444P:
 182         frame->data[0]     = frame->buf[0]->data;
 183         frame->data[1]     = frame->data[0] + ctx->width * ctx->height;
 184         frame->data[2]     = frame->data[1] + ctx->width * ctx->height;
 185         frame->linesize[0] = ctx->width;
 186         frame->linesize[1] = ctx->width;
 187         frame->linesize[2] = ctx->width;
 188         break;
 189     case AV_PIX_FMT_YUV444P16:
 190         frame->data[0]     = frame->buf[0]->data;
 191         frame->data[1]     = frame->data[0] + 2 * ctx->width * ctx->height;
 192         frame->data[2]     = frame->data[1] + 2 * ctx->width * ctx->height;
 193         frame->linesize[0] = 2 * ctx->width;
 194         frame->linesize[1] = 2 * ctx->width;
 195         frame->linesize[2] = 2 * ctx->width;
 196         break;
 197     default:
 198         av_frame_unref(frame);
 199         return AVERROR_BUG;
 200     }
 201
 202     frame->format = AV_PIX_FMT_CUDA;
 203     frame->width  = ctx->width;
 204     frame->height = ctx->height;
 205
 206     return 0;
 207 }
 208
 209 static int cuda_transfer_get_formats(AVHWFramesContext *ctx,
 210                                      enum AVHWFrameTransferDirection dir,
 211                                      enum AVPixelFormat **formats)
 212 {
 213     enum AVPixelFormat *fmts;
 214
 215     fmts = av_malloc_array(2, sizeof(*fmts));
 216     if (!fmts)
 217         return AVERROR(ENOMEM);
 218
 219     fmts[0] = ctx->sw_format;
 220     fmts[1] = AV_PIX_FMT_NONE;
 221
 222     *formats = fmts;
 223
 224     return 0;
 225 }
 226
 227 static int cuda_transfer_data_from(AVHWFramesContext *ctx, AVFrame *dst,
 228                                    const AVFrame *src)
 229 {
 230     CUDAFramesContext           *priv = ctx->internal->priv;
 231     AVCUDADeviceContext *device_hwctx = ctx->device_ctx->hwctx;
 232
 233     CUcontext dummy;
 234     CUresult err;
 235     int i;
 236
 237     err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
 238     if (err != CUDA_SUCCESS)
 239         return AVERROR_UNKNOWN;
 240
 241     for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
 242         CUDA_MEMCPY2D cpy = {
 243             .srcMemoryType = CU_MEMORYTYPE_DEVICE,
 244             .dstMemoryType = CU_MEMORYTYPE_HOST,
 245             .srcDevice     = (CUdeviceptr)src->data[i],
 246             .dstHost       = dst->data[i],
 247             .srcPitch      = src->linesize[i],
 248             .dstPitch      = dst->linesize[i],
 249             .WidthInBytes  = FFMIN(src->linesize[i], dst->linesize[i]),
 250             .Height        = src->height >> (i ? priv->shift_height : 0),
 251         };
 252
 253         err = cuMemcpy2D(&cpy);
 254         if (err != CUDA_SUCCESS) {
 255             av_log(ctx, AV_LOG_ERROR, "Error transferring the data from the CUDA frame\n");
 256             return AVERROR_UNKNOWN;
 257         }
 258     }
 259
 260     cuCtxPopCurrent(&dummy);
 261
 262     return 0;
 263 }
 264
 265 static int cuda_transfer_data_to(AVHWFramesContext *ctx, AVFrame *dst,
 266                                  const AVFrame *src)
 267 {
 268     CUDAFramesContext           *priv = ctx->internal->priv;
 269     AVCUDADeviceContext *device_hwctx = ctx->device_ctx->hwctx;
 270
 271     CUcontext dummy;
 272     CUresult err;
 273     int i;
 274
 275     err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
 276     if (err != CUDA_SUCCESS)
 277         return AVERROR_UNKNOWN;
 278
 279     for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
 280         CUDA_MEMCPY2D cpy = {
 281             .srcMemoryType = CU_MEMORYTYPE_HOST,
 282             .dstMemoryType = CU_MEMORYTYPE_DEVICE,
 283             .srcHost       = src->data[i],
 284             .dstDevice     = (CUdeviceptr)dst->data[i],
 285             .srcPitch      = src->linesize[i],
 286             .dstPitch      = dst->linesize[i],
 287             .WidthInBytes  = FFMIN(src->linesize[i], dst->linesize[i]),
 288             .Height        = src->height >> (i ? priv->shift_height : 0),
 289         };
 290
 291         err = cuMemcpy2D(&cpy);
 292         if (err != CUDA_SUCCESS) {
 293             av_log(ctx, AV_LOG_ERROR, "Error transferring the data from the CUDA frame\n");
 294             return AVERROR_UNKNOWN;
 295         }
 296     }
 297
 298     cuCtxPopCurrent(&dummy);
 299
 300     return 0;
 301 }
 302
 303 static void cuda_device_free(AVHWDeviceContext *ctx)
 304 {
 305     AVCUDADeviceContext *hwctx = ctx->hwctx;
 306     cuCtxDestroy(hwctx->cuda_ctx);
 307 }
 308
 309 static int cuda_device_create(AVHWDeviceContext *ctx, const char *device,
 310                               AVDictionary *opts, int flags)
 311 {
 312     AVCUDADeviceContext *hwctx = ctx->hwctx;
 313     CUdevice cu_device;
 314     CUcontext dummy;
 315     CUresult err;
 316     int device_idx = 0;
 317
 318     if (device)
 319         device_idx = strtol(device, NULL, 0);
 320
 321     err = cuInit(0);
 322     if (err != CUDA_SUCCESS) {
 323         av_log(ctx, AV_LOG_ERROR, "Could not initialize the CUDA driver API\n");
 324         return AVERROR_UNKNOWN;
 325     }
 326
 327     err = cuDeviceGet(&cu_device, device_idx);
 328     if (err != CUDA_SUCCESS) {
 329         av_log(ctx, AV_LOG_ERROR, "Could not get the device number %d\n", device_idx);
 330         return AVERROR_UNKNOWN;
 331     }
 332
 333     err = cuCtxCreate(&hwctx->cuda_ctx, 0, cu_device);
 334     if (err != CUDA_SUCCESS) {
 335         av_log(ctx, AV_LOG_ERROR, "Error creating a CUDA context\n");
 336         return AVERROR_UNKNOWN;
 337     }
 338
 339     cuCtxPopCurrent(&dummy);
 340
 341     ctx->free = cuda_device_free;
 342
 343     return 0;
 344 }
 345
 346 const HWContextType ff_hwcontext_type_cuda = {
 347     .type                 = AV_HWDEVICE_TYPE_CUDA,
 348     .name                 = "CUDA",
 349
 350     .device_hwctx_size    = sizeof(AVCUDADeviceContext),
 351     .frames_priv_size     = sizeof(CUDAFramesContext),
 352
 353     .device_create        = cuda_device_create,
 354     .frames_get_constraints = cuda_frames_get_constraints,
 355     .frames_init          = cuda_frames_init,
 356     .frames_get_buffer    = cuda_get_buffer,
 357     .transfer_get_formats = cuda_transfer_get_formats,
 358     .transfer_data_to     = cuda_transfer_data_to,
 359     .transfer_data_from   = cuda_transfer_data_from,
 360
 361     .pix_fmts             = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE },
 362 };