[FFmpeg-devel] [PATCH 5/6] lavc/apv: AVX2 transquant for x86-64

James Almer jamrial at gmail.com
Sun Apr 20 04:48:29 EEST 2025


On 4/19/2025 4:07 PM, Mark Thompson wrote:
> Typical checkasm result on Alder Lake:
> 
> decode_transquant_8_c:                                 408.7 ( 1.00x)
> decode_transquant_8_avx2:                               94.2 ( 4.34x)
> decode_transquant_10_c:                                413.1 ( 1.00x)
> decode_transquant_10_avx2:                              87.5 ( 4.72x)
> ---
>   libavcodec/apv_dsp.c          |   4 +
>   libavcodec/apv_dsp.h          |   2 +
>   libavcodec/x86/Makefile       |   2 +
>   libavcodec/x86/apv_dsp.asm    | 243 ++++++++++++++++++++++++++++++++++
>   libavcodec/x86/apv_dsp_init.c |  41 ++++++
>   tests/checkasm/Makefile       |   1 +
>   tests/checkasm/apv_dsp.c      | 113 ++++++++++++++++
>   tests/checkasm/checkasm.c     |   3 +
>   tests/checkasm/checkasm.h     |   1 +
>   9 files changed, 410 insertions(+)
>   create mode 100644 libavcodec/x86/apv_dsp.asm
>   create mode 100644 libavcodec/x86/apv_dsp_init.c
>   create mode 100644 tests/checkasm/apv_dsp.c

[...]

> diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile
> index d5c50e5599..193c1e4633 100644
> --- a/tests/checkasm/Makefile
> +++ b/tests/checkasm/Makefile
> @@ -28,6 +28,7 @@ AVCODECOBJS-$(CONFIG_AAC_DECODER)       += aacpsdsp.o \
>                                              sbrdsp.o
>   AVCODECOBJS-$(CONFIG_AAC_ENCODER)       += aacencdsp.o
>   AVCODECOBJS-$(CONFIG_ALAC_DECODER)      += alacdsp.o
> +AVCODECOBJS-$(CONFIG_APV_DECODER)       += apv_dsp.o
>   AVCODECOBJS-$(CONFIG_DCA_DECODER)       += synth_filter.o
>   AVCODECOBJS-$(CONFIG_DIRAC_DECODER)     += diracdsp.o
>   AVCODECOBJS-$(CONFIG_EXR_DECODER)       += exrdsp.o
> diff --git a/tests/checkasm/apv_dsp.c b/tests/checkasm/apv_dsp.c
> new file mode 100644
> index 0000000000..a0272d8edc
> --- /dev/null
> +++ b/tests/checkasm/apv_dsp.c
> @@ -0,0 +1,113 @@
> +/*
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +
> +#include <stdint.h>
> +
> +#include "checkasm.h"
> +
> +#include "libavutil/attributes.h"
> +#include "libavutil/mem_internal.h"
> +#include "libavcodec/apv_dsp.h"
> +
> +
> +static void check_decode_transquant_8(void)
> +{
> +    LOCAL_ALIGNED_16(int16_t, input,      [64]);
> +    LOCAL_ALIGNED_16(int16_t, qmatrix,    [64]);
> +    LOCAL_ALIGNED_16(uint8_t, new_output, [64]);
> +    LOCAL_ALIGNED_16(uint8_t, ref_output, [64]);
> +
> +    declare_func(void,
> +                 uint8_t *output,
> +                 ptrdiff_t pitch,
> +                 const int16_t *input,
> +                 const int16_t *qmatrix,
> +                 int64_t bit_depth,
> +                 int64_t qp_shift);
> +
> +    for (int i = 0; i < 64; i++) {
> +        // Any signed 12-bit integer.
> +        input[i] = rnd() % 2048 - 1024;
> +
> +        // qmatrix input is premultiplied by level_scale, so
> +        // range is 1 to 255 * 71.  Interesting values are all
> +        // at the low end of that, though.
> +        qmatrix[i] = rnd() % 16 + 16;
> +    }
> +
> +    call_ref(ref_output, 8, input, qmatrix, 8, 4);
> +    call_new(new_output, 8, input, qmatrix, 8, 4);
> +
> +    for (int i = 0; i < 64; i++) {
> +        if (ref_output[i] != new_output[i])

memcmp?

> +            fail();
> +    }
> +
> +    bench_new(new_output, 8, input, qmatrix, 8, 4);
> +}
> +
> +static void check_decode_transquant_10(void)
> +{
> +    LOCAL_ALIGNED_16( int16_t, input,      [64]);
> +    LOCAL_ALIGNED_16( int16_t, qmatrix,    [64]);
> +    LOCAL_ALIGNED_16(uint16_t, new_output, [64]);
> +    LOCAL_ALIGNED_16(uint16_t, ref_output, [64]);
> +
> +    declare_func(void,
> +                 uint16_t *output,
> +                 ptrdiff_t pitch,
> +                 const int16_t *input,
> +                 const int16_t *qmatrix,
> +                 int64_t bit_depth,
> +                 int64_t qp_shift);
> +
> +    for (int i = 0; i < 64; i++) {
> +        // Any signed 14-bit integer.
> +        input[i] = rnd() % 16384 - 8192;
> +
> +        // qmatrix input is premultiplied by level_scale, so
> +        // range is 1 to 255 * 71.  Interesting values are all
> +        // at the low end of that, though.
> +        qmatrix[i] = 16; //rnd() % 16 + 16;
> +    }
> +
> +    call_ref(ref_output, 16, input, qmatrix, 10, 4);
> +    call_new(new_output, 16, input, qmatrix, 10, 4);
> +
> +    for (int i = 0; i < 64; i++) {
> +        if (ref_output[i] != new_output[i])
> +            fail();
> +    }
> +
> +    bench_new(new_output, 16, input, qmatrix, 10, 4);
> +}
> +
> +void checkasm_check_apv_dsp(void)
> +{
> +    APVDSPContext dsp;
> +
> +    ff_apv_dsp_init(&dsp);
> +
> +    if (check_func(dsp.decode_transquant, "decode_transquant_8"))
> +        check_decode_transquant_8();
> +
> +    if (check_func(dsp.decode_transquant, "decode_transquant_10"))
> +        check_decode_transquant_10();
> +
> +    report("apv_dsp");
> +}
> diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
> index 412b8b2cd1..3bb82ed0e5 100644
> --- a/tests/checkasm/checkasm.c
> +++ b/tests/checkasm/checkasm.c
> @@ -129,6 +129,9 @@ static const struct {
>       #if CONFIG_ALAC_DECODER
>           { "alacdsp", checkasm_check_alacdsp },
>       #endif
> +    #if CONFIG_APV_DECODER
> +        { "apv_dsp", checkasm_check_apv_dsp },
> +    #endif
>       #if CONFIG_AUDIODSP
>           { "audiodsp", checkasm_check_audiodsp },
>       #endif
> diff --git a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h
> index ad239fb2a4..a6b5965e02 100644
> --- a/tests/checkasm/checkasm.h
> +++ b/tests/checkasm/checkasm.h
> @@ -83,6 +83,7 @@ void checkasm_check_ac3dsp(void);
>   void checkasm_check_aes(void);
>   void checkasm_check_afir(void);
>   void checkasm_check_alacdsp(void);
> +void checkasm_check_apv_dsp(void);
>   void checkasm_check_audiodsp(void);
>   void checkasm_check_av_tx(void);
>   void checkasm_check_blend(void);

Also add an entry to tests/fate/checkasm.mak while at it.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature.asc
Type: application/pgp-signature
Size: 495 bytes
Desc: OpenPGP digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20250419/b22dc188/attachment.sig>


More information about the ffmpeg-devel mailing list