| Message ID | 20180414155339.22839-5-sw@jkqxz.net |
|---|---|
| State | New |
| Headers | show |
> -----Original Message----- > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of Mark Thompson > Sent: Saturday, April 14, 2018 6:54 PM > To: ffmpeg-devel@ffmpeg.org > Subject: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious initialisations > Hi Mark, I reviewed and tested all 5 patches in set all together. Now the code looks cleaner. My local tests passed for dx9 and dx11 with and without -hwaccel_output_format. I am waiting patches to be applied to propose new patch with hwcontext_amf in libavutil. Thanks, Alexander
On 15/04/18 00:17, Alexander Kravchenko wrote: >> -----Original Message----- >> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of Mark Thompson >> Sent: Saturday, April 14, 2018 6:54 PM >> To: ffmpeg-devel@ffmpeg.org >> Subject: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious initialisations >> > > Hi Mark, > I reviewed and tested all 5 patches in set all together. > Now the code looks cleaner. > My local tests passed for dx9 and dx11 with and without -hwaccel_output_format. Ok, applied. Thank you! > I am waiting patches to be applied to propose new patch with hwcontext_amf in libavutil. Can you explain what you're intending to use that for? It's not clear to me how an extra wrapper around the D3D(9|11) surfaces is going to help, given that the support for them with AMF is already pretty good. (Compare the Intel libmfx stuff (the misleadingly-named "qsv") where the extra wrapping does help for some cases because the underlying library has weird constraints, but overall adds a lot of complexity (and failure modes) for rather unclear benefit. It's also inconvenient in that it promotes the existence of antifeatures like the "_qsv" decoders which are inferior to the builtin hwaccels in pretty much every respect.) - Mark
> -----Original Message----- > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of Mark Thompson > Sent: Sunday, April 15, 2018 7:31 PM > To: ffmpeg-devel@ffmpeg.org > Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious initialisations > > > > I am waiting patches to be applied to propose new patch with hwcontext_amf in libavutil. > > Can you explain what you're intending to use that for? It's not clear to me how an extra wrapper around the D3D(9|11) surfaces is > going to help, given that the support for them with AMF is already pretty good. (Compare the Intel libmfx stuff (the misleadingly- > named "qsv") where the extra wrapping does help for some cases because the underlying library has weird constraints, but overall > adds a lot of complexity (and failure modes) for rather unclear benefit. It's also inconvenient in that it promotes the existence of > antifeatures like the "_qsv" decoders which are inferior to the builtin hwaccels in pretty much every respect.) > Hi Mark, I am intending to create amf common helpers(tools) in libavutil. They will be used in amfenc and vf_scaleamf. (vf_scaleamf is hw-accelerated color-space converter and scaler) amf helpers should provide: * amf_library: functions to load/unload amf dll based on reference count mechanism * amf_context: functions to create AMFContext derived from DXVA2, D3D11, opencl and Vulcan in future * av_frame <-> AMFSurface map functions (move from amfenc.c) amfav_context can be implemented like hwdevice_ctx (AVAMFDeviceContext) and can be managed by av_hwdevice_ctx_create_derived (in case of incoming hw frames) and av_hwdevice_ctx_create or it can be implemented not using of av_hwdevice_ctx* mechanism I think don’t need AVAMFFrameContext, and amf components will send/receive hwframes based on existing framecontexts (dxva/opencl...) Could you recommend the best way how to do this fit in current architecture? Thanks, Alexander
On 15/04/18 20:45, Alexander Kravchenko wrote: >> -----Original Message----- >> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of Mark Thompson >> Sent: Sunday, April 15, 2018 7:31 PM >> To: ffmpeg-devel@ffmpeg.org >> Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious initialisations >> >> >>> I am waiting patches to be applied to propose new patch with hwcontext_amf in libavutil. >> >> Can you explain what you're intending to use that for? It's not clear to me how an extra wrapper around the D3D(9|11) surfaces is >> going to help, given that the support for them with AMF is already pretty good. (Compare the Intel libmfx stuff (the misleadingly- >> named "qsv") where the extra wrapping does help for some cases because the underlying library has weird constraints, but overall >> adds a lot of complexity (and failure modes) for rather unclear benefit. It's also inconvenient in that it promotes the existence of >> antifeatures like the "_qsv" decoders which are inferior to the builtin hwaccels in pretty much every respect.) >> > > Hi Mark, > I am intending to create amf common helpers(tools) in libavutil. > They will be used in amfenc and vf_scaleamf. (vf_scaleamf is hw-accelerated color-space converter and scaler) > > amf helpers should provide: > * amf_library: functions to load/unload amf dll based on reference count mechanism > * amf_context: functions to create AMFContext derived from DXVA2, D3D11, opencl and Vulcan in future > * av_frame <-> AMFSurface map functions (move from amfenc.c) > > amfav_context can be implemented like hwdevice_ctx (AVAMFDeviceContext) and can be managed by av_hwdevice_ctx_create_derived (in case of incoming hw frames) and av_hwdevice_ctx_create or it can be implemented not using of av_hwdevice_ctx* mechanism > > I think don’t need AVAMFFrameContext, and amf components will send/receive hwframes based on existing framecontexts (dxva/opencl...) > > Could you recommend the best way how to do this fit in current architecture? I agree that using a hwdevice context here makes sense, since it wraps all of the right properties (in particular, derivation from other devices). It's less clear to me what to do with the frames. A hwframes context could work just for derivation because you don't actually need to implement the allocation stuff (the existing DRM hwcontext does this, since it's only for interop). What other approach would you think of taking? Adding special external API to use internally between libraries is not nice and we try to avoid it quite strongly. - Mark
> -----Original Message----- > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of Mark Thompson > Sent: Sunday, April 22, 2018 6:37 PM > To: ffmpeg-devel@ffmpeg.org > Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious initialisations > > On 15/04/18 20:45, Alexander Kravchenko wrote: > >> -----Original Message----- > >> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf > >> Of Mark Thompson > >> Sent: Sunday, April 15, 2018 7:31 PM > >> To: ffmpeg-devel@ffmpeg.org > >> Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious > >> initialisations > >> > >> > >>> I am waiting patches to be applied to propose new patch with hwcontext_amf in libavutil. > >> > >> Can you explain what you're intending to use that for? It's not > >> clear to me how an extra wrapper around the D3D(9|11) surfaces is > >> going to help, given that the support for them with AMF is already > >> pretty good. (Compare the Intel libmfx stuff (the misleadingly- > >> named "qsv") where the extra wrapping does help for some cases > >> because the underlying library has weird constraints, but overall > >> adds a lot of complexity (and failure modes) for rather unclear > >> benefit. It's also inconvenient in that it promotes the existence of > >> antifeatures like the "_qsv" decoders which are inferior to the > >> builtin hwaccels in pretty much every respect.) > >> > > > > Hi Mark, > > I am intending to create amf common helpers(tools) in libavutil. > > They will be used in amfenc and vf_scaleamf. (vf_scaleamf is > > hw-accelerated color-space converter and scaler) > > > > amf helpers should provide: > > * amf_library: functions to load/unload amf dll based on reference > > count mechanism > > * amf_context: functions to create AMFContext derived from DXVA2, > > D3D11, opencl and Vulcan in future > > * av_frame <-> AMFSurface map functions (move from amfenc.c) > > > > amfav_context can be implemented like hwdevice_ctx > > (AVAMFDeviceContext) and can be managed by > > av_hwdevice_ctx_create_derived (in case of incoming hw frames) and > > av_hwdevice_ctx_create or it can be implemented not using of > > av_hwdevice_ctx* mechanism > > > > I think don’t need AVAMFFrameContext, and amf components will > > send/receive hwframes based on existing framecontexts (dxva/opencl...) > > > > Could you recommend the best way how to do this fit in current architecture? > > I agree that using a hwdevice context here makes sense, since it wraps all of the right properties (in particular, derivation from other > devices). > > It's less clear to me what to do with the frames. A hwframes context could work just for derivation because you don't actually need to > implement the allocation stuff (the existing DRM hwcontext does this, since it's only for interop). What other approach would you > think of taking? Adding special external API to use internally between libraries is not nice and we try to avoid it quite strongly. > Hi Mark, I agree it is good to stay within current API (hwdevice and hwframes), but I am not sure it is always possible. is it ok create hwcontext_amf_internal.h which can be placeholder for the API like the following, or probably some helper functions can be published as pointers in AVAMFDeviceContext structure: amf_set_property_buffer amf_get_property_buffer amf_create_buffer_with_frame_ref amf_release_buffer_with_frame_ref Thanks, Alexander
On 24 April 2018 at 12:29, Alexander Kravchenko <akravchenko188@gmail.com> wrote: > > > > -----Original Message----- > > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf > Of Mark Thompson > > Sent: Sunday, April 22, 2018 6:37 PM > > To: ffmpeg-devel@ffmpeg.org > > Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious > initialisations > > > > On 15/04/18 20:45, Alexander Kravchenko wrote: > > >> -----Original Message----- > > >> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf > > >> Of Mark Thompson > > >> Sent: Sunday, April 15, 2018 7:31 PM > > >> To: ffmpeg-devel@ffmpeg.org > > >> Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious > > >> initialisations > > >> > > >> > > >>> I am waiting patches to be applied to propose new patch with > hwcontext_amf in libavutil. > > >> > > >> Can you explain what you're intending to use that for? It's not > > >> clear to me how an extra wrapper around the D3D(9|11) surfaces is > > >> going to help, given that the support for them with AMF is already > > >> pretty good. (Compare the Intel libmfx stuff (the misleadingly- > > >> named "qsv") where the extra wrapping does help for some cases > > >> because the underlying library has weird constraints, but overall > > >> adds a lot of complexity (and failure modes) for rather unclear > > >> benefit. It's also inconvenient in that it promotes the existence of > > >> antifeatures like the "_qsv" decoders which are inferior to the > > >> builtin hwaccels in pretty much every respect.) > > >> > > > > > > Hi Mark, > > > I am intending to create amf common helpers(tools) in libavutil. > > > They will be used in amfenc and vf_scaleamf. (vf_scaleamf is > > > hw-accelerated color-space converter and scaler) > > > > > > amf helpers should provide: > > > * amf_library: functions to load/unload amf dll based on reference > > > count mechanism > > > * amf_context: functions to create AMFContext derived from DXVA2, > > > D3D11, opencl and Vulcan in future > > > * av_frame <-> AMFSurface map functions (move from amfenc.c) > > > > > > amfav_context can be implemented like hwdevice_ctx > > > (AVAMFDeviceContext) and can be managed by > > > av_hwdevice_ctx_create_derived (in case of incoming hw frames) and > > > av_hwdevice_ctx_create or it can be implemented not using of > > > av_hwdevice_ctx* mechanism > > > > > > I think don’t need AVAMFFrameContext, and amf components will > > > send/receive hwframes based on existing framecontexts (dxva/opencl...) > > > > > > Could you recommend the best way how to do this fit in current > architecture? > > > > I agree that using a hwdevice context here makes sense, since it wraps > all of the right properties (in particular, derivation from other > > devices). > > > > It's less clear to me what to do with the frames. A hwframes context > could work just for derivation because you don't actually need to > > implement the allocation stuff (the existing DRM hwcontext does this, > since it's only for interop). What other approach would you > > think of taking? Adding special external API to use internally between > libraries is not nice and we try to avoid it quite strongly. > > > > Hi Mark, > I agree it is good to stay within current API (hwdevice and hwframes), but > I am not sure it is always possible. > > is it ok create hwcontext_amf_internal.h which can be placeholder for the > API like the following, or probably some helper functions can be published > as pointers in AVAMFDeviceContext structure: > amf_set_property_buffer > amf_get_property_buffer > amf_create_buffer_with_frame_ref > amf_release_buffer_with_frame_ref > No, I think there should be no avpriv functions for hwcontexts, nor any public API changes. Better duplicate that code in lavfi.
> -----Original Message----- > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of Rostislav Pehlivanov > Sent: Tuesday, April 24, 2018 4:04 PM > To: FFmpeg development discussions and patches <ffmpeg-devel@ffmpeg.org> > Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious initialisations > > On 24 April 2018 at 12:29, Alexander Kravchenko <akravchenko188@gmail.com> > wrote: > > > > > > > > -----Original Message----- > > > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On > > > Behalf > > Of Mark Thompson > > > Sent: Sunday, April 22, 2018 6:37 PM > > > To: ffmpeg-devel@ffmpeg.org > > > Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious > > initialisations > > > > > > On 15/04/18 20:45, Alexander Kravchenko wrote: > > > >> -----Original Message----- > > > >> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On > > > >> Behalf Of Mark Thompson > > > >> Sent: Sunday, April 15, 2018 7:31 PM > > > >> To: ffmpeg-devel@ffmpeg.org > > > >> Subject: Re: [FFmpeg-devel] [PATCH 5/5] amfenc: Remove spurious > > > >> initialisations > > > >> > > > >> > > > >>> I am waiting patches to be applied to propose new patch with > > hwcontext_amf in libavutil. > > > >> > > > >> Can you explain what you're intending to use that for? It's not > > > >> clear to me how an extra wrapper around the D3D(9|11) surfaces is > > > >> going to help, given that the support for them with AMF is > > > >> already pretty good. (Compare the Intel libmfx stuff (the > > > >> misleadingly- named "qsv") where the extra wrapping does help for > > > >> some cases because the underlying library has weird constraints, > > > >> but overall adds a lot of complexity (and failure modes) for > > > >> rather unclear benefit. It's also inconvenient in that it > > > >> promotes the existence of antifeatures like the "_qsv" decoders > > > >> which are inferior to the builtin hwaccels in pretty much every > > > >> respect.) > > > >> > > > > > > > > Hi Mark, > > > > I am intending to create amf common helpers(tools) in libavutil. > > > > They will be used in amfenc and vf_scaleamf. (vf_scaleamf is > > > > hw-accelerated color-space converter and scaler) > > > > > > > > amf helpers should provide: > > > > * amf_library: functions to load/unload amf dll based on reference > > > > count mechanism > > > > * amf_context: functions to create AMFContext derived from DXVA2, > > > > D3D11, opencl and Vulcan in future > > > > * av_frame <-> AMFSurface map functions (move from amfenc.c) > > > > > > > > amfav_context can be implemented like hwdevice_ctx > > > > (AVAMFDeviceContext) and can be managed by > > > > av_hwdevice_ctx_create_derived (in case of incoming hw frames) and > > > > av_hwdevice_ctx_create or it can be implemented not using of > > > > av_hwdevice_ctx* mechanism > > > > > > > > I think don’t need AVAMFFrameContext, and amf components will > > > > send/receive hwframes based on existing framecontexts > > > > (dxva/opencl...) > > > > > > > > Could you recommend the best way how to do this fit in current > > architecture? > > > > > > I agree that using a hwdevice context here makes sense, since it > > > wraps > > all of the right properties (in particular, derivation from other > > > devices). > > > > > > It's less clear to me what to do with the frames. A hwframes > > > context > > could work just for derivation because you don't actually need to > > > implement the allocation stuff (the existing DRM hwcontext does > > > this, > > since it's only for interop). What other approach would you > > > think of taking? Adding special external API to use internally > > > between > > libraries is not nice and we try to avoid it quite strongly. > > > > > > > Hi Mark, > > I agree it is good to stay within current API (hwdevice and hwframes), > > but I am not sure it is always possible. > > > > is it ok create hwcontext_amf_internal.h which can be placeholder for > > the API like the following, or probably some helper functions can be > > published as pointers in AVAMFDeviceContext structure: > > amf_set_property_buffer > > amf_get_property_buffer > > amf_create_buffer_with_frame_ref > > amf_release_buffer_with_frame_ref > > > > No, I think there should be no avpriv functions for hwcontexts, nor any public API changes. Better duplicate that code in lavfi. Ok, if avutil is not proper place to put shared code between libavcodec and libavfilter, may be there should be another place? Duplicating code will increase cost of maintenance, which is not good. Probably there should be some other place to put shared code? May be shared code can be just set of static functions and macroses are put in headers? Thanks, Alexander
diff --git a/libavcodec/amfenc.c b/libavcodec/amfenc.c index e641048532..93fcee9480 100644 --- a/libavcodec/amfenc.c +++ b/libavcodec/amfenc.c @@ -107,16 +107,11 @@ static AMFTraceWriterVtbl tracer_vtbl = static int amf_load_library(AVCodecContext *avctx) { - AmfContext *ctx = avctx->priv_data; - AMFInit_Fn init_fun = NULL; - AMFQueryVersion_Fn version_fun = NULL; - AMF_RESULT res = AMF_OK; + AmfContext *ctx = avctx->priv_data; + AMFInit_Fn init_fun; + AMFQueryVersion_Fn version_fun; + AMF_RESULT res; - ctx->eof = 0; - ctx->delayed_drain = 0; - ctx->hw_frames_ctx = NULL; - ctx->hw_device_ctx = NULL; - ctx->delayed_surface = NULL; ctx->delayed_frame = av_frame_alloc(); if (!ctx->delayed_frame) { return AVERROR(ENOMEM); @@ -326,10 +321,10 @@ static int amf_init_context(AVCodecContext *avctx) static int amf_init_encoder(AVCodecContext *avctx) { - AmfContext *ctx = avctx->priv_data; - const wchar_t *codec_id = NULL; - AMF_RESULT res = AMF_OK; - enum AVPixelFormat pix_fmt; + AmfContext *ctx = avctx->priv_data; + const wchar_t *codec_id = NULL; + AMF_RESULT res; + enum AVPixelFormat pix_fmt; switch (avctx->codec->id) { case AV_CODEC_ID_H264: @@ -360,9 +355,9 @@ static int amf_init_encoder(AVCodecContext *avctx) int av_cold ff_amf_encode_close(AVCodecContext *avctx) { - AmfContext *ctx = avctx->priv_data; - if (ctx->delayed_surface) - { + AmfContext *ctx = avctx->priv_data; + + if (ctx->delayed_surface) { ctx->delayed_surface->pVtbl->Release(ctx->delayed_surface); ctx->delayed_surface = NULL; } @@ -402,11 +397,11 @@ int av_cold ff_amf_encode_close(AVCodecContext *avctx) static int amf_copy_surface(AVCodecContext *avctx, const AVFrame *frame, AMFSurface* surface) { - AMFPlane *plane = NULL; - uint8_t *dst_data[4]; - int dst_linesize[4]; - int planes; - int i; + AMFPlane *plane; + uint8_t *dst_data[4]; + int dst_linesize[4]; + int planes; + int i; planes = surface->pVtbl->GetPlanesCount(surface); av_assert0(planes < FF_ARRAY_ELEMS(dst_data)); @@ -437,11 +432,11 @@ static inline int timestamp_queue_enqueue(AVCodecContext *avctx, int64_t timesta static int amf_copy_buffer(AVCodecContext *avctx, AVPacket *pkt, AMFBuffer *buffer) { - AmfContext *ctx = avctx->priv_data; - int ret; - AMFVariantStruct var = {0}; - int64_t timestamp = AV_NOPTS_VALUE; - int64_t size = buffer->pVtbl->GetSize(buffer); + AmfContext *ctx = avctx->priv_data; + int ret; + AMFVariantStruct var = {0}; + int64_t timestamp = AV_NOPTS_VALUE; + int64_t size = buffer->pVtbl->GetSize(buffer); if ((ret = ff_alloc_packet2(avctx, pkt, size, 0)) < 0) { return ret; @@ -497,20 +492,7 @@ static int amf_copy_buffer(AVCodecContext *avctx, AVPacket *pkt, AMFBuffer *buff // amfenc API implementation int ff_amf_encode_init(AVCodecContext *avctx) { - AmfContext *ctx = avctx->priv_data; - int ret; - - ctx->factory = NULL; - ctx->debug = NULL; - ctx->trace = NULL; - ctx->context = NULL; - ctx->encoder = NULL; - ctx->library = NULL; - ctx->version = 0; - ctx->eof = 0; - ctx->format = 0; - ctx->tracer.vtbl = NULL; - ctx->tracer.avctx = NULL; + int ret; if ((ret = amf_load_library(avctx)) == 0) { if ((ret = amf_init_context(avctx)) == 0) { @@ -595,10 +577,10 @@ static void amf_release_buffer_with_frame_ref(AMFBuffer *frame_ref_storage_buffe int ff_amf_send_frame(AVCodecContext *avctx, const AVFrame *frame) { - AMF_RESULT res = AMF_OK; - AmfContext *ctx = avctx->priv_data; - AMFSurface *surface = NULL; - int ret; + AmfContext *ctx = avctx->priv_data; + AMFSurface *surface; + AMF_RESULT res; + int ret; if (!ctx->encoder) return AVERROR(EINVAL);