[FFmpeg-devel] [PATCH] avcodec/dxvenc: Use proper alignment, write endian-independent output

[FFmpeg-devel] [PATCH] avcodec/dxvenc: Use proper alignment, write endian-independent output

[Andreas Rheinhardt]

Fixes the dxv3enc-dxt1 FATE test with UBSan.

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
---
I have not actually tested whether the output is actually wrong
on BE systems. Would be nice if someone could.

 libavcodec/dxvenc.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/libavcodec/dxvenc.c b/libavcodec/dxvenc.c
index 1ce2b1d014..3f28fb696a 100644
--- a/libavcodec/dxvenc.c
+++ b/libavcodec/dxvenc.c
@@ -149,7 +149,7 @@ typedef struct DXVEncContext {
         } else {                                                              \
             op = 0;                                                           \
         }                                                                     \
-        *value |= (op << (state * 2));                                        \
+        AV_WL32(value, AV_RL32(value) | (op << (state * 2)));                 \
         state++;                                                              \
     } while (0)
 
@@ -157,7 +157,7 @@ static int dxv_compress_dxt1(AVCodecContext *avctx)
 {
     DXVEncContext *ctx = avctx->priv_data;
     PutByteContext *pbc = &ctx->pbc;
-    uint32_t *value;
+    void *value;
     uint32_t color, lut, idx, color_idx, lut_idx, prev_pos, state = 16, pos = 2, op = 0;
 
     ht_init(ctx->color_lookback_ht);
-- 
2.40.1

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Re: [FFmpeg-devel] [PATCH] avcodec/dxvenc: Use proper alignment, write endian-independent output

[James Almer]

On 2/26/2024 8:50 PM, Andreas Rheinhardt wrote:
> Fixes the dxv3enc-dxt1 FATE test with UBSan.
> 
> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
> ---
> I have not actually tested whether the output is actually wrong
> on BE systems. Would be nice if someone could.
> 
>   libavcodec/dxvenc.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/libavcodec/dxvenc.c b/libavcodec/dxvenc.c
> index 1ce2b1d014..3f28fb696a 100644
> --- a/libavcodec/dxvenc.c
> +++ b/libavcodec/dxvenc.c
> @@ -149,7 +149,7 @@ typedef struct DXVEncContext {
>           } else {                                                              \
>               op = 0;                                                           \
>           }                                                                     \
> -        *value |= (op << (state * 2));                                        \
> +        AV_WL32(value, AV_RL32(value) | (op << (state * 2)));                 \
>           state++;                                                              \
>       } while (0)
>   
> @@ -157,7 +157,7 @@ static int dxv_compress_dxt1(AVCodecContext *avctx)
>   {
>       DXVEncContext *ctx = avctx->priv_data;
>       PutByteContext *pbc = &ctx->pbc;
> -    uint32_t *value;
> +    void *value;

You should probably also remove the uint32_t* cast in the above macro.

>       uint32_t color, lut, idx, color_idx, lut_idx, prev_pos, state = 16, pos = 2, op = 0;
>   
>       ht_init(ctx->color_lookback_ht);
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Re: [FFmpeg-devel] [PATCH] avcodec/dxvenc: Use proper alignment, write endian-independent output

[Sean McGovern]

Hi,

On Mon, Feb 26, 2024 at 6:48 PM Andreas Rheinhardt
<andreas.rheinhardt@outlook.com> wrote:
>
> Fixes the dxv3enc-dxt1 FATE test with UBSan.
>
> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
> ---
> I have not actually tested whether the output is actually wrong
> on BE systems. Would be nice if someone could.
>
>  libavcodec/dxvenc.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/libavcodec/dxvenc.c b/libavcodec/dxvenc.c
> index 1ce2b1d014..3f28fb696a 100644
> --- a/libavcodec/dxvenc.c
> +++ b/libavcodec/dxvenc.c
> @@ -149,7 +149,7 @@ typedef struct DXVEncContext {
>          } else {                                                              \
>              op = 0;                                                           \
>          }                                                                     \
> -        *value |= (op << (state * 2));                                        \
> +        AV_WL32(value, AV_RL32(value) | (op << (state * 2)));                 \
>          state++;                                                              \
>      } while (0)
>
> @@ -157,7 +157,7 @@ static int dxv_compress_dxt1(AVCodecContext *avctx)
>  {
>      DXVEncContext *ctx = avctx->priv_data;
>      PutByteContext *pbc = &ctx->pbc;
> -    uint32_t *value;
> +    void *value;
>      uint32_t color, lut, idx, color_idx, lut_idx, prev_pos, state = 16, pos = 2, op = 0;
>
>      ht_init(ctx->color_lookback_ht);

Confirming that this does fix the failing test on a big-endian PowerPC
+ Altivec virtual machine.

Thanks,
Sean McGovern
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Re: [FFmpeg-devel] [PATCH] avcodec/dxvenc: Use proper alignment, write endian-independent output

[Andreas Rheinhardt]

Sean McGovern:
> Hi,
> 
> On Mon, Feb 26, 2024 at 6:48 PM Andreas Rheinhardt
> <andreas.rheinhardt@outlook.com> wrote:
>>
>> Fixes the dxv3enc-dxt1 FATE test with UBSan.
>>
>> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
>> ---
>> I have not actually tested whether the output is actually wrong
>> on BE systems. Would be nice if someone could.
>>
>>  libavcodec/dxvenc.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/libavcodec/dxvenc.c b/libavcodec/dxvenc.c
>> index 1ce2b1d014..3f28fb696a 100644
>> --- a/libavcodec/dxvenc.c
>> +++ b/libavcodec/dxvenc.c
>> @@ -149,7 +149,7 @@ typedef struct DXVEncContext {
>>          } else {                                                              \
>>              op = 0;                                                           \
>>          }                                                                     \
>> -        *value |= (op << (state * 2));                                        \
>> +        AV_WL32(value, AV_RL32(value) | (op << (state * 2)));                 \
>>          state++;                                                              \
>>      } while (0)
>>
>> @@ -157,7 +157,7 @@ static int dxv_compress_dxt1(AVCodecContext *avctx)
>>  {
>>      DXVEncContext *ctx = avctx->priv_data;
>>      PutByteContext *pbc = &ctx->pbc;
>> -    uint32_t *value;
>> +    void *value;
>>      uint32_t color, lut, idx, color_idx, lut_idx, prev_pos, state = 16, pos = 2, op = 0;
>>
>>      ht_init(ctx->color_lookback_ht);
> 
> Confirming that this does fix the failing test on a big-endian PowerPC
> + Altivec virtual machine.
> 

Thanks for testing, applied the patch with James' suggested modification.

- Andreas

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".