From: Tucker DiNapoli
Also pulled QP initialization out of inner loop, which removed some redundent
code.
Added some dummy fields to PPContext to allow current code to work while
changing the rest of the postprocessing code to support the arrays.
I also increased alignment requirements for some fields in the PPContext struct
to
support future avx2 code.
---
libpostproc/postprocess_internal.h | 10 -
libpostproc/postprocess_template.c | 81 ++
2 files changed, 46 insertions(+), 45 deletions(-)
diff --git a/libpostproc/postprocess_internal.h
b/libpostproc/postprocess_internal.h
index 1ebd974..c1a306d 100644
--- a/libpostproc/postprocess_internal.h
+++ b/libpostproc/postprocess_internal.h
@@ -143,8 +143,11 @@ typedef struct PPContext{
DECLARE_ALIGNED(8, uint64_t, pQPb);
DECLARE_ALIGNED(8, uint64_t, pQPb2);
-DECLARE_ALIGNED(8, uint64_t, mmxDcOffset)[64];
-DECLARE_ALIGNED(8, uint64_t, mmxDcThreshold)[64];
+DECLARE_ALIGNED(32, uint64_t, pQPb_block)[4];
+DECLARE_ALIGNED(32, uint64_t, pQPb2_block)[4];
+
+DECLARE_ALIGNED(32, uint64_t, mmxDcOffset)[64];
+DECLARE_ALIGNED(32, uint64_t, mmxDcThreshold)[64];
QP_STORE_T *stdQPTable; ///< used to fix MPEG2 style qscale
QP_STORE_T *nonBQPTable;
@@ -153,6 +156,9 @@ typedef struct PPContext{
int QP;
int nonBQP;
+DECLARE_ALIGNED(32, int, QP_block)[4];
+DECLARE_ALIGNED(32, int, nonBQP_block)[4];
+
int frameNum;
int cpuCaps;
diff --git a/libpostproc/postprocess_template.c
b/libpostproc/postprocess_template.c
index e153b13..b7296c4 100644
--- a/libpostproc/postprocess_template.c
+++ b/libpostproc/postprocess_template.c
@@ -3479,7 +3479,7 @@ static void RENAME(postProcess)(const uint8_t src[], int
srcStride, uint8_t dst[
#endif
const int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)];
-int QP=0;
+int QP=0, nonBQP=0;
/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and
1 line upwards
if not than use a temporary buffer */
if(y+15 >= height){
@@ -3512,6 +3512,29 @@ static void RENAME(postProcess)(const uint8_t src[], int
srcStride, uint8_t dst[
int endx = FFMIN(width, x+32);
uint8_t *dstBlockStart = dstBlock;
const uint8_t *srcBlockStart = srcBlock;
+int qp_index = 0;
+for(qp_index=0; qp_index < (endx-startx)/BLOCK_SIZE; qp_index++){
+QP = QPptr[(x+qp_index*BLOCK_SIZE)>>qpHShift];
+nonBQP = nonBQPptr[(x+qp_index*BLOCK_SIZE)>>qpHShift];
+if(!isColor){
+QP= (QP* QPCorrecture + 256*128)>>16;
+nonBQP= (nonBQP* QPCorrecture + 256*128)>>16;
+yHistogram[(srcBlock+qp_index*8)[srcStride*12 + 4]]++;
+}
+c.QP_block[qp_index] = QP;
+c.nonBQP_block[qp_index] = nonBQP;
+#if TEMPLATE_PP_MMX
+__asm__ volatile(
+"movd %1, %%mm7 \n\t"
+"packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
+"packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
+"packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
+"movq %%mm7, %0 \n\t"
+: "=m" (c.pQPb_block[qp_index])
+: "r" (QP)
+);
+#endif
+}
for(; x < endx; x+=BLOCK_SIZE){
RENAME(prefetchnta)(srcBlock + (((x>>2)&6) + copyAhead)*srcStride
+ 32);
RENAME(prefetchnta)(srcBlock + (((x>>2)&6) +
copyAhead+1)*srcStride + 32);
@@ -3543,27 +3566,15 @@ static void RENAME(postProcess)(const uint8_t src[],
int srcStride, uint8_t dst[
dstBlock = dstBlockStart;
srcBlock = srcBlockStart;
- for(x = startx; x < endx; x+=BLOCK_SIZE){
+ for(x = startx, qp_index = 0; x < endx; x+=BLOCK_SIZE, qp_index++){
const int stride= dstStride;
-QP = QPptr[x>>qpHShift];
-c.nonBQP = nonBQPptr[x>>qpHShift];
-if(!isColor){
-QP= (QP* QPCorrecture + 256*128)>>16;
-c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
-yHistogram[srcBlock[srcStride*12 + 4]]++;
-}
-c.QP= QP;
-#if TEMPLATE_PP_MMX
-__asm__ volatile(
-"movd %1, %%mm7 \n\t"
-"packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
-"packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
-"packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
-"movq %%mm7, %0 \n\t"
-: "=m" (c.pQPb)
-: "r" (QP)
-);
-#endif
+//temporary while changing QP stuff to make things continue to work
+//eventually QP,nonBQP,etc will be arrays and this will be
unnecessary
+c.Q
