1
0
mirror of https://github.com/opencv/opencv.git synced 2026-07-31 08:13:04 +04:00

Merge pull request #19755 from mikhail-nikolskiy:ffmpeg-umat

cv::UMat output/input in VideoCapture/VideoWriter (data stays in GPU memory)

* FFMPEG with UMat input/output

* OpenCL_D3D* context

* fix Linux build

* cosmetic changes

* fix build if USE_AV_HW_CODECS=0

* simplify how child context pointer stored in parent context

* QSV interop with OpenCL on Windows

* detect_msdk.cmake via pkg-config

* fix av_buffer_ref() usage

* revert windows-decode-mfx whitelisting; remove debug msg

* address review comments

* rename property to HW_ACCELERATION_USE_OPENCL

* fix issue with "cl_khr_d3d11_sharing" extension not reported by OpenCL GPU+CPU platform

* core(ocl): add OpenCL stubs for configurations without OpenCL

* videoio(ffmpeg): update #if guards

* Put OpenCL related code under HAVE_OPENCL; simplify reuse of media context from OpenCL context

* videoio(test): skip unsupported tests

- plugins don't support OpenCL/UMat yet
- change handling of *_USE_OPENCL flag

* videoio(ffmpeg): OpenCL dependency

* videoio(ffmpeg): MediaSDK/oneVPL dependency

* cleanup, logging

* cmake: fix handling of 3rdparty interface targets

Co-authored-by: Alexander Alekhin <alexander.a.alekhin@gmail.com>
This commit is contained in:
Mikhail Nikolskii
2021-05-14 19:48:50 +03:00
committed by GitHub
parent bb92eb5a93
commit a604d44d06
17 changed files with 1347 additions and 783 deletions
+18 -2
View File
@@ -92,8 +92,17 @@ public:
unsigned char* data = 0;
int step=0, width=0, height=0, cn=0;
if (!ffmpegCapture ||
!icvRetrieveFrame_FFMPEG_p(ffmpegCapture, &data, &step, &width, &height, &cn))
if (!ffmpegCapture)
return false;
// if UMat, try GPU to GPU copy using OpenCL extensions
if (frame.isUMat()) {
if (ffmpegCapture->retrieveHWFrame(frame)) {
return true;
}
}
if (!icvRetrieveFrame_FFMPEG_p(ffmpegCapture, &data, &step, &width, &height, &cn))
return false;
cv::Mat tmp(height, width, CV_MAKETYPE(CV_8U, cn), data, step);
@@ -176,6 +185,13 @@ public:
return;
CV_Assert(image.depth() == CV_8U);
// if UMat, try GPU to GPU copy using OpenCL extensions
if (image.isUMat()) {
if (ffmpegWriter->writeHWFrame(image)) {
return;
}
}
icvWriteFrame_FFMPEG_p(ffmpegWriter, (const uchar*)image.getMat().ptr(), (int)image.step(), image.cols(), image.rows(), image.channels(), 0);
}
virtual bool open( const cv::String& filename, int fourcc, double fps, cv::Size frameSize, const VideoWriterParameters& params )
+518 -80
View File
@@ -5,7 +5,10 @@
// Copyright (C) 2020-2021 Intel Corporation
#include "opencv2/videoio.hpp"
#if defined(__OPENCV_BUILD) || defined(OPENCV_HAVE_CVCONFIG_H) // TODO Properly detect and add D3D11 / LIBVA dependencies for standalone plugins
#ifdef HAVE_OPENCL
#include "opencv2/core/ocl.hpp"
#endif
#if defined(__OPENCV_BUILD) && !defined(BUILD_PLUGIN) // TODO Properly detect and add D3D11 / LIBVA dependencies for standalone plugins
#include "cvconfig.h"
#endif
#include <sstream>
@@ -14,16 +17,31 @@
#define D3D11_NO_HELPERS
#include <d3d11.h>
#include <codecvt>
#include "opencv2/core/directx.hpp"
#ifdef HAVE_OPENCL
#include <CL/cl_d3d11.h>
#endif
#endif // HAVE_D3D11
#ifdef HAVE_VA
#include <va/va_backend.h>
#ifdef HAVE_VA_INTEL
#include "opencv2/core/va_intel.hpp"
#ifndef CL_TARGET_OPENCL_VERSION
#define CL_TARGET_OPENCL_VERSION 120
#endif
#ifdef HAVE_VA_INTEL_OLD_HEADER
#include <CL/va_ext.h>
#else
#include <CL/cl_va_api_media_sharing_intel.h>
#endif
#endif
#endif // HAVE_VA
// FFMPEG "C" headers
extern "C" {
#include <libavcodec/avcodec.h>
#include <libavutil/avutil.h>
#include <libavutil/hwcontext.h>
#ifdef HAVE_D3D11
#include <libavutil/hwcontext_d3d11va.h>
@@ -31,8 +49,23 @@ extern "C" {
#ifdef HAVE_VA
#include <libavutil/hwcontext_vaapi.h>
#endif
#ifdef HAVE_MFX // dependency only on MFX header files, no linkage dependency
#include <libavutil/hwcontext_qsv.h>
#endif
}
#define HW_DEFAULT_POOL_SIZE 32
#define HW_DEFAULT_SW_FORMAT AV_PIX_FMT_NV12
using namespace cv;
static AVCodec *hw_find_codec(AVCodecID id, AVHWDeviceType hw_type, int (*check_category)(const AVCodec *),
const char *disabled_codecs, AVPixelFormat *hw_pix_fmt);
static AVBufferRef* hw_create_device(AVHWDeviceType hw_type, int hw_device, const std::string& device_subname, bool use_opencl);
static AVBufferRef* hw_create_frames(struct AVCodecContext* ctx, AVBufferRef *hw_device_ctx, int width, int height, AVPixelFormat hw_format);
static AVPixelFormat hw_get_format_callback(struct AVCodecContext *ctx, const enum AVPixelFormat * fmt);
static VideoAccelerationType hw_type_to_va_type(AVHWDeviceType hw_type);
static
const char* getVideoAccelerationName(VideoAccelerationType va_type)
{
@@ -70,7 +103,7 @@ std::string getDecoderConfiguration(VideoAccelerationType va_type, AVDictionary
case VIDEO_ACCELERATION_ANY: return "d3d11va";
case VIDEO_ACCELERATION_D3D11: return "d3d11va";
case VIDEO_ACCELERATION_VAAPI: return "";
case VIDEO_ACCELERATION_MFX: return "";
case VIDEO_ACCELERATION_MFX: return ""; // "qsv" fails if non-Intel D3D11 device
}
return "";
#else
@@ -80,7 +113,7 @@ std::string getDecoderConfiguration(VideoAccelerationType va_type, AVDictionary
case VIDEO_ACCELERATION_ANY: return "vaapi.iHD";
case VIDEO_ACCELERATION_D3D11: return "";
case VIDEO_ACCELERATION_VAAPI: return "vaapi.iHD";
case VIDEO_ACCELERATION_MFX: return "";
case VIDEO_ACCELERATION_MFX: return "qsv.iHD";
}
return "";
#endif
@@ -125,7 +158,6 @@ std::string getEncoderConfiguration(VideoAccelerationType va_type, AVDictionary
#endif
}
static
std::string getDecoderDisabledCodecs(AVDictionary *dict)
{
@@ -170,19 +202,6 @@ std::string getEncoderDisabledCodecs(AVDictionary *dict)
#endif
}
#define HW_DEFAULT_POOL_SIZE 32
#define HW_DEFAULT_SW_FORMAT AV_PIX_FMT_NV12
using namespace cv;
static AVCodec *hw_find_codec(AVCodecID id, AVHWDeviceType hw_type, int (*check_category)(const AVCodec *),
const char *disabled_codecs, AVPixelFormat *hw_pix_fmt);
static AVBufferRef* hw_create_device(AVHWDeviceType hw_type, int hw_device, const std::string& device_subname);
static AVBufferRef* hw_create_frames(struct AVCodecContext* ctx, AVBufferRef *hw_device_ctx, int width, int height, AVPixelFormat hw_format);
static AVPixelFormat hw_get_format_callback(struct AVCodecContext *ctx, const enum AVPixelFormat * fmt);
static VideoAccelerationType hw_type_to_va_type(AVHWDeviceType hw_type);
static
bool hw_check_device(AVBufferRef* ctx, AVHWDeviceType hw_type, const std::string& device_subname) {
if (!ctx)
@@ -259,75 +278,343 @@ bool hw_check_device(AVBufferRef* ctx, AVHWDeviceType hw_type, const std::string
}
static
AVBufferRef* hw_create_device(AVHWDeviceType hw_type, int hw_device, const std::string& device_subname) {
if (AV_HWDEVICE_TYPE_NONE == hw_type)
return NULL;
AVHWDeviceType child_type = hw_type;
if (hw_type == AV_HWDEVICE_TYPE_QSV) {
#ifdef _WIN32
child_type = AV_HWDEVICE_TYPE_DXVA2;
#else
child_type = AV_HWDEVICE_TYPE_VAAPI;
#endif
}
AVBufferRef* hw_device_ctx = NULL;
char device[128] = "";
char* pdevice = NULL;
if (hw_device >= 0 && hw_device < 100000) {
if (child_type == AV_HWDEVICE_TYPE_VAAPI) {
snprintf(device, sizeof(device), "/dev/dri/renderD%d", 128 + hw_device);
} else {
snprintf(device, sizeof(device), "%d", hw_device);
}
pdevice = device;
}
const char *hw_child_name = av_hwdevice_get_type_name(child_type);
const char *device_name = pdevice ? pdevice : "'default'";
int err = av_hwdevice_ctx_create(&hw_device_ctx, child_type, pdevice, NULL, 0);
if (hw_device_ctx && err >= 0)
AVBufferRef* hw_create_derived_context(AVHWDeviceType hw_type, AVBufferRef* hw_device_ctx) {
AVBufferRef* derived_ctx = NULL;
const char* hw_name = av_hwdevice_get_type_name(hw_type);
int err = av_hwdevice_ctx_create_derived(&derived_ctx, hw_type, hw_device_ctx, 0);
if (!derived_ctx || err < 0)
{
CV_LOG_DEBUG(NULL, "FFMPEG: Created video acceleration context (av_hwdevice_ctx_create) for " << hw_child_name << " on device " << device_name);
if (!hw_check_device(hw_device_ctx, hw_type, device_subname)) {
av_buffer_unref(&hw_device_ctx);
return NULL;
}
if (hw_type != child_type) {
AVBufferRef *derived_ctx = NULL;
const char *hw_name = av_hwdevice_get_type_name(hw_type);
err = av_hwdevice_ctx_create_derived(&derived_ctx, hw_type, hw_device_ctx, 0);
if (!derived_ctx || err < 0)
{
if (derived_ctx)
av_buffer_unref(&derived_ctx);
CV_LOG_INFO(NULL, "FFMPEG: Failed to create derived video acceleration (av_hwdevice_ctx_create_derived) for " << hw_name << ". Error=" << err);
}
else
{
CV_LOG_DEBUG(NULL, "FFMPEG: Created derived video acceleration context (av_hwdevice_ctx_create_derived) for " << hw_name);
}
av_buffer_unref(&hw_device_ctx);
return derived_ctx;
} else {
return hw_device_ctx;
}
if (derived_ctx)
av_buffer_unref(&derived_ctx);
CV_LOG_INFO(NULL, "FFMPEG: Failed to create derived video acceleration (av_hwdevice_ctx_create_derived) for " << hw_name << ". Error=" << err);
return NULL;
}
else
{
const char *hw_name = hw_child_name;
CV_LOG_INFO(NULL, "FFMPEG: Failed to create " << hw_name << " video acceleration (av_hwdevice_ctx_create) on device " << device_name);
// Store child context in 'user_opaque' field of parent context.
struct FreeChildContext {
static void free(struct AVHWDeviceContext* ctx) {
AVBufferRef* child_ctx = (AVBufferRef*)ctx->user_opaque;
if (child_ctx)
av_buffer_unref(&child_ctx);
}
};
AVHWDeviceContext* ctx = (AVHWDeviceContext*)derived_ctx->data;
ctx->user_opaque = av_buffer_ref(hw_device_ctx);
ctx->free = FreeChildContext::free;
CV_LOG_INFO(NULL, "FFMPEG: Created derived video acceleration context (av_hwdevice_ctx_create_derived) for " << hw_name);
return derived_ctx;
}
}
#ifdef HAVE_OPENCL // GPU buffer interop with cv::UMat
// FFmpeg context attached to OpenCL context
class OpenCL_FFMPEG_Context : public ocl::Context::UserContext {
public:
OpenCL_FFMPEG_Context(AVBufferRef* ctx) {
ctx_ = av_buffer_ref(ctx);
}
virtual ~OpenCL_FFMPEG_Context() {
av_buffer_unref(&ctx_);
}
AVBufferRef* GetAVHWDevice() {
return ctx_;
}
private:
AVBufferRef* ctx_;
};
#ifdef HAVE_MFX
static
int hw_find_qsv_surface_index(AVFrame* hw_frame)
{
if (AV_PIX_FMT_QSV != hw_frame->format)
return -1;
mfxFrameSurface1* surface = (mfxFrameSurface1*)hw_frame->data[3]; // As defined by AV_PIX_FMT_QSV
AVHWFramesContext* frames_ctx = (AVHWFramesContext*)hw_frame->hw_frames_ctx->data;
AVQSVFramesContext* qsv_ctx = (AVQSVFramesContext*)frames_ctx->hwctx;
for (int i = 0; i < qsv_ctx->nb_surfaces; i++) {
if (surface == qsv_ctx->surfaces + i) {
return i;
}
}
return -1;
}
#endif
#ifdef HAVE_VA
static
VADisplay hw_get_va_display(AVHWDeviceContext* hw_device_ctx)
{
if (hw_device_ctx->type == AV_HWDEVICE_TYPE_QSV) { // we stored pointer to child context in 'user_opaque' field
AVBufferRef* ctx = (AVBufferRef*)hw_device_ctx->user_opaque;
hw_device_ctx = (AVHWDeviceContext*)ctx->data;
}
if (hw_device_ctx && hw_device_ctx->type == AV_HWDEVICE_TYPE_VAAPI) {
return ((AVVAAPIDeviceContext*)hw_device_ctx->hwctx)->display;
}
return NULL;
}
#endif // HAVE_VA
#ifdef HAVE_VA_INTEL
static
VASurfaceID hw_get_va_surface(AVFrame* hw_frame) {
if (AV_PIX_FMT_VAAPI == hw_frame->format) {
return (VASurfaceID)(size_t)hw_frame->data[3]; // As defined by AV_PIX_FMT_VAAPI
}
#ifdef HAVE_MFX
else if (AV_PIX_FMT_QSV == hw_frame->format) {
int frame_idx = hw_find_qsv_surface_index(hw_frame);
if (frame_idx >= 0) { // frame index is same in parent (QSV) and child (VAAPI) frame context
AVHWFramesContext *frames_ctx = (AVHWFramesContext *) hw_frame->hw_frames_ctx->data;
AVHWFramesContext *child_ctx = (AVHWFramesContext *) frames_ctx->user_opaque;
if (child_ctx && AV_HWDEVICE_TYPE_VAAPI == child_ctx->device_ctx->type) {
AVVAAPIFramesContext *vaapi_ctx = (AVVAAPIFramesContext *) child_ctx->hwctx;
CV_Assert(frame_idx < vaapi_ctx->nb_surfaces);
return vaapi_ctx->surface_ids[frame_idx];
}
}
}
#endif // HAVE_MFX
return VA_INVALID_SURFACE;
}
#endif // HAVE_VA_INTEL
#ifdef HAVE_D3D11
static
AVD3D11VADeviceContext* hw_get_d3d11_device_ctx(AVHWDeviceContext* hw_device_ctx) {
if (AV_HWDEVICE_TYPE_QSV == hw_device_ctx->type) { // we stored pointer to child context in 'user_opaque' field
AVBufferRef* ctx = (AVBufferRef*)hw_device_ctx->user_opaque;
hw_device_ctx = (AVHWDeviceContext*)ctx->data;
}
if (AV_HWDEVICE_TYPE_D3D11VA == hw_device_ctx->type) {
return (AVD3D11VADeviceContext*)hw_device_ctx->hwctx;
}
return NULL;
}
ID3D11Texture2D* hw_get_d3d11_texture(AVFrame* hw_frame, int* subresource) {
ID3D11Texture2D* texture = NULL;
if (AV_PIX_FMT_D3D11 == hw_frame->format) {
texture = (ID3D11Texture2D*)hw_frame->data[0]; // As defined by AV_PIX_FMT_D3D11
*subresource = (intptr_t)hw_frame->data[1]; // As defined by AV_PIX_FMT_D3D11
}
#ifdef HAVE_MFX
else if (AV_PIX_FMT_QSV == hw_frame->format) {
AVHWFramesContext *frames_ctx = (AVHWFramesContext *) hw_frame->hw_frames_ctx->data;
AVHWFramesContext *child_ctx = (AVHWFramesContext *) frames_ctx->user_opaque;
if (child_ctx && AV_HWDEVICE_TYPE_D3D11VA == child_ctx->device_ctx->type) {
texture = ((AVD3D11VAFramesContext*)child_ctx->hwctx)->texture;
}
*subresource = hw_find_qsv_surface_index(hw_frame);
CV_Assert(*subresource >= 0);
}
#endif
return texture;
}
// In D3D11 case we allocate additional texture as single texture (not texture array) because
// OpenCL interop with D3D11 doesn't support/work with NV12 sub-texture of texture array.
ID3D11Texture2D* hw_get_d3d11_single_texture(AVFrame* hw_frame, AVD3D11VADeviceContext* d3d11_device_ctx, ID3D11Texture2D* texture) {
AVHWFramesContext* frames_ctx = (AVHWFramesContext*)hw_frame->hw_frames_ctx->data;
if (AV_HWDEVICE_TYPE_QSV == frames_ctx->device_ctx->type) {
frames_ctx = (AVHWFramesContext*)frames_ctx->user_opaque; // we stored pointer to child context in 'user_opaque' field
}
if (!frames_ctx || AV_HWDEVICE_TYPE_D3D11VA != frames_ctx->device_ctx->type) {
return NULL;
}
ID3D11Texture2D* singleTexture = (ID3D11Texture2D*)frames_ctx->user_opaque;
if (!singleTexture && d3d11_device_ctx && texture) {
D3D11_TEXTURE2D_DESC desc = {};
texture->GetDesc(&desc);
desc.ArraySize = 1;
desc.BindFlags |= D3D11_BIND_SHADER_RESOURCE;
desc.MiscFlags |= D3D11_RESOURCE_MISC_SHARED;
if (SUCCEEDED(d3d11_device_ctx->device->CreateTexture2D(&desc, NULL, &singleTexture))) {
frames_ctx->user_opaque = singleTexture;
}
}
return singleTexture;
}
#endif // HAVE_D3D11
static
AVHWDeviceType hw_check_opencl_context(AVHWDeviceContext* ctx) {
ocl::OpenCLExecutionContext& ocl_context = ocl::OpenCLExecutionContext::getCurrentRef();
if (!ctx || ocl_context.empty())
return AV_HWDEVICE_TYPE_NONE;
#ifdef HAVE_VA_INTEL
VADisplay vadisplay_ocl = ocl_context.getContext().getOpenCLContextProperty(CL_CONTEXT_VA_API_DISPLAY_INTEL);
VADisplay vadisplay_ctx = hw_get_va_display(ctx);
if (vadisplay_ocl && vadisplay_ocl == vadisplay_ctx)
return AV_HWDEVICE_TYPE_VAAPI;
#endif
#ifdef HAVE_D3D11
ID3D11Device* d3d11device_ocl = (ID3D11Device*)ocl_context.getContext().getOpenCLContextProperty(CL_CONTEXT_D3D11_DEVICE_KHR);
AVD3D11VADeviceContext* d3d11_device_ctx = hw_get_d3d11_device_ctx(ctx);
if (d3d11_device_ctx && d3d11device_ocl && d3d11_device_ctx->device == d3d11device_ocl)
return AV_HWDEVICE_TYPE_D3D11VA;
#endif
return AV_HWDEVICE_TYPE_NONE;
}
static
void hw_init_opencl(AVBufferRef* ctx) {
if (!ctx)
return;
AVHWDeviceContext* hw_device_ctx = (AVHWDeviceContext*)ctx->data;
if (!hw_device_ctx)
return;
#ifdef HAVE_VA_INTEL
VADisplay va_display = hw_get_va_display(hw_device_ctx);
if (va_display) {
va_intel::ocl::initializeContextFromVA(va_display);
}
#endif
#ifdef HAVE_D3D11
AVD3D11VADeviceContext* d3d11_device_ctx = hw_get_d3d11_device_ctx(hw_device_ctx);
if (d3d11_device_ctx) {
directx::ocl::initializeContextFromD3D11Device(d3d11_device_ctx->device);
}
#endif
if (hw_check_opencl_context(hw_device_ctx) != AV_HWDEVICE_TYPE_NONE) {
// Attach AVHWDeviceContext to OpenCL context
ocl::Context &ocl_context = ocl::OpenCLExecutionContext::getCurrent().getContext();
ocl_context.setUserContext(std::make_shared<OpenCL_FFMPEG_Context>(ctx));
}
}
static
AVBufferRef* hw_create_frames(struct AVCodecContext* ctx, AVBufferRef *hw_device_ctx, int width, int height, AVPixelFormat hw_format)
{
AVBufferRef *hw_frames_ref = nullptr;
AVBufferRef* hw_create_context_from_opencl(ocl::OpenCLExecutionContext& ocl_context, AVHWDeviceType hw_type) {
if (ocl_context.empty())
return NULL;
auto ocl_ffmpeg_context = ocl_context.getContext().getUserContext<OpenCL_FFMPEG_Context>();
if (!ocl_ffmpeg_context)
return NULL;
AVBufferRef* ctx = ocl_ffmpeg_context->GetAVHWDevice();
if (hw_type != ((AVHWDeviceContext*)ctx->data)->type) {
ctx = hw_create_derived_context(hw_type, ctx);
}
else {
ctx = av_buffer_ref(ctx);
}
if (ctx)
CV_LOG_INFO(NULL, "FFMPEG: Using " << av_hwdevice_get_type_name(hw_type) << " video acceleration context attached to OpenCL context");
return ctx;
}
#endif // HAVE_OPENCL
static
AVBufferRef* hw_create_device(AVHWDeviceType hw_type, int hw_device, const std::string& device_subname, bool use_opencl) {
AVBufferRef* hw_device_ctx = NULL;
if (AV_HWDEVICE_TYPE_NONE == hw_type)
return NULL;
#ifdef HAVE_OPENCL
// Check if OpenCL context has AVHWDeviceContext attached to it
ocl::OpenCLExecutionContext& ocl_context = ocl::OpenCLExecutionContext::getCurrentRef();
try {
hw_device_ctx = hw_create_context_from_opencl(ocl_context, hw_type);
if (hw_device_ctx) {
if (hw_device >= 0)
CV_LOG_ERROR(NULL, "VIDEOIO/FFMPEG: ignoring property HW_DEVICE as device context already created and attached to OpenCL context");
return hw_device_ctx;
}
}
catch (...) {
CV_LOG_INFO(NULL, "FFMPEG: Exception creating Video Acceleration context using current OpenCL context");
}
#endif
// Create new media context. In QSV case, first create 'child' context.
std::vector<AVHWDeviceType> child_types = { hw_type };
if (hw_type == AV_HWDEVICE_TYPE_QSV) {
#ifdef _WIN32
child_types = { AV_HWDEVICE_TYPE_D3D11VA, AV_HWDEVICE_TYPE_DXVA2 };
#else
child_types = { AV_HWDEVICE_TYPE_VAAPI };
#endif
}
for (AVHWDeviceType child_type : child_types) {
char device[128] = "";
char* pdevice = NULL;
if (hw_device >= 0 && hw_device < 100000) {
if (child_type == AV_HWDEVICE_TYPE_VAAPI) {
snprintf(device, sizeof(device), "/dev/dri/renderD%d", 128 + hw_device);
}
else {
snprintf(device, sizeof(device), "%d", hw_device);
}
pdevice = device;
}
const char* hw_child_name = av_hwdevice_get_type_name(child_type);
const char* device_name = pdevice ? pdevice : "'default'";
int err = av_hwdevice_ctx_create(&hw_device_ctx, child_type, pdevice, NULL, 0);
if (hw_device_ctx && err >= 0)
{
if (!hw_check_device(hw_device_ctx, hw_type, device_subname)) {
av_buffer_unref(&hw_device_ctx);
continue;
}
CV_LOG_INFO(NULL, "FFMPEG: Created video acceleration context (av_hwdevice_ctx_create) for " << hw_child_name << " on device " << device_name);
#ifdef HAVE_OPENCL
// if OpenCL context not created yet or property HW_ACCELERATION_USE_OPENCL set, create OpenCL context with binding to video acceleration context
if (ocl::haveOpenCL()) {
if (ocl_context.empty() || use_opencl) {
try {
hw_init_opencl(hw_device_ctx);
ocl_context = ocl::OpenCLExecutionContext::getCurrentRef();
if (!ocl_context.empty()) {
CV_LOG_INFO(NULL, "FFMPEG: Created OpenCL context with " << hw_child_name <<
" video acceleration on OpenCL device: " << ocl_context.getDevice().name());
}
} catch (...) {
CV_LOG_INFO(NULL, "FFMPEG: Exception creating OpenCL context with " << hw_child_name << " video acceleration");
}
}
else {
CV_LOG_INFO(NULL, "FFMPEG: Can't bind " << hw_child_name << " video acceleration context to already created OpenCL context");
}
}
#else
CV_UNUSED(use_opencl);
#endif
if (hw_type != child_type) {
AVBufferRef* derived_ctx = hw_create_derived_context(hw_type, hw_device_ctx);
av_buffer_unref(&hw_device_ctx);
return derived_ctx;
} else {
return hw_device_ctx;
}
}
else
{
const char* hw_name = hw_child_name;
CV_LOG_INFO(NULL, "FFMPEG: Failed to create " << hw_name << " video acceleration (av_hwdevice_ctx_create) on device " << device_name);
}
}
return NULL;
}
static
AVBufferRef* hw_create_frames(struct AVCodecContext* codec_ctx, AVBufferRef *hw_device_ctx, int width, int height, AVPixelFormat hw_format)
{
AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)hw_device_ctx->data;
AVBufferRef* child_ctx = hw_device_ctx;
// In QSV case we first allocate child D3D11/VAAPI frames (except DXVA2 as no OpenCL interop), then derive to parent QSV frames
if (AV_HWDEVICE_TYPE_QSV == device_ctx->type) {
AVBufferRef *ctx = (AVBufferRef *) device_ctx->user_opaque; // child context stored during creation of derived context
if (ctx && AV_HWDEVICE_TYPE_DXVA2 != ((AVHWDeviceContext *) ctx->data)->type) {
child_ctx = ctx;
}
}
AVBufferRef *hw_frames_ref = nullptr;
if (codec_ctx)
{
int res = avcodec_get_hw_frames_parameters(ctx, hw_device_ctx, hw_format, &hw_frames_ref);
int res = avcodec_get_hw_frames_parameters(codec_ctx, child_ctx, hw_format, &hw_frames_ref);
if (res < 0)
{
CV_LOG_DEBUG(NULL, "FFMPEG: avcodec_get_hw_frames_parameters() call failed: " << res)
@@ -335,7 +622,7 @@ AVBufferRef* hw_create_frames(struct AVCodecContext* ctx, AVBufferRef *hw_device
}
if (!hw_frames_ref)
{
hw_frames_ref = av_hwframe_ctx_alloc(hw_device_ctx);
hw_frames_ref = av_hwframe_ctx_alloc(child_ctx);
}
if (!hw_frames_ref)
{
@@ -345,12 +632,41 @@ AVBufferRef* hw_create_frames(struct AVCodecContext* ctx, AVBufferRef *hw_device
AVHWFramesContext *frames_ctx = (AVHWFramesContext *)(hw_frames_ref->data);
frames_ctx->width = width;
frames_ctx->height = height;
if (frames_ctx->format == AV_PIX_FMT_NONE)
frames_ctx->format = hw_format;
if (frames_ctx->format == AV_PIX_FMT_NONE) {
if (child_ctx == hw_device_ctx) {
frames_ctx->format = hw_format;
}
else {
AVHWFramesConstraints* constraints = av_hwdevice_get_hwframe_constraints(child_ctx, NULL);
if (constraints) {
frames_ctx->format = constraints->valid_hw_formats[0];
av_hwframe_constraints_free(&constraints);
}
}
}
if (frames_ctx->sw_format == AV_PIX_FMT_NONE)
frames_ctx->sw_format = HW_DEFAULT_SW_FORMAT;
if (frames_ctx->initial_pool_size == 0)
frames_ctx->initial_pool_size = HW_DEFAULT_POOL_SIZE;
#ifdef HAVE_D3D11
if (frames_ctx->device_ctx && AV_HWDEVICE_TYPE_D3D11VA == frames_ctx->device_ctx->type) {
// BindFlags
AVD3D11VAFramesContext* frames_hwctx = (AVD3D11VAFramesContext*)frames_ctx->hwctx;
frames_hwctx->BindFlags |= D3D11_BIND_DECODER | D3D11_BIND_VIDEO_ENCODER;
// See function hw_get_d3d11_single_texture(), it allocates additional ID3D11Texture2D texture and
// attaches it as 'user_opaque' field. We have to set free() callback before av_hwframe_ctx_init() call.
struct D3D11SingleTexture {
static void free(struct AVHWFramesContext* ctx) {
ID3D11Texture2D* singleTexture = (ID3D11Texture2D*)ctx->user_opaque;
if (ctx->user_opaque)
singleTexture->Release();
}
};
frames_ctx->free = D3D11SingleTexture::free;
}
#endif
int res = av_hwframe_ctx_init(hw_frames_ref);
if (res < 0)
{
@@ -358,7 +674,25 @@ AVBufferRef* hw_create_frames(struct AVCodecContext* ctx, AVBufferRef *hw_device
av_buffer_unref(&hw_frames_ref);
return NULL;
}
return hw_frames_ref;
if (child_ctx != hw_device_ctx) {
AVBufferRef* derived_frame_ctx = NULL;
int flags = AV_HWFRAME_MAP_READ | AV_HWFRAME_MAP_WRITE;
res = av_hwframe_ctx_create_derived(&derived_frame_ctx, hw_format, hw_device_ctx, hw_frames_ref, flags);
av_buffer_unref(&hw_frames_ref);
if (res < 0)
{
CV_LOG_INFO(NULL, "FFMPEG: Failed to create derived HW frame context (av_hwframe_ctx_create_derived): " << res);
return NULL;
}
else {
((AVHWFramesContext*)derived_frame_ctx->data)->user_opaque = frames_ctx;
return derived_frame_ctx;
}
}
else {
return hw_frames_ref;
}
}
static
@@ -455,6 +789,110 @@ AVPixelFormat hw_get_format_callback(struct AVCodecContext *ctx, const enum AVPi
return fmt[0];
}
// GPU color conversion NV12->BGRA via OpenCL extensions
static bool
hw_copy_frame_to_umat(AVBufferRef* ctx, AVFrame* hw_frame, cv::OutputArray output) {
CV_UNUSED(hw_frame);
CV_UNUSED(output);
if (!ctx)
return false;
#ifdef HAVE_OPENCL
try {
// check that current OpenCL context initilized with binding to same VAAPI/D3D11 context
AVHWDeviceContext *hw_device_ctx = (AVHWDeviceContext *) ctx->data;
AVHWDeviceType child_type = hw_check_opencl_context(hw_device_ctx);
if (child_type == AV_HWDEVICE_TYPE_NONE)
return false;
#ifdef HAVE_VA_INTEL
if (child_type == AV_HWDEVICE_TYPE_VAAPI) {
VADisplay va_display = hw_get_va_display(hw_device_ctx);
VASurfaceID va_surface = hw_get_va_surface(hw_frame);
if (va_display && va_surface != VA_INVALID_SURFACE) {
va_intel::convertFromVASurface(va_display, va_surface, {hw_frame->width, hw_frame->height}, output);
return true;
}
}
#endif
#ifdef HAVE_D3D11
if (child_type == AV_HWDEVICE_TYPE_D3D11VA) {
AVD3D11VADeviceContext* d3d11_device_ctx = hw_get_d3d11_device_ctx(hw_device_ctx);
int subresource = 0;
ID3D11Texture2D* texture = hw_get_d3d11_texture(hw_frame, &subresource);
ID3D11Texture2D* singleTexture = hw_get_d3d11_single_texture(hw_frame, d3d11_device_ctx, texture);
if (texture && singleTexture) {
// Copy D3D11 sub-texture to D3D11 single texture
d3d11_device_ctx->device_context->CopySubresourceRegion(singleTexture, 0, 0, 0, 0, texture, subresource, NULL);
// Copy D3D11 single texture to cv::UMat
directx::convertFromD3D11Texture2D(singleTexture, output);
return true;
}
}
#endif
}
catch (...)
{
return false;
}
#endif // HAVE_OPENCL
return false;
}
// GPU color conversion BGRA->NV12 via OpenCL extensions
static bool
hw_copy_umat_to_frame(AVBufferRef* ctx, cv::InputArray input, AVFrame* hw_frame) {
CV_UNUSED(input);
CV_UNUSED(hw_frame);
if (!ctx)
return false;
#ifdef HAVE_OPENCL
try {
// check that current OpenCL context initilized with binding to same VAAPI/D3D11 context
AVHWDeviceContext *hw_device_ctx = (AVHWDeviceContext *) ctx->data;
AVHWDeviceType child_type = hw_check_opencl_context(hw_device_ctx);
if (child_type == AV_HWDEVICE_TYPE_NONE)
return false;
#ifdef HAVE_VA_INTEL
if (child_type == AV_HWDEVICE_TYPE_VAAPI) {
VADisplay va_display = hw_get_va_display(hw_device_ctx);
VASurfaceID va_surface = hw_get_va_surface(hw_frame);
if (va_display != NULL && va_surface != VA_INVALID_SURFACE) {
va_intel::convertToVASurface(va_display, input, va_surface, {hw_frame->width, hw_frame->height});
return true;
}
}
#endif
#ifdef HAVE_D3D11
if (child_type == AV_HWDEVICE_TYPE_D3D11VA) {
AVD3D11VADeviceContext* d3d11_device_ctx = hw_get_d3d11_device_ctx(hw_device_ctx);
int subresource = 0;
ID3D11Texture2D* texture = hw_get_d3d11_texture(hw_frame, &subresource);
ID3D11Texture2D* singleTexture = hw_get_d3d11_single_texture(hw_frame, d3d11_device_ctx, texture);
if (texture && singleTexture) {
// Copy cv::UMat to D3D11 single texture
directx::convertToD3D11Texture2D(input, singleTexture);
// Copy D3D11 single texture to D3D11 sub-texture
d3d11_device_ctx->device_context->CopySubresourceRegion(texture, subresource, 0, 0, 0, singleTexture, 0, NULL);
return true;
}
}
#endif
}
catch (...)
{
return false;
}
#endif // HAVE_OPENCL
return false;
}
static
VideoAccelerationType hw_type_to_va_type(AVHWDeviceType hw_type) {
struct HWTypeFFMPEG {
+71 -2
View File
@@ -476,6 +476,7 @@ struct CvCapture_FFMPEG
bool setProperty(int, double);
bool grabFrame();
bool retrieveFrame(int, unsigned char** data, int* step, int* width, int* height, int* cn);
bool retrieveHWFrame(cv::OutputArray output);
void rotateFrame(cv::Mat &mat) const;
void init();
@@ -537,6 +538,7 @@ struct CvCapture_FFMPEG
#endif
VideoAccelerationType va_type;
int hw_device;
int use_opencl;
};
void CvCapture_FFMPEG::init()
@@ -574,6 +576,7 @@ void CvCapture_FFMPEG::init()
bsfc = NULL;
va_type = cv::VIDEO_ACCELERATION_NONE; // TODO OpenCV 5.0: change to _ANY?
hw_device = -1;
use_opencl = 0;
}
@@ -922,6 +925,9 @@ bool CvCapture_FFMPEG::open(const char* _filename, const VideoCaptureParameters&
return false;
}
}
if (params.has(CAP_PROP_HW_ACCELERATION_USE_OPENCL)) {
use_opencl = params.get<int>(CAP_PROP_HW_ACCELERATION_USE_OPENCL);
}
if (params.warnUnusedParameters())
{
CV_LOG_ERROR(NULL, "VIDEOIO/FFMPEG: unsupported parameters in .open(), see logger INFO channel for details. Bailout");
@@ -1051,7 +1057,7 @@ bool CvCapture_FFMPEG::open(const char* _filename, const VideoCaptureParameters&
if (codec) {
if (hw_pix_fmt != AV_PIX_FMT_NONE)
enc->get_format = hw_get_format_callback; // set callback to select HW pixel format, not SW format
enc->hw_device_ctx = hw_create_device(hw_type, hw_device, accel_iter.device_subname());
enc->hw_device_ctx = hw_create_device(hw_type, hw_device, accel_iter.device_subname(), use_opencl != 0);
if (!enc->hw_device_ctx)
{
CV_LOG_DEBUG(NULL, "FFMPEG: ... can't create H/W device: '" << accel_iter.hw_type_device_string() << "'");
@@ -1476,6 +1482,22 @@ bool CvCapture_FFMPEG::retrieveFrame(int, unsigned char** data, int* step, int*
return true;
}
bool CvCapture_FFMPEG::retrieveHWFrame(cv::OutputArray output)
{
#if USE_AV_HW_CODECS
// check that we have HW frame in GPU memory
if (!picture || !picture->hw_frames_ctx) {
return false;
}
// GPU color conversion NV12->BGRA, from GPU media buffer to GPU OpenCL buffer
return hw_copy_frame_to_umat(video_st->codec->hw_device_ctx, picture, output);
#else
CV_UNUSED(output);
return false;
#endif
}
double CvCapture_FFMPEG::getProperty( int property_id ) const
{
if( !video_st ) return 0;
@@ -1549,6 +1571,8 @@ double CvCapture_FFMPEG::getProperty( int property_id ) const
return static_cast<double>(va_type);
case CAP_PROP_HW_DEVICE:
return static_cast<double>(hw_device);
case CAP_PROP_HW_ACCELERATION_USE_OPENCL:
return static_cast<double>(use_opencl);
#endif // USE_AV_HW_CODECS
default:
break;
@@ -1752,6 +1776,7 @@ struct CvVideoWriter_FFMPEG
double fps, int width, int height, const VideoWriterParameters& params );
void close();
bool writeFrame( const unsigned char* data, int step, int width, int height, int cn, int origin );
bool writeHWFrame(cv::InputArray input);
double getProperty(int propId) const;
void init();
@@ -1774,6 +1799,7 @@ struct CvVideoWriter_FFMPEG
struct SwsContext *img_convert_ctx;
VideoAccelerationType va_type;
int hw_device;
int use_opencl;
};
static const char * icvFFMPEGErrStr(int err)
@@ -1836,6 +1862,7 @@ void CvVideoWriter_FFMPEG::init()
frame_idx = 0;
va_type = VIDEO_ACCELERATION_NONE;
hw_device = -1;
use_opencl = 0;
ok = false;
}
@@ -2210,6 +2237,41 @@ bool CvVideoWriter_FFMPEG::writeFrame( const unsigned char* data, int step, int
return ret;
}
bool CvVideoWriter_FFMPEG::writeHWFrame(cv::InputArray input) {
#if USE_AV_HW_CODECS
if (!video_st->codec->hw_frames_ctx)
return false;
// Get hardware frame from frame pool
AVFrame* hw_frame = av_frame_alloc();
if (!hw_frame) {
return false;
}
if (av_hwframe_get_buffer(video_st->codec->hw_frames_ctx, hw_frame, 0) < 0) {
av_frame_free(&hw_frame);
return false;
}
// GPU to GPU copy
if (!hw_copy_umat_to_frame(video_st->codec->hw_device_ctx, input, hw_frame)) {
av_frame_free(&hw_frame);
return false;
}
// encode
hw_frame->pts = frame_idx;
icv_av_write_frame_FFMPEG( oc, video_st, outbuf, outbuf_size, hw_frame, frame_idx);
frame_idx++;
av_frame_free(&hw_frame);
return true;
#else
CV_UNUSED(input);
return false;
#endif
}
double CvVideoWriter_FFMPEG::getProperty(int propId) const
{
CV_UNUSED(propId);
@@ -2222,6 +2284,10 @@ double CvVideoWriter_FFMPEG::getProperty(int propId) const
{
return static_cast<double>(hw_device);
}
else if (propId == VIDEOWRITER_PROP_HW_ACCELERATION_USE_OPENCL)
{
return static_cast<double>(use_opencl);
}
#endif
return 0;
}
@@ -2375,6 +2441,9 @@ bool CvVideoWriter_FFMPEG::open( const char * filename, int fourcc,
return false;
}
}
if (params.has(VIDEOWRITER_PROP_HW_ACCELERATION_USE_OPENCL)) {
use_opencl = params.get<int>(VIDEOWRITER_PROP_HW_ACCELERATION_USE_OPENCL);
}
if (params.warnUnusedParameters())
{
@@ -2638,7 +2707,7 @@ bool CvVideoWriter_FFMPEG::open( const char * filename, int fourcc,
if (!codec)
continue;
hw_device_ctx = hw_create_device(hw_type, hw_device, accel_iter.device_subname());
hw_device_ctx = hw_create_device(hw_type, hw_device, accel_iter.device_subname(), use_opencl != 0);
if (!hw_device_ctx)
continue;
}