On 4/4/2017 10:53 PM, James Darnley wrote:
> Haswell:
> - 1.11x faster (522±0.4 vs. 469±1.8 decicycles) compared with mmxext
>
> Skylake-U:
> - 1.21x faster (671±5.5 vs. 555±1.4 decicycles) compared with mmxext
Again, you should add an SSE2 version first, then an AVX one if it's
measurably faster than the SSE2 one.
> ---
> libavcodec/x86/h264_idct.asm | 33 ++++++++++++++++++++++++++++++++-
> libavcodec/x86/h264dsp_init.c | 3 +++
> 2 files changed, 35 insertions(+), 1 deletion(-)
>
> diff --git a/libavcodec/x86/h264_idct.asm b/libavcodec/x86/h264_idct.asm
> index bc4dce4..24fb4d2 100644
> --- a/libavcodec/x86/h264_idct.asm
> +++ b/libavcodec/x86/h264_idct.asm
> @@ -65,7 +65,15 @@ SECTION .text
>
> IDCT4_1D w, 0, 1, 2, 3, 4, 5
> mova m6, [pw_32]
> - TRANSPOSE4x4W 0, 1, 2, 3, 4
> + %if mmsize == 8
> + TRANSPOSE4x4W 0, 1, 2, 3, 4
> + %else
> + punpcklwd m0, m1
> + punpcklwd m2, m3
> + SBUTTERFLY dq, 0, 2, 4
> + MOVHL m1, m0
> + MOVHL m3, m2
> + %endif
> paddw m0, m6
> IDCT4_1D w, 0, 1, 2, 3, 4, 5
> pxor m7, m7
> @@ -1131,3 +1139,26 @@ INIT_MMX mmx
> IDCT_DC_DEQUANT 0
> INIT_MMX sse2
> IDCT_DC_DEQUANT 7
> +
> +INIT_XMM avx
> +
> +; %unmacro STORE_DIFFx2 8 ; remove macro from x86util.asm but yasm doesn't
> have this yet
> +%macro STORE_DIFFx2 8 ; add1, add2, reg1, reg2, zero, shift, source, stride
> + movd %3, [%7]
> + movd %4, [%7+%8]
> + psraw %1, %6
> + psraw %2, %6
> + punpcklbw %3, %5
> + punpcklbw %4, %5
> + paddw %3, %1
> + paddw %4, %2
> + packuswb %3, %5
> + packuswb %4, %5
> + movd [%7], %3
> + movd [%7+%8], %4
> +%endmacro
> +
> +cglobal h264_idct_add_8, 3, 3, 8, dst_, block_, stride_
> + movsxdifnidn stride_q, stride_d
> + IDCT4_ADD dst_q, block_q, stride_q
> +RET
> diff --git a/libavcodec/x86/h264dsp_init.c b/libavcodec/x86/h264dsp_init.c
> index 0643b37..8ba085f 100644
> --- a/libavcodec/x86/h264dsp_init.c
> +++ b/libavcodec/x86/h264dsp_init.c
> @@ -32,6 +32,7 @@ void ff_h264_idct ## NUM ## _add_ ## DEPTH ## _ ##
> OPT(uint8_t *dst, \
> int stride);
>
> IDCT_ADD_FUNC(, 8, mmx)
> +IDCT_ADD_FUNC(, 8, avx)
> IDCT_ADD_FUNC(, 10, sse2)
> IDCT_ADD_FUNC(_dc, 8, mmxext)
> IDCT_ADD_FUNC(_dc, 10, mmxext)
> @@ -337,6 +338,8 @@ av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const
> int bit_depth,
> c->h264_h_loop_filter_chroma =
> ff_deblock_h_chroma422_8_avx;
> c->h264_h_loop_filter_chroma_intra =
> ff_deblock_h_chroma422_intra_8_avx;
> }
> +
> + c->h264_idct_add = ff_h264_idct_add_8_avx;
> }
> } else if (bit_depth == 10) {
> if (EXTERNAL_MMXEXT(cpu_flags)) {
>
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
