Hi, 2014-08-12 23:22 GMT+02:00 James Darnley <james.darnley@gmail.com>:
--- libavcodec/flacdsp.h | 7 +++++++ 1 files changed, 7 insertions(+), 0 deletions(-)
diff --git a/libavcodec/flacdsp.h b/libavcodec/flacdsp.h index 272cf2a..36cd904 100644 --- a/libavcodec/flacdsp.h +++ b/libavcodec/flacdsp.h @@ -27,6 +27,13 @@ typedef struct FLACDSPContext { int len, int shift); void (*lpc)(int32_t *samples, const int coeffs[32], int order, int qlevel, int len); + /** + * This function has some limitations with various configurations: + * - when CONFIG_SMALL is 0 there is an unrolled loop which assumes the + * maximum value of order is 32. + * - when SSE4 (or newer) is available on x86 there is an unrolled copy + * which also assumes the maximum value of order is 0. + */ void (*lpc_encode)(int32_t *res, const int32_t *smp, int len, int order, const int32_t *coefs, int shift); } FLACDSPContext;
Some comments: - The decoder passes int coeffs[32]; I don't see how order > 32 would then work - Why not do the same for lpc_encode, and document that maximum order value is 32 - The decoder actually reads the order (see decode_subframe and flacdec.c:366) on 6 bits (there's a type >= 32 check though but I don't see what I'm missing) - The encoder has min and max bounds in (min,max)_prediction_order in arrays (see flacenc.c:312) that are below 32 so the encoder should never use an order above 32 (as seen with again the coding on 6bits on 1126) I guess some of your comments in the asm and here could be dropped then ? -- Christophe