[FFmpeg-devel] [PATCH 10/10] lavc/flacenc: partially unroll loop in flac_enc_lpc_32

[James Darnley]

Now does 6 samples per iteration, up from 2.

>From 1.6 to 2.1 times faster again.  2.5 to 3.9 times faster overall.
Runtime is reduced by a further 4 to 17%.  Reduced by 9 to 65% overall.

Same conditions as previously.
---
 libavcodec/x86/flac_dsp_gpl.asm |   30 +++++++++++++++++++++++++++---
 1 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/libavcodec/x86/flac_dsp_gpl.asm b/libavcodec/x86/flac_dsp_gpl.asm
index e36c76b..704e461 100644
--- a/libavcodec/x86/flac_dsp_gpl.asm
+++ b/libavcodec/x86/flac_dsp_gpl.asm
@@ -157,15 +157,25 @@ mova [rsp], m4 ; sign extend mask
 
 .looplen:
     pxor m0,   m0
+    pxor m4,   m4
+    pxor m6,   m6
     xor  posj,  posj
     xor  negj,  negj
     .looporder:
         movd   m2,  [coefsq+posj*4] ; c = coefs[j]
         SPLATD m2
         movh   m1,  [smpq+negj*4-4] ; s = smp[i-j-1]
+        movh   m5,  [smpq+negj*4-4+mmsize/2]
+        movh   m7,  [smpq+negj*4-4+mmsize]
         pshufd m1,   m1, q3130
+        pshufd m5,   m5, q3130
+        pshufd m7,   m7, q3130
         pmuldq m1,   m2
+        pmuldq m5,   m2
+        pmuldq m7,   m2
         paddq  m0,   m1             ; p += c * s
+        paddq  m4,   m5
+        paddq  m6,   m7
 
         add    posj, 1
         sub    negj, 1
@@ -175,14 +185,28 @@ mova [rsp], m4 ; sign extend mask
 
     pxor m2, m2
     HACK_PSRAQ m0, m3, [rsp], m2
+    pxor m2, m2
+    HACK_PSRAQ m4, m3, [rsp], m2
+    pxor m2, m2
+    HACK_PSRAQ m6, m3, [rsp], m2
     CLIPQ   m0, [pq_int_min], [pq_int_max], m2
+    CLIPQ   m4, [pq_int_min], [pq_int_max], m2
+    CLIPQ   m6, [pq_int_min], [pq_int_max], m2
     pshufd  m0,    m0, q0020 ; pack into first 2 dwords
+    pshufd  m4,    m4, q0020
+    pshufd  m6,    m6, q0020
     movh    m1,   [smpq]
+    movh    m5,   [smpq+mmsize/2]
+    movh    m7,   [smpq+mmsize]
     psubd   m1,    m0                ; smp[i] - p
+    psubd   m5,    m4
+    psubd   m7,    m6
     movh   [resq], m1                ; res[i] = smp[i] - (p >> shift)
+    movh   [resq+mmsize/2], m5
+    movh   [resq+mmsize], m7
 
-    add    resq,  mmsize/2
-    add    smpq,  mmsize/2
-    sub    lenmp, mmsize/8
+    add    resq,  (3*mmsize)/2
+    add    smpq,  (3*mmsize)/2
+    sub    lenmp, (3*mmsize)/8
 jg .looplen
 RET
-- 
1.7.9