[FFmpeg-devel] [PATCH] checkasm: add aacencdsp.quant_bands test

Rémi Denis-Courmont remi at remlab.net
Sat Jun 1 09:51:29 EEST 2024



Le 1 juin 2024 04:03:37 GMT+03:00, James Almer <jamrial at gmail.com> a écrit :
>On 5/31/2024 3:18 PM, Rémi Denis-Courmont wrote:
>> ---
>>   tests/checkasm/aacencdsp.c | 37 +++++++++++++++++++++++++++++++++++++
>>   1 file changed, 37 insertions(+)
>> 
>> diff --git a/tests/checkasm/aacencdsp.c b/tests/checkasm/aacencdsp.c
>> index 1756c4ecd5..756f92fd8f 100644
>> --- a/tests/checkasm/aacencdsp.c
>> +++ b/tests/checkasm/aacencdsp.c
>> @@ -22,7 +22,9 @@
>>     #include "libavutil/mem_internal.h"
>>   +#include "libavcodec/aacenc_utils.h"
>>   #include "libavcodec/aacencdsp.h"
>> +#include "libavcodec/aactab.h"
>>     #include "checkasm.h"
>>   @@ -35,6 +37,8 @@
>>           }                                                       \
>>       } while (0)
>>   +#define randomize_elem(tab) (tab[rnd() % FF_ARRAY_ELEMS(tab)])
>> +
>>   static void test_abs_pow34(AACEncDSPContext *s)
>>   {
>>   #define BUF_SIZE 1024
>> @@ -60,6 +64,38 @@ static void test_abs_pow34(AACEncDSPContext *s)
>>       report("abs_pow34");
>>   }
>>   +static void test_quant_bands(AACEncDSPContext *s)
>> +{
>> +    int maxval = randomize_elem(aac_cb_maxval);
>> +    float q34 = randomize_elem(ff_aac_pow34sf_tab);
>> +    float rounding = (rnd() & 1) ? ROUND_TO_ZERO : ROUND_STANDARD;
>> +    LOCAL_ALIGNED_32(float, in, [BUF_SIZE]);
>> +    LOCAL_ALIGNED_32(float, scaled, [BUF_SIZE]);
>> +
>> +    declare_func(void, int *, const float *, const float *, int, int, int,
>> +                 const float, const float);
>> +
>> +    randomize_float(in, BUF_SIZE);
>> +    randomize_float(scaled, BUF_SIZE);
>> +
>> +    for (int sign = 0; sign <= 1; sign++) {
>> +        if (check_func(s->quant_bands, "quant_bands_%s",
>> +                       sign ? "signed" : "unsigned")) {
>> +            LOCAL_ALIGNED_32(int, out, [BUF_SIZE]);
>> +            LOCAL_ALIGNED_32(int, out2, [BUF_SIZE]);
>
>Does your RVV implementation expect out to be 32 byte aligned? Because looking at aaccoder.c, quant_bands() is called with AACEncContext.qcoefs as argument for out, which is 16 byte aligned.

The RVV code requires 32-*bit* alignment for all pointers, i.e. natural alignment for int and float. I will fix the test case.

(IMO 32 byte alignment is counter-productive on non-x86, but that's a more general problem.)


More information about the ffmpeg-devel mailing list