Thanks for feedback!
On Thu, Nov 14, 2024 at 9:40 PM Ronald S. Bultje <rsbul...@gmail.com> wrote:
> In terms of binary size, the above will triplicate nearly identical code in
> the binary, and that doesn't really seem necessary. You should be able to
> write the above out as a function and call it with appropriate xblen and
> func_index arguments (check_add_obmc(0, 8); check_add_obmc(1, 16);
> check_add_obmc(2, 32)) without duplicating these in the binary. Done for many
> tests, this reduces binary size and compile time significantly.
Sounds good. I'm going to implement these since I have some free time now.
> You can also get bonus points if you randomize yblen, since the bitstream
> allows multiple values for this (at least 4, 12, 16, 24, but possibly any
> number), so maybe 1 + (random % 64) would be a good choice?
I'll try this one too.
Thanks.
Kyosuke
On Thu, Nov 14, 2024 at 9:40 PM Ronald S. Bultje <rsbul...@gmail.com> wrote:
>
> Hi,
>
> thanks for adding the test! This looks pretty good. Minor suggestion:
>
> On Wed, Nov 13, 2024 at 5:39 PM Kyosuke Kawakami <kawakami150...@gmail.com>
> wrote:
>>
>> +#define RANDOMIZE_DESTS(name, size) \
>> + do { \
>> + int i; \
>> + for (i = 0; i < size; ++i) { \
>> + uint16_t r = rnd(); \
>> + AV_WN16A(name##0 + i, r); \
>> + AV_WN16A(name##1 + i, r); \
>> + } \
>> + } while (0)
>> +
>> +#define RANDOMIZE_BUFFER8(name, size) \
>> + do { \
>> + int i; \
>> + for (i = 0; i < size; ++i) { \
>> + uint8_t r = rnd(); \
>> + name[i] = r; \
>> + } \
>> + } while (0)
>> +
>> +#define OBMC_STRIDE 32
>> +
>> +#define CHECK_ADD_OBMC(func_index, yblen, xblen)
>> \
>> + static void check_add_obmc ## xblen(void)
>> \
>> + {
>> \
>> + LOCAL_ALIGNED_8(uint8_t, src, [yblen * xblen]);
>> \
>> + LOCAL_ALIGNED_16(uint16_t, dst0, [yblen * xblen]);
>> \
>> + LOCAL_ALIGNED_16(uint16_t, dst1, [yblen * xblen]);
>> \
>> + LOCAL_ALIGNED_8(uint8_t, obmc_weight, [yblen * OBMC_STRIDE]);
>> \
>> + DiracDSPContext h;
>> \
>> + ff_diracdsp_init(&h);
>> \
>> + if (check_func(h.add_dirac_obmc[func_index],
>> "diracdsp.add_dirac_obmc_%d", xblen)) { \
>> + declare_func(void, uint16_t*, const uint8_t*, int, const
>> uint8_t *, int); \
>> + RANDOMIZE_BUFFER8(src, yblen * xblen);
>> \
>> + RANDOMIZE_DESTS(dst, yblen * xblen);
>> \
>> + RANDOMIZE_BUFFER8(obmc_weight, yblen * OBMC_STRIDE);
>> \
>> + call_ref(dst0, src, xblen, obmc_weight, yblen);
>> \
>> + call_new(dst1, src, xblen, obmc_weight, yblen);
>> \
>> + if (memcmp(dst0, dst1, yblen * xblen))
>> \
>> + fail();
>> \
>> + bench_new(dst1, src, xblen, obmc_weight, yblen);
>> \
>> + }
>> \
>> + }
>> \
>> +
>> +CHECK_ADD_OBMC(0, 64, 8)
>> +CHECK_ADD_OBMC(1, 64, 16)
>> +CHECK_ADD_OBMC(2, 64, 32)
>> +
>> +void checkasm_check_diracdsp(void)
>> +{
>> + check_add_obmc8();
>> + check_add_obmc16();
>> + check_add_obmc32();
>> + report("diracdsp");
>> +}
>
>
> In terms of binary size, the above will triplicate nearly identical code in
> the binary, and that doesn't really seem necessary. You should be able to
> write the above out as a function and call it with appropriate xblen and
> func_index arguments (check_add_obmc(0, 8); check_add_obmc(1, 16);
> check_add_obmc(2, 32)) without duplicating these in the binary. Done for many
> tests, this reduces binary size and compile time significantly.
>
> You can also get bonus points if you randomize yblen, since the bitstream
> allows multiple values for this (at least 4, 12, 16, 24, but possibly any
> number), so maybe 1 + (random % 64) would be a good choice?
>
> If you don't care for the above, we can merge this as-is, they're merely
> suggestions, let me know what you think.
>
> Ronald
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
