[FFmpeg-devel] [PATCH 3/4] checkasm/h263dsp: test dct_unquantize_{intra, inter}

[Andreas Rheinhardt]

Rémi Denis-Courmont:
> ---
>  tests/checkasm/h263dsp.c | 57 +++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 56 insertions(+), 1 deletion(-)
> 
> diff --git a/tests/checkasm/h263dsp.c b/tests/checkasm/h263dsp.c
> index 2d0957a90b..26020211dc 100644
> --- a/tests/checkasm/h263dsp.c
> +++ b/tests/checkasm/h263dsp.c
> @@ -18,13 +18,65 @@
>   * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
>   */
>  
> +#include <stdbool.h>
>  #include <string.h>
>  
>  #include "checkasm.h"
>  
> -#include "libavcodec/h263dsp.h"
> +#include "libavutil/avassert.h"
>  #include "libavutil/mem.h"
>  #include "libavutil/mem_internal.h"
> +#include "libavcodec/h263dsp.h"
> +#include "libavcodec/mpegvideodata.h"
> +
> +static uint_fast8_t mpeg_qscale_rnd(void)
> +{
> +    int n = rnd(), q = (n >> 1) & 31;
> +
> +    if (n & 1)
> +        return ff_mpeg2_non_linear_qscale[q];
> +    else
> +        return q << 1;

This branch reminds me of the q_scale_type check from the
mpeg2-unquantize functions. Why is it here?
(For H.263, qscale is simply a value in 1..31. This code will test
values outside of this range.)

> +}
> +
> +typedef void (*unquantizer)(int16_t *, ptrdiff_t, int, int);
> +
> +static void check_dct_unquantize(unquantizer func, bool inter)
> +{
> +    const char *name = inter ? "inter" : "intra";
> +#define LEN 64
> +    LOCAL_ALIGNED_16(int16_t, block0, [LEN]);
> +    LOCAL_ALIGNED_16(int16_t, block1, [LEN]);
> +    size_t len = rnd() % LEN;
> +    const int qscale = mpeg_qscale_rnd();
> +    const int qmul = qscale << 1;
> +    const int qadd = (rnd() & 1) ? (qscale - 1) | 1 : 0;
> +
> +    declare_func(void, int16_t *, ptrdiff_t, int, int);
> +
> +    for (size_t i = 0; i < LEN; i++) {
> +        int r = rnd();
> +
> +        block1[i] = block0[i] = (r & 1) ? (r >> 1) : 0;

This will potentially set all elements to some nonzero value. In
reality, all elements not processed in the (current) C version are zero;
in other words, one could add

    for (; i <= 63; i++)
        av_assert0(!block[i]);

at the end of dct_unquantize_h263_(inter|intra)_c. This is what makes it
possible to switch to a do-while loop. It would also make it possible to
unroll the loop/omit special handling for the tail when vectorizing, so
this test is a hindrance to potential optimizations.

> +    }
> +
> +    if (check_func(func, "h263dsp.dct_unquantize_%s", name)) {
> +        call_ref(block0, 0, qmul, qadd);
> +        call_new(block1, 0, qmul, qadd);
> +
> +        if (memcmp(block0, block1, LEN * sizeof (int16_t)))
> +            fail();
> +
> +        av_assert0(len < LEN);
> +        call_ref(block0, len, qmul, qadd);
> +        call_new(block1, len, qmul, qadd);
> +
> +        if (memcmp(block0, block1, LEN * sizeof (int16_t)))
> +            fail();
> +
> +        bench_new(block1, LEN, qmul, qadd);
> +    }
> +}
>  
>  typedef void (*filter)(uint8_t *src, int stride, int qscale);
>  
> @@ -56,6 +108,9 @@ void checkasm_check_h263dsp(void)
>      H263DSPContext ctx;
>  
>      ff_h263dsp_init(&ctx);
> +    check_dct_unquantize(ctx.h263_dct_unquantize_intra, false);
> +    check_dct_unquantize(ctx.h263_dct_unquantize_inter, true);
> +    report("dct_unquantize");
>      check_loop_filter('h', ctx.h263_h_loop_filter);
>      check_loop_filter('v', ctx.h263_v_loop_filter);
>      report("loop_filter");