checkasm: Add tests for h264 idct

Message ID 1466166800-97644-1-git-send-email-martin@martin.st
State Committed
Headers show

Commit Message

Martin Storsjö June 17, 2016, 12:33 p.m.
The idct tests are inspired by similar tests for vp9 by
Ronald Bultje.
---
Now using a fixed stride as a multiple of 16 (16 exactly),
and testing all valid alignments within that.
---
 tests/checkasm/Makefile   |   1 +
 tests/checkasm/checkasm.c |   3 +
 tests/checkasm/checkasm.h |   1 +
 tests/checkasm/h264dsp.c  | 232 ++++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 237 insertions(+)
 create mode 100644 tests/checkasm/h264dsp.c

Comments

Janne Grunau June 17, 2016, 1:42 p.m. | #1
On 2016-06-17 15:33:20 +0300, Martin Storsjö wrote:
> The idct tests are inspired by similar tests for vp9 by
> Ronald Bultje.
> ---
> Now using a fixed stride as a multiple of 16 (16 exactly),
> and testing all valid alignments within that.
> ---
>  tests/checkasm/Makefile   |   1 +
>  tests/checkasm/checkasm.c |   3 +
>  tests/checkasm/checkasm.h |   1 +
>  tests/checkasm/h264dsp.c  | 232 ++++++++++++++++++++++++++++++++++++++++++++++
>  4 files changed, 237 insertions(+)
>  create mode 100644 tests/checkasm/h264dsp.c
> 
> diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile
> index 8a0cee5..3c23853 100644
> --- a/tests/checkasm/Makefile
> +++ b/tests/checkasm/Makefile
> @@ -2,6 +2,7 @@
>  # subsystems
>  AVCODECOBJS-$(CONFIG_BSWAPDSP)          += bswapdsp.o
>  AVCODECOBJS-$(CONFIG_FMTCONVERT)        += fmtconvert.o
> +AVCODECOBJS-$(CONFIG_H264DSP)           += h264dsp.o
>  AVCODECOBJS-$(CONFIG_H264PRED)          += h264pred.o
>  AVCODECOBJS-$(CONFIG_H264QPEL)          += h264qpel.o
>  
> diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
> index c75e431..15f9f68 100644
> --- a/tests/checkasm/checkasm.c
> +++ b/tests/checkasm/checkasm.c
> @@ -74,6 +74,9 @@ static const struct {
>  #if CONFIG_FMTCONVERT
>      { "fmtconvert", checkasm_check_fmtconvert },
>  #endif
> +#if CONFIG_H264DSP
> +    { "h264dsp", checkasm_check_h264dsp },
> +#endif
>  #if CONFIG_H264PRED
>      { "h264pred", checkasm_check_h264pred },
>  #endif
> diff --git a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h
> index c7aa19c..619ebc7 100644
> --- a/tests/checkasm/checkasm.h
> +++ b/tests/checkasm/checkasm.h
> @@ -33,6 +33,7 @@
>  void checkasm_check_bswapdsp(void);
>  void checkasm_check_dcadsp(void);
>  void checkasm_check_fmtconvert(void);
> +void checkasm_check_h264dsp(void);
>  void checkasm_check_h264pred(void);
>  void checkasm_check_h264qpel(void);
>  void checkasm_check_hevc_mc(void);
> diff --git a/tests/checkasm/h264dsp.c b/tests/checkasm/h264dsp.c
> new file mode 100644
> index 0000000..88beed0
> --- /dev/null
> +++ b/tests/checkasm/h264dsp.c
> @@ -0,0 +1,232 @@
> +/*
> + * Copyright (c) 2016 Martin Storsjo
> + *
> + * This file is part of Libav.
> + *
> + * Libav is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License as published by
> + * the Free Software Foundation; either version 2 of the License, or
> + * (at your option) any later version.
> + *
> + * Libav 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 General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public License along
> + * with Libav; if not, write to the Free Software Foundation, Inc.,
> + * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
> + */
> +
> +#include <string.h>
> +#include "checkasm.h"
> +#include "libavcodec/avcodec.h"
> +#include "libavcodec/h264dsp.h"
> +#include "libavutil/common.h"
> +#include "libavutil/internal.h"
> +#include "libavutil/intreadwrite.h"
> +
> +static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
> +
> +#define SIZEOF_PIXEL ((bit_depth + 7) / 8)
> +#define SIZEOF_COEF  (2 * ((bit_depth + 7) / 8))
> +#define PIXEL_STRIDE 16
> +
> +#define randomize_buffers()                                                  \
> +    do {                                                                     \
> +        uint32_t mask = pixel_mask[bit_depth - 8];                           \
> +        for (y = 0; y < sz; y++) {                                           \
> +            for (x = 0; x < sz * SIZEOF_PIXEL; x += 4) {                     \
> +                AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask);          \
> +                AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask);          \
> +            }                                                                \
> +            for (x = 0; x < sz; x++) {                                       \
> +                if (bit_depth == 8) {                                        \
> +                    coef[y * sz + x] = src[y * PIXEL_STRIDE + x] -           \
> +                                       dst[y * PIXEL_STRIDE + x];            \
> +                } else {                                                     \
> +                    ((int32_t *)coef)[y * sz + x] =                          \
> +                        ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] -        \
> +                        ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x];         \
> +                }                                                            \
> +            }                                                                \
> +        }                                                                    \
> +    } while (0)
> +
> +#define dct4x4_impl(size, dctcoef)                                           \
> +static void dct4x4_##size(dctcoef *coef)                                     \
> +{                                                                            \
> +    int i, y, x;                                                             \
> +    dctcoef tmp[16];                                                         \
> +    for (i = 0; i < 4; i++) {                                                \
> +        const int z0 = coef[i*4 + 0] + coef[i*4 + 3];                        \
> +        const int z1 = coef[i*4 + 1] + coef[i*4 + 2];                        \
> +        const int z2 = coef[i*4 + 0] - coef[i*4 + 3];                        \
> +        const int z3 = coef[i*4 + 1] - coef[i*4 + 2];                        \
> +        tmp[i + 4*0] =   z0 +   z1;                                          \
> +        tmp[i + 4*1] = 2*z2 +   z3;                                          \
> +        tmp[i + 4*2] =   z0 -   z1;                                          \
> +        tmp[i + 4*3] =   z2 - 2*z3;                                          \
> +    }                                                                        \
> +    for (i = 0; i < 4; i++) {                                                \
> +        const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3];                          \
> +        const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2];                          \
> +        const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3];                          \
> +        const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2];                          \
> +        coef[i*4 + 0] =   z0 +   z1;                                         \
> +        coef[i*4 + 1] = 2*z2 +   z3;                                         \
> +        coef[i*4 + 2] =   z0 -   z1;                                         \
> +        coef[i*4 + 3] =   z2 - 2*z3;                                         \
> +    }                                                                        \
> +    for (y = 0; y < 4; y++) {                                                \
> +        for (x = 0; x < 4; x++) {                                            \
> +            static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
> +            const int idx = (y & 1) + (x & 1);                               \
> +            coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15;  \
> +        }                                                                    \
> +    }                                                                        \
> +}
> +
> +#define DCT8_1D(src, srcstride, dst, dststride) do {                         \
> +    const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7];              \
> +    const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7];              \
> +    const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6];              \
> +    const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6];              \
> +    const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5];              \
> +    const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5];              \
> +    const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4];              \
> +    const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4];              \
> +    const int b0 = a0 + a6;                                                  \
> +    const int b1 = a2 + a4;                                                  \
> +    const int b2 = a0 - a6;                                                  \
> +    const int b3 = a2 - a4;                                                  \
> +    const int b4 = a3 + a5 + (a1 + (a1 >> 1));                               \
> +    const int b5 = a1 - a7 - (a5 + (a5 >> 1));                               \
> +    const int b6 = a1 + a7 - (a3 + (a3 >> 1));                               \
> +    const int b7 = a3 - a5 + (a7 + (a7 >> 1));                               \
> +    (dst)[dststride * 0] =  b0 +  b1;                                        \
> +    (dst)[dststride * 1] =  b4 + (b7 >> 2);                                  \
> +    (dst)[dststride * 2] =  b2 + (b3 >> 1);                                  \
> +    (dst)[dststride * 3] =  b5 + (b6 >> 2);                                  \
> +    (dst)[dststride * 4] =  b0  - b1;                                        \
> +    (dst)[dststride * 5] =  b6 - (b5 >> 2);                                  \
> +    (dst)[dststride * 6] = (b2 >> 1) - b3;                                   \
> +    (dst)[dststride * 7] = (b4 >> 2) - b7;                                   \
> +} while (0)
> +
> +#define dct8x8_impl(size, dctcoef)                                           \
> +static void dct8x8_##size(dctcoef *coef)                                     \
> +{                                                                            \
> +    int i, x, y;                                                             \
> +    dctcoef tmp[64];                                                         \
> +    for (i = 0; i < 8; i++)                                                  \
> +        DCT8_1D(coef + i, 8, tmp + i, 8);                                    \
> +                                                                             \
> +    for (i = 0; i < 8; i++)                                                  \
> +        DCT8_1D(tmp + 8*i, 1, coef + i, 8);                                  \
> +                                                                             \
> +    for (y = 0; y < 8; y++) {                                                \
> +        for (x = 0; x < 8; x++) {                                            \
> +            static const int scale[] = {                                     \
> +                13107 * 20, 11428 * 18, 20972 * 32,                          \
> +                12222 * 19, 16777 * 25, 15481 * 24,                          \
> +            };                                                               \
> +            static const int idxmap[] = {                                    \
> +                0, 3, 4, 3,                                                  \
> +                3, 1, 5, 1,                                                  \
> +                4, 5, 2, 5,                                                  \
> +                3, 1, 5, 1,                                                  \
> +            };                                                               \
> +            const int idx = idxmap[(y & 3) * 4 + (x & 3)];                   \
> +            coef[y*8 + x] = ((int64_t)coef[y*8 + x] *                        \
> +                             scale[idx] + (1 << 17)) >> 18;                  \
> +        }                                                                    \
> +    }                                                                        \
> +}
> +
> +dct4x4_impl(16, int16_t)
> +dct4x4_impl(32, int32_t)
> +
> +dct8x8_impl(16, int16_t)
> +dct8x8_impl(32, int32_t)
> +
> +static void dct4x4(int16_t *coef, int bit_depth)
> +{
> +    if (bit_depth == 8)
> +        dct4x4_16(coef);
> +    else
> +        dct4x4_32((int32_t *) coef);
> +}
> +
> +static void dct8x8(int16_t *coef, int bit_depth)
> +{
> +    if (bit_depth == 8) {
> +        dct8x8_16(coef);
> +    } else {
> +        dct8x8_32((int32_t *) coef);
> +    }
> +}
> +
> +
> +static void check_idct(void)
> +{
> +    LOCAL_ALIGNED_16(uint8_t, src,  [8 * (16 + 8 * 2)]);
> +    LOCAL_ALIGNED_16(uint8_t, dst,  [8 * (16 + 8 * 2)]);
> +    LOCAL_ALIGNED_16(uint8_t, dst0, [8 * (16 + 8 * 2)]);
> +    LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * (16 + 8 * 2) + 16]);

What's the reason for making the stride 32? 8 * 16 + 32 should be 
enough. + 16 for the varying alignment, + 16 padding for the altivec 
unaligned stores.

> +    LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
> +    LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
> +    LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
> +    H264DSPContext h;
> +    int bit_depth, sz, align;
> +    int x, y, dc;
> +    declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
> +
> +    for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
> +        ff_h264dsp_init(&h, bit_depth, 1);
> +        for (sz = 4; sz <= 8; sz += 4) {
> +            randomize_buffers();
> +
> +            if (sz == 4)
> +                dct4x4(coef, bit_depth);
> +            else
> +                dct8x8(coef, bit_depth);
> +
> +            for (dc = 0; dc <= 1; dc++) {
> +                void (*idct)(uint8_t *, int16_t *, int) = NULL;
> +                switch ((sz << 1) | dc) {
> +                case (4 << 1) | 0: idct = h.h264_idct_add; break;
> +                case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
> +                case (8 << 1) | 0: idct = h.h264_idct8_add; break;
> +                case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
> +                }
> +                if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
> +                    for (align = 0; align < 16; align += sz * PIXEL_STRIDE) {

s/PIXEL_STRIDE/PIXEL_SIZE/

> +                        uint8_t *dst1 = dst1_base + align;
> +                        if (dc) {
> +                            memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
> +                            memcpy(subcoef0, coef, SIZEOF_COEF);
> +                        } else {
> +                            memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
> +                        }
> +                        memcpy(dst0, dst, sz * PIXEL_STRIDE);
> +                        memcpy(dst1, dst, sz * PIXEL_STRIDE);
> +                        memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
> +                        call_ref(dst0, subcoef0, PIXEL_STRIDE);
> +                        call_new(dst1, subcoef1, PIXEL_STRIDE);
> +                        if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
> +                            memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
> +                            fail();
> +                        bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
> +                    }
> +                }
> +            }
> +        }
> +    }
> +    report("idct");
> +}
> +
> +void checkasm_check_h264dsp(void)
> +{
> +    check_idct();
> +}
> -- 

otherwise ok

Janne
Martin Storsjö June 17, 2016, 4:40 p.m. | #2
On Fri, 17 Jun 2016, Janne Grunau wrote:

> On 2016-06-17 15:33:20 +0300, Martin Storsjö wrote:
>> The idct tests are inspired by similar tests for vp9 by
>> Ronald Bultje.
>> ---
>> Now using a fixed stride as a multiple of 16 (16 exactly),
>> and testing all valid alignments within that.
>> ---
>>  tests/checkasm/Makefile   |   1 +
>>  tests/checkasm/checkasm.c |   3 +
>>  tests/checkasm/checkasm.h |   1 +
>>  tests/checkasm/h264dsp.c  | 232 ++++++++++++++++++++++++++++++++++++++++++++++
>>  4 files changed, 237 insertions(+)
>>  create mode 100644 tests/checkasm/h264dsp.c
>>
>> diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile
>> index 8a0cee5..3c23853 100644
>> --- a/tests/checkasm/Makefile
>> +++ b/tests/checkasm/Makefile
>> @@ -2,6 +2,7 @@
>>  # subsystems
>>  AVCODECOBJS-$(CONFIG_BSWAPDSP)          += bswapdsp.o
>>  AVCODECOBJS-$(CONFIG_FMTCONVERT)        += fmtconvert.o
>> +AVCODECOBJS-$(CONFIG_H264DSP)           += h264dsp.o
>>  AVCODECOBJS-$(CONFIG_H264PRED)          += h264pred.o
>>  AVCODECOBJS-$(CONFIG_H264QPEL)          += h264qpel.o
>>
>> diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
>> index c75e431..15f9f68 100644
>> --- a/tests/checkasm/checkasm.c
>> +++ b/tests/checkasm/checkasm.c
>> @@ -74,6 +74,9 @@ static const struct {
>>  #if CONFIG_FMTCONVERT
>>      { "fmtconvert", checkasm_check_fmtconvert },
>>  #endif
>> +#if CONFIG_H264DSP
>> +    { "h264dsp", checkasm_check_h264dsp },
>> +#endif
>>  #if CONFIG_H264PRED
>>      { "h264pred", checkasm_check_h264pred },
>>  #endif
>> diff --git a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h
>> index c7aa19c..619ebc7 100644
>> --- a/tests/checkasm/checkasm.h
>> +++ b/tests/checkasm/checkasm.h
>> @@ -33,6 +33,7 @@
>>  void checkasm_check_bswapdsp(void);
>>  void checkasm_check_dcadsp(void);
>>  void checkasm_check_fmtconvert(void);
>> +void checkasm_check_h264dsp(void);
>>  void checkasm_check_h264pred(void);
>>  void checkasm_check_h264qpel(void);
>>  void checkasm_check_hevc_mc(void);
>> diff --git a/tests/checkasm/h264dsp.c b/tests/checkasm/h264dsp.c
>> new file mode 100644
>> index 0000000..88beed0
>> --- /dev/null
>> +++ b/tests/checkasm/h264dsp.c
>> @@ -0,0 +1,232 @@
>> +/*
>> + * Copyright (c) 2016 Martin Storsjo
>> + *
>> + * This file is part of Libav.
>> + *
>> + * Libav is free software; you can redistribute it and/or modify
>> + * it under the terms of the GNU General Public License as published by
>> + * the Free Software Foundation; either version 2 of the License, or
>> + * (at your option) any later version.
>> + *
>> + * Libav 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 General Public License for more details.
>> + *
>> + * You should have received a copy of the GNU General Public License along
>> + * with Libav; if not, write to the Free Software Foundation, Inc.,
>> + * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
>> + */
>> +
>> +#include <string.h>
>> +#include "checkasm.h"
>> +#include "libavcodec/avcodec.h"
>> +#include "libavcodec/h264dsp.h"
>> +#include "libavutil/common.h"
>> +#include "libavutil/internal.h"
>> +#include "libavutil/intreadwrite.h"
>> +
>> +static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
>> +
>> +#define SIZEOF_PIXEL ((bit_depth + 7) / 8)
>> +#define SIZEOF_COEF  (2 * ((bit_depth + 7) / 8))
>> +#define PIXEL_STRIDE 16
>> +
>> +#define randomize_buffers()                                                  \
>> +    do {                                                                     \
>> +        uint32_t mask = pixel_mask[bit_depth - 8];                           \
>> +        for (y = 0; y < sz; y++) {                                           \
>> +            for (x = 0; x < sz * SIZEOF_PIXEL; x += 4) {                     \
>> +                AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask);          \
>> +                AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask);          \
>> +            }                                                                \
>> +            for (x = 0; x < sz; x++) {                                       \
>> +                if (bit_depth == 8) {                                        \
>> +                    coef[y * sz + x] = src[y * PIXEL_STRIDE + x] -           \
>> +                                       dst[y * PIXEL_STRIDE + x];            \
>> +                } else {                                                     \
>> +                    ((int32_t *)coef)[y * sz + x] =                          \
>> +                        ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] -        \
>> +                        ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x];         \
>> +                }                                                            \
>> +            }                                                                \
>> +        }                                                                    \
>> +    } while (0)
>> +
>> +#define dct4x4_impl(size, dctcoef)                                           \
>> +static void dct4x4_##size(dctcoef *coef)                                     \
>> +{                                                                            \
>> +    int i, y, x;                                                             \
>> +    dctcoef tmp[16];                                                         \
>> +    for (i = 0; i < 4; i++) {                                                \
>> +        const int z0 = coef[i*4 + 0] + coef[i*4 + 3];                        \
>> +        const int z1 = coef[i*4 + 1] + coef[i*4 + 2];                        \
>> +        const int z2 = coef[i*4 + 0] - coef[i*4 + 3];                        \
>> +        const int z3 = coef[i*4 + 1] - coef[i*4 + 2];                        \
>> +        tmp[i + 4*0] =   z0 +   z1;                                          \
>> +        tmp[i + 4*1] = 2*z2 +   z3;                                          \
>> +        tmp[i + 4*2] =   z0 -   z1;                                          \
>> +        tmp[i + 4*3] =   z2 - 2*z3;                                          \
>> +    }                                                                        \
>> +    for (i = 0; i < 4; i++) {                                                \
>> +        const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3];                          \
>> +        const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2];                          \
>> +        const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3];                          \
>> +        const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2];                          \
>> +        coef[i*4 + 0] =   z0 +   z1;                                         \
>> +        coef[i*4 + 1] = 2*z2 +   z3;                                         \
>> +        coef[i*4 + 2] =   z0 -   z1;                                         \
>> +        coef[i*4 + 3] =   z2 - 2*z3;                                         \
>> +    }                                                                        \
>> +    for (y = 0; y < 4; y++) {                                                \
>> +        for (x = 0; x < 4; x++) {                                            \
>> +            static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
>> +            const int idx = (y & 1) + (x & 1);                               \
>> +            coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15;  \
>> +        }                                                                    \
>> +    }                                                                        \
>> +}
>> +
>> +#define DCT8_1D(src, srcstride, dst, dststride) do {                         \
>> +    const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7];              \
>> +    const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7];              \
>> +    const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6];              \
>> +    const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6];              \
>> +    const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5];              \
>> +    const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5];              \
>> +    const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4];              \
>> +    const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4];              \
>> +    const int b0 = a0 + a6;                                                  \
>> +    const int b1 = a2 + a4;                                                  \
>> +    const int b2 = a0 - a6;                                                  \
>> +    const int b3 = a2 - a4;                                                  \
>> +    const int b4 = a3 + a5 + (a1 + (a1 >> 1));                               \
>> +    const int b5 = a1 - a7 - (a5 + (a5 >> 1));                               \
>> +    const int b6 = a1 + a7 - (a3 + (a3 >> 1));                               \
>> +    const int b7 = a3 - a5 + (a7 + (a7 >> 1));                               \
>> +    (dst)[dststride * 0] =  b0 +  b1;                                        \
>> +    (dst)[dststride * 1] =  b4 + (b7 >> 2);                                  \
>> +    (dst)[dststride * 2] =  b2 + (b3 >> 1);                                  \
>> +    (dst)[dststride * 3] =  b5 + (b6 >> 2);                                  \
>> +    (dst)[dststride * 4] =  b0  - b1;                                        \
>> +    (dst)[dststride * 5] =  b6 - (b5 >> 2);                                  \
>> +    (dst)[dststride * 6] = (b2 >> 1) - b3;                                   \
>> +    (dst)[dststride * 7] = (b4 >> 2) - b7;                                   \
>> +} while (0)
>> +
>> +#define dct8x8_impl(size, dctcoef)                                           \
>> +static void dct8x8_##size(dctcoef *coef)                                     \
>> +{                                                                            \
>> +    int i, x, y;                                                             \
>> +    dctcoef tmp[64];                                                         \
>> +    for (i = 0; i < 8; i++)                                                  \
>> +        DCT8_1D(coef + i, 8, tmp + i, 8);                                    \
>> +                                                                             \
>> +    for (i = 0; i < 8; i++)                                                  \
>> +        DCT8_1D(tmp + 8*i, 1, coef + i, 8);                                  \
>> +                                                                             \
>> +    for (y = 0; y < 8; y++) {                                                \
>> +        for (x = 0; x < 8; x++) {                                            \
>> +            static const int scale[] = {                                     \
>> +                13107 * 20, 11428 * 18, 20972 * 32,                          \
>> +                12222 * 19, 16777 * 25, 15481 * 24,                          \
>> +            };                                                               \
>> +            static const int idxmap[] = {                                    \
>> +                0, 3, 4, 3,                                                  \
>> +                3, 1, 5, 1,                                                  \
>> +                4, 5, 2, 5,                                                  \
>> +                3, 1, 5, 1,                                                  \
>> +            };                                                               \
>> +            const int idx = idxmap[(y & 3) * 4 + (x & 3)];                   \
>> +            coef[y*8 + x] = ((int64_t)coef[y*8 + x] *                        \
>> +                             scale[idx] + (1 << 17)) >> 18;                  \
>> +        }                                                                    \
>> +    }                                                                        \
>> +}
>> +
>> +dct4x4_impl(16, int16_t)
>> +dct4x4_impl(32, int32_t)
>> +
>> +dct8x8_impl(16, int16_t)
>> +dct8x8_impl(32, int32_t)
>> +
>> +static void dct4x4(int16_t *coef, int bit_depth)
>> +{
>> +    if (bit_depth == 8)
>> +        dct4x4_16(coef);
>> +    else
>> +        dct4x4_32((int32_t *) coef);
>> +}
>> +
>> +static void dct8x8(int16_t *coef, int bit_depth)
>> +{
>> +    if (bit_depth == 8) {
>> +        dct8x8_16(coef);
>> +    } else {
>> +        dct8x8_32((int32_t *) coef);
>> +    }
>> +}
>> +
>> +
>> +static void check_idct(void)
>> +{
>> +    LOCAL_ALIGNED_16(uint8_t, src,  [8 * (16 + 8 * 2)]);
>> +    LOCAL_ALIGNED_16(uint8_t, dst,  [8 * (16 + 8 * 2)]);
>> +    LOCAL_ALIGNED_16(uint8_t, dst0, [8 * (16 + 8 * 2)]);
>> +    LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * (16 + 8 * 2) + 16]);
>
> What's the reason for making the stride 32? 8 * 16 + 32 should be
> enough. + 16 for the varying alignment, + 16 padding for the altivec
> unaligned stores.

Oops, that's leftovers from testing strides 16 + sz * SIZEOF_PIXEL, will 
fix

>> +    LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
>> +    LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
>> +    LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
>> +    H264DSPContext h;
>> +    int bit_depth, sz, align;
>> +    int x, y, dc;
>> +    declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
>> +
>> +    for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
>> +        ff_h264dsp_init(&h, bit_depth, 1);
>> +        for (sz = 4; sz <= 8; sz += 4) {
>> +            randomize_buffers();
>> +
>> +            if (sz == 4)
>> +                dct4x4(coef, bit_depth);
>> +            else
>> +                dct8x8(coef, bit_depth);
>> +
>> +            for (dc = 0; dc <= 1; dc++) {
>> +                void (*idct)(uint8_t *, int16_t *, int) = NULL;
>> +                switch ((sz << 1) | dc) {
>> +                case (4 << 1) | 0: idct = h.h264_idct_add; break;
>> +                case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
>> +                case (8 << 1) | 0: idct = h.h264_idct8_add; break;
>> +                case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
>> +                }
>> +                if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
>> +                    for (align = 0; align < 16; align += sz * PIXEL_STRIDE) {
>
> s/PIXEL_STRIDE/PIXEL_SIZE/

Indeed, will fix

>> +                        uint8_t *dst1 = dst1_base + align;
>> +                        if (dc) {
>> +                            memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
>> +                            memcpy(subcoef0, coef, SIZEOF_COEF);
>> +                        } else {
>> +                            memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
>> +                        }
>> +                        memcpy(dst0, dst, sz * PIXEL_STRIDE);
>> +                        memcpy(dst1, dst, sz * PIXEL_STRIDE);
>> +                        memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
>> +                        call_ref(dst0, subcoef0, PIXEL_STRIDE);
>> +                        call_new(dst1, subcoef1, PIXEL_STRIDE);
>> +                        if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
>> +                            memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
>> +                            fail();
>> +                        bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
>> +                    }
>> +                }
>> +            }
>> +        }
>> +    }
>> +    report("idct");
>> +}
>> +
>> +void checkasm_check_h264dsp(void)
>> +{
>> +    check_idct();
>> +}
>> --
>
> otherwise ok

Ok, will fix, retest and push. Thanks for the reviews!

// Martin

Patch

diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile
index 8a0cee5..3c23853 100644
--- a/tests/checkasm/Makefile
+++ b/tests/checkasm/Makefile
@@ -2,6 +2,7 @@ 
 # subsystems
 AVCODECOBJS-$(CONFIG_BSWAPDSP)          += bswapdsp.o
 AVCODECOBJS-$(CONFIG_FMTCONVERT)        += fmtconvert.o
+AVCODECOBJS-$(CONFIG_H264DSP)           += h264dsp.o
 AVCODECOBJS-$(CONFIG_H264PRED)          += h264pred.o
 AVCODECOBJS-$(CONFIG_H264QPEL)          += h264qpel.o
 
diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
index c75e431..15f9f68 100644
--- a/tests/checkasm/checkasm.c
+++ b/tests/checkasm/checkasm.c
@@ -74,6 +74,9 @@  static const struct {
 #if CONFIG_FMTCONVERT
     { "fmtconvert", checkasm_check_fmtconvert },
 #endif
+#if CONFIG_H264DSP
+    { "h264dsp", checkasm_check_h264dsp },
+#endif
 #if CONFIG_H264PRED
     { "h264pred", checkasm_check_h264pred },
 #endif
diff --git a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h
index c7aa19c..619ebc7 100644
--- a/tests/checkasm/checkasm.h
+++ b/tests/checkasm/checkasm.h
@@ -33,6 +33,7 @@ 
 void checkasm_check_bswapdsp(void);
 void checkasm_check_dcadsp(void);
 void checkasm_check_fmtconvert(void);
+void checkasm_check_h264dsp(void);
 void checkasm_check_h264pred(void);
 void checkasm_check_h264qpel(void);
 void checkasm_check_hevc_mc(void);
diff --git a/tests/checkasm/h264dsp.c b/tests/checkasm/h264dsp.c
new file mode 100644
index 0000000..88beed0
--- /dev/null
+++ b/tests/checkasm/h264dsp.c
@@ -0,0 +1,232 @@ 
+/*
+ * Copyright (c) 2016 Martin Storsjo
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * Libav 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with Libav; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#include <string.h>
+#include "checkasm.h"
+#include "libavcodec/avcodec.h"
+#include "libavcodec/h264dsp.h"
+#include "libavutil/common.h"
+#include "libavutil/internal.h"
+#include "libavutil/intreadwrite.h"
+
+static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
+
+#define SIZEOF_PIXEL ((bit_depth + 7) / 8)
+#define SIZEOF_COEF  (2 * ((bit_depth + 7) / 8))
+#define PIXEL_STRIDE 16
+
+#define randomize_buffers()                                                  \
+    do {                                                                     \
+        uint32_t mask = pixel_mask[bit_depth - 8];                           \
+        for (y = 0; y < sz; y++) {                                           \
+            for (x = 0; x < sz * SIZEOF_PIXEL; x += 4) {                     \
+                AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask);          \
+                AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask);          \
+            }                                                                \
+            for (x = 0; x < sz; x++) {                                       \
+                if (bit_depth == 8) {                                        \
+                    coef[y * sz + x] = src[y * PIXEL_STRIDE + x] -           \
+                                       dst[y * PIXEL_STRIDE + x];            \
+                } else {                                                     \
+                    ((int32_t *)coef)[y * sz + x] =                          \
+                        ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] -        \
+                        ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x];         \
+                }                                                            \
+            }                                                                \
+        }                                                                    \
+    } while (0)
+
+#define dct4x4_impl(size, dctcoef)                                           \
+static void dct4x4_##size(dctcoef *coef)                                     \
+{                                                                            \
+    int i, y, x;                                                             \
+    dctcoef tmp[16];                                                         \
+    for (i = 0; i < 4; i++) {                                                \
+        const int z0 = coef[i*4 + 0] + coef[i*4 + 3];                        \
+        const int z1 = coef[i*4 + 1] + coef[i*4 + 2];                        \
+        const int z2 = coef[i*4 + 0] - coef[i*4 + 3];                        \
+        const int z3 = coef[i*4 + 1] - coef[i*4 + 2];                        \
+        tmp[i + 4*0] =   z0 +   z1;                                          \
+        tmp[i + 4*1] = 2*z2 +   z3;                                          \
+        tmp[i + 4*2] =   z0 -   z1;                                          \
+        tmp[i + 4*3] =   z2 - 2*z3;                                          \
+    }                                                                        \
+    for (i = 0; i < 4; i++) {                                                \
+        const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3];                          \
+        const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2];                          \
+        const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3];                          \
+        const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2];                          \
+        coef[i*4 + 0] =   z0 +   z1;                                         \
+        coef[i*4 + 1] = 2*z2 +   z3;                                         \
+        coef[i*4 + 2] =   z0 -   z1;                                         \
+        coef[i*4 + 3] =   z2 - 2*z3;                                         \
+    }                                                                        \
+    for (y = 0; y < 4; y++) {                                                \
+        for (x = 0; x < 4; x++) {                                            \
+            static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
+            const int idx = (y & 1) + (x & 1);                               \
+            coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15;  \
+        }                                                                    \
+    }                                                                        \
+}
+
+#define DCT8_1D(src, srcstride, dst, dststride) do {                         \
+    const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7];              \
+    const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7];              \
+    const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6];              \
+    const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6];              \
+    const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5];              \
+    const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5];              \
+    const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4];              \
+    const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4];              \
+    const int b0 = a0 + a6;                                                  \
+    const int b1 = a2 + a4;                                                  \
+    const int b2 = a0 - a6;                                                  \
+    const int b3 = a2 - a4;                                                  \
+    const int b4 = a3 + a5 + (a1 + (a1 >> 1));                               \
+    const int b5 = a1 - a7 - (a5 + (a5 >> 1));                               \
+    const int b6 = a1 + a7 - (a3 + (a3 >> 1));                               \
+    const int b7 = a3 - a5 + (a7 + (a7 >> 1));                               \
+    (dst)[dststride * 0] =  b0 +  b1;                                        \
+    (dst)[dststride * 1] =  b4 + (b7 >> 2);                                  \
+    (dst)[dststride * 2] =  b2 + (b3 >> 1);                                  \
+    (dst)[dststride * 3] =  b5 + (b6 >> 2);                                  \
+    (dst)[dststride * 4] =  b0  - b1;                                        \
+    (dst)[dststride * 5] =  b6 - (b5 >> 2);                                  \
+    (dst)[dststride * 6] = (b2 >> 1) - b3;                                   \
+    (dst)[dststride * 7] = (b4 >> 2) - b7;                                   \
+} while (0)
+
+#define dct8x8_impl(size, dctcoef)                                           \
+static void dct8x8_##size(dctcoef *coef)                                     \
+{                                                                            \
+    int i, x, y;                                                             \
+    dctcoef tmp[64];                                                         \
+    for (i = 0; i < 8; i++)                                                  \
+        DCT8_1D(coef + i, 8, tmp + i, 8);                                    \
+                                                                             \
+    for (i = 0; i < 8; i++)                                                  \
+        DCT8_1D(tmp + 8*i, 1, coef + i, 8);                                  \
+                                                                             \
+    for (y = 0; y < 8; y++) {                                                \
+        for (x = 0; x < 8; x++) {                                            \
+            static const int scale[] = {                                     \
+                13107 * 20, 11428 * 18, 20972 * 32,                          \
+                12222 * 19, 16777 * 25, 15481 * 24,                          \
+            };                                                               \
+            static const int idxmap[] = {                                    \
+                0, 3, 4, 3,                                                  \
+                3, 1, 5, 1,                                                  \
+                4, 5, 2, 5,                                                  \
+                3, 1, 5, 1,                                                  \
+            };                                                               \
+            const int idx = idxmap[(y & 3) * 4 + (x & 3)];                   \
+            coef[y*8 + x] = ((int64_t)coef[y*8 + x] *                        \
+                             scale[idx] + (1 << 17)) >> 18;                  \
+        }                                                                    \
+    }                                                                        \
+}
+
+dct4x4_impl(16, int16_t)
+dct4x4_impl(32, int32_t)
+
+dct8x8_impl(16, int16_t)
+dct8x8_impl(32, int32_t)
+
+static void dct4x4(int16_t *coef, int bit_depth)
+{
+    if (bit_depth == 8)
+        dct4x4_16(coef);
+    else
+        dct4x4_32((int32_t *) coef);
+}
+
+static void dct8x8(int16_t *coef, int bit_depth)
+{
+    if (bit_depth == 8) {
+        dct8x8_16(coef);
+    } else {
+        dct8x8_32((int32_t *) coef);
+    }
+}
+
+
+static void check_idct(void)
+{
+    LOCAL_ALIGNED_16(uint8_t, src,  [8 * (16 + 8 * 2)]);
+    LOCAL_ALIGNED_16(uint8_t, dst,  [8 * (16 + 8 * 2)]);
+    LOCAL_ALIGNED_16(uint8_t, dst0, [8 * (16 + 8 * 2)]);
+    LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * (16 + 8 * 2) + 16]);
+    LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
+    LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
+    LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
+    H264DSPContext h;
+    int bit_depth, sz, align;
+    int x, y, dc;
+    declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
+
+    for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
+        ff_h264dsp_init(&h, bit_depth, 1);
+        for (sz = 4; sz <= 8; sz += 4) {
+            randomize_buffers();
+
+            if (sz == 4)
+                dct4x4(coef, bit_depth);
+            else
+                dct8x8(coef, bit_depth);
+
+            for (dc = 0; dc <= 1; dc++) {
+                void (*idct)(uint8_t *, int16_t *, int) = NULL;
+                switch ((sz << 1) | dc) {
+                case (4 << 1) | 0: idct = h.h264_idct_add; break;
+                case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
+                case (8 << 1) | 0: idct = h.h264_idct8_add; break;
+                case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
+                }
+                if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
+                    for (align = 0; align < 16; align += sz * PIXEL_STRIDE) {
+                        uint8_t *dst1 = dst1_base + align;
+                        if (dc) {
+                            memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
+                            memcpy(subcoef0, coef, SIZEOF_COEF);
+                        } else {
+                            memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
+                        }
+                        memcpy(dst0, dst, sz * PIXEL_STRIDE);
+                        memcpy(dst1, dst, sz * PIXEL_STRIDE);
+                        memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
+                        call_ref(dst0, subcoef0, PIXEL_STRIDE);
+                        call_new(dst1, subcoef1, PIXEL_STRIDE);
+                        if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
+                            memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
+                            fail();
+                        bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
+                    }
+                }
+            }
+        }
+    }
+    report("idct");
+}
+
+void checkasm_check_h264dsp(void)
+{
+    check_idct();
+}