[FFmpeg-devel] [PATCH]doc/platform: Mention musl where x86_32 is not supported

[Carl Eugen Hoyos]

2016-10-03 13:57 GMT+02:00 Hendrik Leppkes <h.leppkes at gmail.com>:

> The underlying problem is that mmx code is mixed with allocations,

Definitely.

> which seems like an unusual case to begin with

I am not sure if I understand this but one instance is calling radix_sort()
in the dnxhd encoder.

Below is what is needed to run the dnxhd fate tests here (the hr hq test
crashes), the second hunk is needed to fix thread joining and avoid an
endless loop. There are also issues with h26x.

You may be interested in this post that contains a link to a work-around
in musl: http://www.openwall.com/lists/musl/2016/09/14/8

No more comments from me, Carl Eugen

diff --git a/libavcodec/dnxhdenc.c b/libavcodec/dnxhdenc.c
index 88edd6b..5b4a36a 100644
--- a/libavcodec/dnxhdenc.c
+++ b/libavcodec/dnxhdenc.c
@@ -1117,6 +1117,7 @@ static int dnxhd_encode_fast(AVCodecContext
*avctx, DNXHDEncContext *ctx)
         if (RC_VARIANCE)
             avctx->execute2(avctx, dnxhd_mb_var_thread,
                             NULL, NULL, ctx->m.mb_height);
+        emms_c();
         radix_sort(ctx->mb_cmp, ctx->m.mb_num);
         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
             int mb = ctx->mb_cmp[x].mb;
@@ -1214,6 +1215,7 @@ FF_ENABLE_DEPRECATION_WARNINGS

     ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA,
NULL, 0, AV_PICTURE_TYPE_I);

+    emms_c();
     pkt->flags |= AV_PKT_FLAG_KEY;
     *got_packet = 1;
     return 0;