[FFmpeg-devel] [PATCH 2/3] checkasm/blockdsp: use smallest allowed aligned buffers for fill_block_tab tests

[James Almer]

On 5/7/2024 12:14 PM, Andreas Rheinhardt wrote:
> James Almer:
>> The requirement is either 8 or 16 bytes alignment, not 32.
>> This should help finding bugs in asm implementations.
>>
>> Signed-off-by: James Almer <jamrial at gmail.com>
>> ---
>>   tests/checkasm/blockdsp.c | 23 +++++++++--------------
>>   1 file changed, 9 insertions(+), 14 deletions(-)
>>
>> diff --git a/tests/checkasm/blockdsp.c b/tests/checkasm/blockdsp.c
>> index ab87fc8fa4..f67a38d302 100644
>> --- a/tests/checkasm/blockdsp.c
>> +++ b/tests/checkasm/blockdsp.c
>> @@ -29,11 +29,6 @@
>>   #include "libavutil/intreadwrite.h"
>>   #include "libavutil/mem_internal.h"
>>   
>> -typedef struct {
>> -    const char *name;
>> -    int size;
>> -} test;
>> -
>>   #define randomize_buffers(size)             \
>>       do {                                    \
>>           int i;                              \
>> @@ -58,18 +53,18 @@ do {                                                                \
>>   } while (0)
>>   
>>   static void check_fill(BlockDSPContext *h){
>> -    const test tests[] = {
>> -        {"fill_block_tab[0]", 16},
>> -        {"fill_block_tab[1]", 8},
>> -    };
>> -    LOCAL_ALIGNED_32(uint8_t, buf0, [16 * 16]);
>> -    LOCAL_ALIGNED_32(uint8_t, buf1, [16 * 16]);
>> +    LOCAL_ALIGNED_16(uint8_t, buf0_16, [16 * 16]);
>> +    LOCAL_ALIGNED_16(uint8_t, buf1_16, [16 * 16]);
>> +    LOCAL_ALIGNED_8(uint8_t, buf0_8, [8 * 8]);
>> +    LOCAL_ALIGNED_8(uint8_t, buf1_8, [8 * 8]);
>>   
>> -    for (size_t t = 0; t < FF_ARRAY_ELEMS(tests); ++t) {
>> -        int n = tests[t].size;
>> +    for (int t = 0; t < 2; ++t) {
>> +        uint8_t *buf0 = t ? buf0_8 : buf0_16;
>> +        uint8_t *buf1 = t ? buf1_8 : buf1_16;
>> +        int n = 16 - 8 * t;
>>           declare_func(void, uint8_t *block, uint8_t value,
>>                        ptrdiff_t line_size, int h);
>> -        if (check_func(h->fill_block_tab[t], "blockdsp.%s", tests[t].name)) {
>> +        if (check_func(h->fill_block_tab[t], "blockdsp.fill_block_tab[%d]", t)) {
>>               uint8_t value = rnd();
>>               memset(buf0, 0, sizeof(*buf0) * n * n);
>>               memset(buf1, 0, sizeof(*buf0) * n * n);
> 
> 1. I wouldn't be surprised if the *_8 buffers were still 16 byte
> aligned. You should probably force 8 byte alignment by using a 16
> byte-aligned buffer with an offset of eight.

Amended the following locally:

> diff --git a/tests/checkasm/blockdsp.c b/tests/checkasm/blockdsp.c
> index 8c1f8281d2..5f4d46b8fa 100644
> --- a/tests/checkasm/blockdsp.c
> +++ b/tests/checkasm/blockdsp.c
> @@ -55,12 +55,10 @@ do {                                                                \
>  static void check_fill(BlockDSPContext *h){
>      LOCAL_ALIGNED_16(uint8_t, buf0_16, [16 * 16]);
>      LOCAL_ALIGNED_16(uint8_t, buf1_16, [16 * 16]);
> -    LOCAL_ALIGNED_8(uint8_t, buf0_8, [8 * 8]);
> -    LOCAL_ALIGNED_8(uint8_t, buf1_8, [8 * 8]);
> 
>      for (int t = 0; t < 2; ++t) {
> -        uint8_t *buf0 = t ? buf0_8 : buf0_16;
> -        uint8_t *buf1 = t ? buf1_8 : buf1_16;
> +        uint8_t *buf0 = buf0_16 + t * /* force 8 byte alignment */ 8;
> +        uint8_t *buf1 = buf1_16 + t * /* force 8 byte alignment */ 8;
>          int n = 16 - 8 * t;
>          declare_func(void, uint8_t *block, uint8_t value,
>                       ptrdiff_t line_size, int h);

> 2. Can you also extend this test to actually test the case of stride !=
> width? (And negative strides.)

Maybe later.