May 2022 - ffmpeg-cvslog - lists.ffmpeg.org

avutil/hwcontext_qsv: fix mapping issue between QSV frames and D3D11VA frames
by Tong Wu 23 May '22

23 May '22

ffmpeg | branch: master | Tong Wu <tong1.wu-at-intel.com(a)ffmpeg.org> | Sat May 7 06:24:41 2022 +0000| [1f9b5fa5815c0ea22d830da6cc4cbfcb6098def4] | committer: Haihao Xiang avutil/hwcontext_qsv: fix mapping issue between QSV frames and D3D11VA frames Fixes: $ ffmpeg.exe -init_hw_device d3d11va=d3d11 -init_hw_device \ qsv=qsv@d3d11 -s:v WxH -pix_fmt nv12 -i input.yuv -vf \ "hwupload=extra_hw_frames=16,hwmap=derive_device=d3d11va,format=d3d11,\ hwmap=derive_device=qsv,format=qsv" -f null - Reviewed-by: Soft Works <softworkz(a)hotmail.com> Signed-off-by: Tong Wu <tong1.wu(a)intel.com> Signed-off-by: Haihao Xiang <haihao.xiang(a)intel.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=1f9b5fa5815c0ea22… --- libavutil/hwcontext_qsv.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/libavutil/hwcontext_qsv.c b/libavutil/hwcontext_qsv.c index 6e4ba9d5d3..f0616f20b9 100644 --- a/libavutil/hwcontext_qsv.c +++ b/libavutil/hwcontext_qsv.c @@ -914,7 +914,7 @@ static int qsv_map_from(AVHWFramesContext *ctx, if (child_frames_ctx->device_ctx->type == AV_HWDEVICE_TYPE_D3D11VA) { mfxHDLPair *pair = (mfxHDLPair*)surf->Data.MemId; dst->data[0] = pair->first; - dst->data[1] = pair->second; + dst->data[1] = pair->second == (mfxMemId)MFX_INFINITE ? (uint8_t *)0 : pair->second; } else { dst->data[3] = child_data; } @@ -944,7 +944,7 @@ static int qsv_map_from(AVHWFramesContext *ctx, if (child_frames_ctx->device_ctx->type == AV_HWDEVICE_TYPE_D3D11VA) { mfxHDLPair *pair = (mfxHDLPair*)surf->Data.MemId; dummy->data[0] = pair->first; - dummy->data[1] = pair->second; + dummy->data[1] = pair->second == (mfxMemId)MFX_INFINITE ? (uint8_t *)0 : pair->second; } else { dummy->data[3] = child_data; } @@ -1362,7 +1362,8 @@ static int qsv_map_to(AVHWFramesContext *dst_ctx, { mfxHDLPair *pair = (mfxHDLPair*)hwctx->surfaces[i].Data.MemId; if (pair->first == src->data[0] - && pair->second == src->data[1]) { + && (pair->second == src->data[1] + || (pair->second == (mfxMemId)MFX_INFINITE && src->data[1] == (uint8_t *)0))) { index = i; break; }

1 0

avutil/hwcontext_d3d11va: pass the format value from outside for staging texture
by Tong Wu 23 May '22

23 May '22

ffmpeg | branch: master | Tong Wu <tong1.wu-at-intel.com(a)ffmpeg.org> | Sat May 7 06:24:40 2022 +0000| [20807a9d6128ac7a872949a8cc6c6b1672cb6942] | committer: Haihao Xiang avutil/hwcontext_d3d11va: pass the format value from outside for staging texture In d3d11va_create_staging_texture(), during the hwmap process, the ctx->internal->priv is not initialized, resulting in the texDesc.Format not initialized. Now pass the format value from d3d11va_transfer_data() to fix it. $ ffmpeg.exe -y -hwaccel qsv -init_hw_device d3d11va=d3d11 \ -init_hw_device qsv=qsv@d3d11 -c:v h264_qsv \ -i input.h264 -vf "hwmap=derive_device=d3d11va,format=d3d11,hwdownload,format=nv12" \ -f null - Reviewed-by: Soft Works <softworkz(a)hotmail.com> Signed-off-by: Tong Wu <tong1.wu(a)intel.com> Signed-off-by: Haihao Xiang <haihao.xiang(a)intel.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=20807a9d6128ac7a8… --- libavutil/hwcontext_d3d11va.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/libavutil/hwcontext_d3d11va.c b/libavutil/hwcontext_d3d11va.c index 8ab96bad25..904d14bbc8 100644 --- a/libavutil/hwcontext_d3d11va.c +++ b/libavutil/hwcontext_d3d11va.c @@ -339,7 +339,7 @@ static int d3d11va_transfer_get_formats(AVHWFramesContext *ctx, return 0; } -static int d3d11va_create_staging_texture(AVHWFramesContext *ctx) +static int d3d11va_create_staging_texture(AVHWFramesContext *ctx, DXGI_FORMAT format) { AVD3D11VADeviceContext *device_hwctx = ctx->device_ctx->hwctx; D3D11VAFramesContext *s = ctx->internal->priv; @@ -348,7 +348,7 @@ static int d3d11va_create_staging_texture(AVHWFramesContext *ctx) .Width = ctx->width, .Height = ctx->height, .MipLevels = 1, - .Format = s->format, + .Format = format, .SampleDesc = { .Count = 1 }, .ArraySize = 1, .Usage = D3D11_USAGE_STAGING, @@ -404,7 +404,8 @@ static int d3d11va_transfer_data(AVHWFramesContext *ctx, AVFrame *dst, device_hwctx->lock(device_hwctx->lock_ctx); if (!s->staging_texture) { - int res = d3d11va_create_staging_texture(ctx); + ID3D11Texture2D_GetDesc((ID3D11Texture2D *)texture, &desc); + int res = d3d11va_create_staging_texture(ctx, desc.Format); if (res < 0) return res; }

1 0

avutil/hwcontext_qsv: derive QSV frames to D3D11VA frames
by Tong Wu 23 May '22

23 May '22

ffmpeg | branch: master | Tong Wu <tong1.wu-at-intel.com(a)ffmpeg.org> | Sat May 7 06:24:39 2022 +0000| [632db3c36d7b81d7e15f02e8915a84e1b8296507] | committer: Haihao Xiang avutil/hwcontext_qsv: derive QSV frames to D3D11VA frames Fixes: $ ffmpeg.exe -y -hwaccel qsv -init_hw_device d3d11va=d3d11 \ -init_hw_device qsv=qsv@d3d11 -c:v h264_qsv -i input.h264 \ -vf "hwmap=derive_device=d3d11va,format=d3d11" -f null - Reviewed-by: Soft Works <softworkz(a)hotmail.com> Signed-off-by: Tong Wu <tong1.wu(a)intel.com> Signed-off-by: Haihao Xiang <haihao.xiang(a)intel.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=632db3c36d7b81d7e… --- libavutil/hwcontext_qsv.c | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) diff --git a/libavutil/hwcontext_qsv.c b/libavutil/hwcontext_qsv.c index 66c0e38955..6e4ba9d5d3 100644 --- a/libavutil/hwcontext_qsv.c +++ b/libavutil/hwcontext_qsv.c @@ -808,12 +808,22 @@ static int qsv_frames_derive_from(AVHWFramesContext *dst_ctx, #if CONFIG_D3D11VA case AV_HWDEVICE_TYPE_D3D11VA: { + D3D11_TEXTURE2D_DESC texDesc; + dst_ctx->initial_pool_size = src_ctx->initial_pool_size; AVD3D11VAFramesContext *dst_hwctx = dst_ctx->hwctx; - mfxHDLPair *pair = (mfxHDLPair*)src_hwctx->surfaces[i].Data.MemId; - dst_hwctx->texture = (ID3D11Texture2D*)pair->first; + dst_hwctx->texture_infos = av_calloc(src_hwctx->nb_surfaces, + sizeof(*dst_hwctx->texture_infos)); + if (!dst_hwctx->texture_infos) + return AVERROR(ENOMEM); if (src_hwctx->frame_type & MFX_MEMTYPE_SHARED_RESOURCE) dst_hwctx->MiscFlags = D3D11_RESOURCE_MISC_SHARED; - dst_hwctx->BindFlags = qsv_get_d3d11va_bind_flags(src_hwctx->frame_type); + for (i = 0; i < src_hwctx->nb_surfaces; i++) { + mfxHDLPair *pair = (mfxHDLPair*)src_hwctx->surfaces[i].Data.MemId; + dst_hwctx->texture_infos[i].texture = (ID3D11Texture2D*)pair->first; + dst_hwctx->texture_infos[i].index = pair->second == (mfxMemId)MFX_INFINITE ? (intptr_t)0 : (intptr_t)pair->second; + } + ID3D11Texture2D_GetDesc(dst_hwctx->texture_infos[0].texture, &texDesc); + dst_hwctx->BindFlags = texDesc.BindFlags; } break; #endif

1 0

avfilter/avf_showfreqs: fix memleak in plot_freqs
by Steven Liu 23 May '22

23 May '22

ffmpeg | branch: master | Steven Liu <lq(a)chinaffmpeg.org> | Fri May 13 10:19:24 2022 +0800| [90007e0b4e1b92674097dcc897a29bf8cb49cc3e] | committer: Steven Liu avfilter/avf_showfreqs: fix memleak in plot_freqs plot_freqs should free colors before return error when ff_get_video_buffer failed Signed-off-by: Steven Liu <lq(a)chinaffmpeg.org> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=90007e0b4e1b92674… --- libavfilter/avf_showfreqs.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/libavfilter/avf_showfreqs.c b/libavfilter/avf_showfreqs.c index 6dcff686b6..86a67c7328 100644 --- a/libavfilter/avf_showfreqs.c +++ b/libavfilter/avf_showfreqs.c @@ -421,8 +421,10 @@ static int plot_freqs(AVFilterLink *inlink, int64_t pts) return AVERROR(ENOMEM); out = ff_get_video_buffer(outlink, outlink->w, outlink->h); - if (!out) + if (!out) { + av_free(colors); return AVERROR(ENOMEM); + } for (n = 0; n < outlink->h; n++) memset(out->data[0] + out->linesize[0] * n, 0, outlink->w * 4);

1 0

avcodec/libuavs3d: fix access uninitialized variable when draining
by Zhao Zhili 23 May '22

23 May '22

ffmpeg | branch: master | Zhao Zhili <quinkblack(a)foxmail.com> | Thu May 12 17:55:31 2022 +0800| [a12ebbbcaa07b0ce77c413146ac0120e3f4f3a8d] | committer: Zhao Zhili avcodec/libuavs3d: fix access uninitialized variable when draining buf_ptr is uninitialized and accessed when function return. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=a12ebbbcaa07b0ce7… --- libavcodec/libuavs3d.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/libavcodec/libuavs3d.c b/libavcodec/libuavs3d.c index 6966e00b62..23de4c8cd5 100644 --- a/libavcodec/libuavs3d.c +++ b/libavcodec/libuavs3d.c @@ -149,7 +149,7 @@ static int libuavs3d_decode_frame(AVCodecContext *avctx, AVFrame *frm, const uint8_t *buf = avpkt->data; int buf_size = avpkt->size; const uint8_t *buf_end; - const uint8_t *buf_ptr; + const uint8_t *buf_ptr = buf; int left_bytes; int ret, finish = 0; @@ -170,7 +170,6 @@ static int libuavs3d_decode_frame(AVCodecContext *avctx, AVFrame *frm, } else { uavs3d_io_frm_t *frm_dec = &h->dec_frame; - buf_ptr = buf; buf_end = buf + buf_size; frm_dec->pkt_pos = avpkt->pos; frm_dec->pkt_size = avpkt->size;

1 0

avcodec/libuavs3d: fix access uninitialized variable when draining
by Zhao Zhili 23 May '22

23 May '22

ffmpeg | branch: master | Zhao Zhili <quinkblack(a)foxmail.com> | Thu May 12 17:55:31 2022 +0800| [a12ebbbcaa07b0ce77c413146ac0120e3f4f3a8d] | committer: Zhao Zhili avcodec/libuavs3d: fix access uninitialized variable when draining buf_ptr is uninitialized and accessed when function return. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=a12ebbbcaa07b0ce7… --- libavcodec/libuavs3d.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/libavcodec/libuavs3d.c b/libavcodec/libuavs3d.c index 6966e00b62..23de4c8cd5 100644 --- a/libavcodec/libuavs3d.c +++ b/libavcodec/libuavs3d.c @@ -149,7 +149,7 @@ static int libuavs3d_decode_frame(AVCodecContext *avctx, AVFrame *frm, const uint8_t *buf = avpkt->data; int buf_size = avpkt->size; const uint8_t *buf_end; - const uint8_t *buf_ptr; + const uint8_t *buf_ptr = buf; int left_bytes; int ret, finish = 0; @@ -170,7 +170,6 @@ static int libuavs3d_decode_frame(AVCodecContext *avctx, AVFrame *frm, } else { uavs3d_io_frm_t *frm_dec = &h->dec_frame; - buf_ptr = buf; buf_end = buf + buf_size; frm_dec->pkt_pos = avpkt->pos; frm_dec->pkt_size = avpkt->size;

1 0

x86/tx_float: replace fft_sr_avx with fft_sr_fma3
by Lynne 21 May '22

21 May '22

ffmpeg | branch: master | Lynne <dev(a)lynne.ee> | Sat May 21 01:39:12 2022 +0200| [27cffd16aadd04bf2311e3114b8aa81e3732552f] | committer: Lynne x86/tx_float: replace fft_sr_avx with fft_sr_fma3 When the SLOW_GATHER flag was added to the AVX2 version, this made FMA3-features not enabled on Zen CPUs. As FMA3 adds 6-7% across all platforms that support it, in the interest of saving space, this commit removes the AVX version and replaces it with an FMA3 version. The only CPUs affected are Sandy Bridge and Bulldozer, which have AVX support, but no FMA3 support. In the future, if there's a demand for it, a version of the function duplicated for AVX can be added. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=27cffd16aadd04bf2… --- libavutil/x86/tx_float.asm | 4 ++-- libavutil/x86/tx_float_init.c | 14 +++++++------- 2 files changed, 9 insertions(+), 9 deletions(-) diff --git a/libavutil/x86/tx_float.asm b/libavutil/x86/tx_float.asm index 511d8d6fa3..21f99d3945 100644 --- a/libavutil/x86/tx_float.asm +++ b/libavutil/x86/tx_float.asm @@ -1285,8 +1285,8 @@ FFT_SPLIT_RADIX_DEF 131072 %endmacro %if ARCH_X86_64 -FFT_SPLIT_RADIX_FN avx, float, 0 -FFT_SPLIT_RADIX_FN avx, ns_float, 1 +FFT_SPLIT_RADIX_FN fma3, float, 0 +FFT_SPLIT_RADIX_FN fma3, ns_float, 1 %if HAVE_AVX2_EXTERNAL FFT_SPLIT_RADIX_FN avx2, float, 0 FFT_SPLIT_RADIX_FN avx2, ns_float, 1 diff --git a/libavutil/x86/tx_float_init.c b/libavutil/x86/tx_float_init.c index 108f9b4b04..5db0b57d13 100644 --- a/libavutil/x86/tx_float_init.c +++ b/libavutil/x86/tx_float_init.c @@ -38,8 +38,8 @@ TX_DECL_FN(fft32, avx) TX_DECL_FN(fft32_ns, avx) TX_DECL_FN(fft32, fma3) TX_DECL_FN(fft32_ns, fma3) -TX_DECL_FN(fft_sr, avx) -TX_DECL_FN(fft_sr_ns, avx) +TX_DECL_FN(fft_sr, fma3) +TX_DECL_FN(fft_sr_ns, fma3) TX_DECL_FN(fft_sr, avx2) TX_DECL_FN(fft_sr_ns, avx2) @@ -88,13 +88,13 @@ const FFTXCodelet * const ff_tx_codelet_list_float_x86[] = { TX_DEF(fft32, FFT, 32, 32, 2, 0, 288, b8_i2, fma3, FMA3, AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW), TX_DEF(fft32_ns, FFT, 32, 32, 2, 0, 352, b8_i2, fma3, FMA3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, AV_CPU_FLAG_AVXSLOW), -#if HAVE_AVX2_EXTERNAL - TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 256, b8_i2, avx, AVX, 0, AV_CPU_FLAG_AVXSLOW), - TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 288, b8_i2, fma3, FMA3, 0, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 352, b8_i2, fma3, FMA3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, AV_CPU_FLAG_AVXSLOW), - TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 288, b8_i2, avx2, AVX2, 0, +#if HAVE_AVX2_EXTERNAL + TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 320, b8_i2, avx2, AVX2, 0, AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER), - TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 352, b8_i2, avx2, AVX2, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 384, b8_i2, avx2, AVX2, AV_TX_INPLACE | FF_TX_PRESHUFFLE, AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER), #endif #endif

1 0

x86/tx_float: improve temporary register allocation for loads
by Lynne 21 May '22

21 May '22

ffmpeg | branch: master | Lynne <dev(a)lynne.ee> | Sat May 21 00:50:09 2022 +0200| [0938ff97016884184796c3dd80bc6fcb729e875b] | committer: Lynne x86/tx_float: improve temporary register allocation for loads On Zen 3: Before: 1484285 decicycles in av_tx (fft), 131072 runs, 0 skips After: 1415243 decicycles in av_tx (fft), 131072 runs, 0 skips > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=0938ff97016884184… --- libavutil/x86/tx_float.asm | 48 +++++++++++++++++++++++----------------------- 1 file changed, 24 insertions(+), 24 deletions(-) diff --git a/libavutil/x86/tx_float.asm b/libavutil/x86/tx_float.asm index bab44b0947..511d8d6fa3 100644 --- a/libavutil/x86/tx_float.asm +++ b/libavutil/x86/tx_float.asm @@ -820,10 +820,10 @@ cglobal fft32_ %+ %2, 4, 4, 16, ctx, out, in, tmp movaps m7, [inq + 7*mmsize] %else mov ctxq, [ctxq + AVTXContext.map] - LOAD64_LUT m4, inq, ctxq, (mmsize/2)*4, tmpq, m8, m9 - LOAD64_LUT m5, inq, ctxq, (mmsize/2)*5, tmpq, m10, m11 - LOAD64_LUT m6, inq, ctxq, (mmsize/2)*6, tmpq, m12, m13 - LOAD64_LUT m7, inq, ctxq, (mmsize/2)*7, tmpq, m14, m15 + LOAD64_LUT m4, inq, ctxq, (mmsize/2)*4, tmpq, m8, m12 + LOAD64_LUT m5, inq, ctxq, (mmsize/2)*5, tmpq, m9, m13 + LOAD64_LUT m6, inq, ctxq, (mmsize/2)*6, tmpq, m10, m14 + LOAD64_LUT m7, inq, ctxq, (mmsize/2)*7, tmpq, m11, m15 %endif FFT8 m4, m5, m6, m7, m8, m9 @@ -834,10 +834,10 @@ cglobal fft32_ %+ %2, 4, 4, 16, ctx, out, in, tmp movaps m2, [inq + 2*mmsize] movaps m3, [inq + 3*mmsize] %else - LOAD64_LUT m0, inq, ctxq, (mmsize/2)*0, tmpq, m8, m9 - LOAD64_LUT m1, inq, ctxq, (mmsize/2)*1, tmpq, m10, m11 - LOAD64_LUT m2, inq, ctxq, (mmsize/2)*2, tmpq, m12, m13 - LOAD64_LUT m3, inq, ctxq, (mmsize/2)*3, tmpq, m14, m15 + LOAD64_LUT m0, inq, ctxq, (mmsize/2)*0, tmpq, m8, m12 + LOAD64_LUT m1, inq, ctxq, (mmsize/2)*1, tmpq, m9, m13 + LOAD64_LUT m2, inq, ctxq, (mmsize/2)*2, tmpq, m10, m14 + LOAD64_LUT m3, inq, ctxq, (mmsize/2)*3, tmpq, m11, m15 %endif movaps m8, [tab_32_float] @@ -939,10 +939,10 @@ ALIGN 16 movaps m6, [inq + 6*mmsize] movaps m7, [inq + 7*mmsize] %else - LOAD64_LUT m4, inq, lutq, (mmsize/2)*4, tmpq, m8, m9 - LOAD64_LUT m5, inq, lutq, (mmsize/2)*5, tmpq, m10, m11 - LOAD64_LUT m6, inq, lutq, (mmsize/2)*6, tmpq, m12, m13 - LOAD64_LUT m7, inq, lutq, (mmsize/2)*7, tmpq, m14, m15 + LOAD64_LUT m4, inq, lutq, (mmsize/2)*4, tmpq, m8, m12 + LOAD64_LUT m5, inq, lutq, (mmsize/2)*5, tmpq, m9, m13 + LOAD64_LUT m6, inq, lutq, (mmsize/2)*6, tmpq, m10, m14 + LOAD64_LUT m7, inq, lutq, (mmsize/2)*7, tmpq, m11, m15 %endif FFT8 m4, m5, m6, m7, m8, m9 @@ -953,10 +953,10 @@ ALIGN 16 movaps m2, [inq + 2*mmsize] movaps m3, [inq + 3*mmsize] %else - LOAD64_LUT m0, inq, lutq, (mmsize/2)*0, tmpq, m8, m9 - LOAD64_LUT m1, inq, lutq, (mmsize/2)*1, tmpq, m10, m11 - LOAD64_LUT m2, inq, lutq, (mmsize/2)*2, tmpq, m12, m13 - LOAD64_LUT m3, inq, lutq, (mmsize/2)*3, tmpq, m14, m15 + LOAD64_LUT m0, inq, lutq, (mmsize/2)*0, tmpq, m8, m12 + LOAD64_LUT m1, inq, lutq, (mmsize/2)*1, tmpq, m9, m13 + LOAD64_LUT m2, inq, lutq, (mmsize/2)*2, tmpq, m10, m14 + LOAD64_LUT m3, inq, lutq, (mmsize/2)*3, tmpq, m11, m15 %endif movaps m8, [tab_32_float] @@ -1013,10 +1013,10 @@ ALIGN 16 movaps tx1_o0, [inq + 2*mmsize] movaps tx1_o1, [inq + 3*mmsize] %else - LOAD64_LUT tx1_e0, inq, lutq, (mmsize/2)*0, tmpq, tw_e, tw_o - LOAD64_LUT tx1_e1, inq, lutq, (mmsize/2)*1, tmpq, tmp1, tmp2 - LOAD64_LUT tx1_o0, inq, lutq, (mmsize/2)*2, tmpq, tw_e, tw_o - LOAD64_LUT tx1_o1, inq, lutq, (mmsize/2)*3, tmpq, tmp1, tmp2 + LOAD64_LUT tx1_e0, inq, lutq, (mmsize/2)*0, tmpq, tw_e, tmp1 + LOAD64_LUT tx1_e1, inq, lutq, (mmsize/2)*1, tmpq, tw_o, tmp2 + LOAD64_LUT tx1_o0, inq, lutq, (mmsize/2)*2, tmpq, tw_e, tmp1 + LOAD64_LUT tx1_o1, inq, lutq, (mmsize/2)*3, tmpq, tw_o, tmp2 %endif FFT16 tx1_e0, tx1_e1, tx1_o0, tx1_o1, tw_e, tw_o, tx2_o0, tx2_o1 @@ -1027,10 +1027,10 @@ ALIGN 16 movaps tx2_o0, [inq + 6*mmsize] movaps tx2_o1, [inq + 7*mmsize] %else - LOAD64_LUT tx2_e0, inq, lutq, (mmsize/2)*4, tmpq, tmp1, tmp2 - LOAD64_LUT tx2_e1, inq, lutq, (mmsize/2)*5, tmpq, tw_e, tw_o - LOAD64_LUT tx2_o0, inq, lutq, (mmsize/2)*6, tmpq, tmp1, tmp2 - LOAD64_LUT tx2_o1, inq, lutq, (mmsize/2)*7, tmpq, tw_e, tw_o + LOAD64_LUT tx2_e0, inq, lutq, (mmsize/2)*4, tmpq, tw_e, tmp1 + LOAD64_LUT tx2_e1, inq, lutq, (mmsize/2)*5, tmpq, tw_o, tmp2 + LOAD64_LUT tx2_o0, inq, lutq, (mmsize/2)*6, tmpq, tw_e, tmp1 + LOAD64_LUT tx2_o1, inq, lutq, (mmsize/2)*7, tmpq, tw_o, tmp2 %endif FFT16 tx2_e0, tx2_e1, tx2_o0, tx2_o1, tmp1, tmp2, tw_e, tw_o

1 0

lavu/tx: make slow ISA extension penalties smarter
by Lynne 21 May '22

21 May '22

ffmpeg | branch: master | Lynne <dev(a)lynne.ee> | Sat May 21 00:04:11 2022 +0200| [805e8d1921438e36cb0c6f92fa42a93298df4053] | committer: Lynne lavu/tx: make slow ISA extension penalties smarter Instead of having a fixed -64 prio penalty, make the penalties more granular. As the prio is based on the register size in bits, decrementing it by 129 makes AVX SLOW functions be avoided in favor of any SSE versions. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=805e8d1921438e36c… --- libavutil/tx.c | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-) diff --git a/libavutil/tx.c b/libavutil/tx.c index c648d4a1cd..4cc3a98751 100644 --- a/libavutil/tx.c +++ b/libavutil/tx.c @@ -469,6 +469,15 @@ av_cold int ff_tx_init_subtx(AVTXContext *s, enum AVTXType type, AV_CPU_FLAG_ATOM | AV_CPU_FLAG_SSSE3SLOW | AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER; + static const int slow_penalties[][2] = { + { AV_CPU_FLAG_SSE2SLOW, 1 + 64 }, + { AV_CPU_FLAG_SSE3SLOW, 1 + 64 }, + { AV_CPU_FLAG_SSSE3SLOW, 1 + 64 }, + { AV_CPU_FLAG_ATOM, 1 + 128 }, + { AV_CPU_FLAG_AVXSLOW, 1 + 128 }, + { AV_CPU_FLAG_SLOW_GATHER, 1 + 32 }, + }; + /* Flags the transform wants */ uint64_t req_flags = flags; @@ -535,8 +544,10 @@ av_cold int ff_tx_init_subtx(AVTXContext *s, enum AVTXType type, /* If the CPU has a SLOW flag, and the instruction is also flagged * as being slow for such, reduce its priority */ - if ((cpu_flags & cd->cpu_flags) & slow_mask) - cd_matches[nb_cd_matches].prio -= 64; + for (int i = 0; i < FF_ARRAY_ELEMS(slow_penalties); i++) { + if ((cpu_flags & cd->cpu_flags) & slow_penalties[i][0]) + cd_matches[nb_cd_matches].prio -= slow_penalties[i][1]; + } /* Prioritize aligned-only codelets */ if ((cd->flags & FF_TX_ALIGNED) && !(cd->flags & AV_TX_UNALIGNED))

1 0

x86/tx_float: add AV_CPU_FLAG_AVXSLOW/SLOW_GATHER flags where appropriate
by Lynne 21 May '22

21 May '22

ffmpeg | branch: master | Lynne <dev(a)lynne.ee> | Sat May 21 00:11:45 2022 +0200| [19c0bb2aa91443d716f12fa3d1e019711243689c] | committer: Lynne x86/tx_float: add AV_CPU_FLAG_AVXSLOW/SLOW_GATHER flags where appropriate > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=19c0bb2aa91443d71… --- libavutil/x86/tx_float_init.c | 35 +++++++++++++++++++++-------------- 1 file changed, 21 insertions(+), 14 deletions(-) diff --git a/libavutil/x86/tx_float_init.c b/libavutil/x86/tx_float_init.c index e60faf7fa6..108f9b4b04 100644 --- a/libavutil/x86/tx_float_init.c +++ b/libavutil/x86/tx_float_init.c @@ -71,24 +71,31 @@ const FFTXCodelet * const ff_tx_codelet_list_float_x86[] = { TX_DEF(fft4_inv, FFT, 4, 4, 2, 0, 128, NULL, sse2, SSE2, AV_TX_INPLACE | FF_TX_INVERSE_ONLY, 0), TX_DEF(fft8, FFT, 8, 8, 2, 0, 128, b8_i0, sse3, SSE3, AV_TX_INPLACE, 0), TX_DEF(fft8_ns, FFT, 8, 8, 2, 0, 192, b8_i0, sse3, SSE3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), - TX_DEF(fft8, FFT, 8, 8, 2, 0, 256, b8_i0, avx, AVX, AV_TX_INPLACE, 0), - TX_DEF(fft8_ns, FFT, 8, 8, 2, 0, 320, b8_i0, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), - TX_DEF(fft16, FFT, 16, 16, 2, 0, 256, b8_i2, avx, AVX, AV_TX_INPLACE, 0), - TX_DEF(fft16_ns, FFT, 16, 16, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), - TX_DEF(fft16, FFT, 16, 16, 2, 0, 288, b8_i2, fma3, FMA3, AV_TX_INPLACE, 0), - TX_DEF(fft16_ns, FFT, 16, 16, 2, 0, 352, b8_i2, fma3, FMA3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), + TX_DEF(fft8, FFT, 8, 8, 2, 0, 256, b8_i0, avx, AVX, AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft8_ns, FFT, 8, 8, 2, 0, 320, b8_i0, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft16, FFT, 16, 16, 2, 0, 256, b8_i2, avx, AVX, AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft16_ns, FFT, 16, 16, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft16, FFT, 16, 16, 2, 0, 288, b8_i2, fma3, FMA3, AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft16_ns, FFT, 16, 16, 2, 0, 352, b8_i2, fma3, FMA3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + AV_CPU_FLAG_AVXSLOW), #if ARCH_X86_64 - TX_DEF(fft32, FFT, 32, 32, 2, 0, 256, b8_i2, avx, AVX, AV_TX_INPLACE, 0), - TX_DEF(fft32_ns, FFT, 32, 32, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), - TX_DEF(fft32, FFT, 32, 32, 2, 0, 288, b8_i2, fma3, FMA3, AV_TX_INPLACE, 0), - TX_DEF(fft32_ns, FFT, 32, 32, 2, 0, 352, b8_i2, fma3, FMA3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), + TX_DEF(fft32, FFT, 32, 32, 2, 0, 256, b8_i2, avx, AVX, AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft32_ns, FFT, 32, 32, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft32, FFT, 32, 32, 2, 0, 288, b8_i2, fma3, FMA3, AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft32_ns, FFT, 32, 32, 2, 0, 352, b8_i2, fma3, FMA3, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + AV_CPU_FLAG_AVXSLOW), #if HAVE_AVX2_EXTERNAL - TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 256, b8_i2, avx, AVX, 0, 0), - TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, 0), - TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 288, b8_i2, avx2, AVX2, 0, AV_CPU_FLAG_AVXSLOW), - TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 352, b8_i2, avx2, AVX2, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 256, b8_i2, avx, AVX, 0, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 320, b8_i2, avx, AVX, AV_TX_INPLACE | FF_TX_PRESHUFFLE, AV_CPU_FLAG_AVXSLOW), + TX_DEF(fft_sr, FFT, 64, 131072, 2, 0, 288, b8_i2, avx2, AVX2, 0, + AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER), + TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 352, b8_i2, avx2, AVX2, AV_TX_INPLACE | FF_TX_PRESHUFFLE, + AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER), #endif #endif

1 0