h264: Don't store intra pcm samples in h->mb

Instead, keep them in the bitstream buffer until we read them verbatim,
this saves a memcpy() and a subsequent clearing of the target buffer.
decode_cabac+decode_mb for a sample file (CAPM3_Sony_D.jsv) goes from
6121.4 to 6095.5 cycles, i.e. 26 cycles faster.

Signed-off-by: Martin Storsjö <martin@martin.st>

4 files changed, 21 insertions, 23 deletions

diff --git a/libavcodec/h264.h b/libavcodec/h264.h
index 45c134e877..426f8b3491 100644
--- a/libavcodec/h264.h
+++ b/libavcodec/h264.h
@@ -416,6 +416,7 @@ typedef struct H264Context {
     GetBitContext *intra_gb_ptr;
     GetBitContext *inter_gb_ptr;
 
+    const uint8_t *intra_pcm_ptr;
     DECLARE_ALIGNED(16, int16_t, mb)[16 * 48 * 2]; ///< as a dct coeffecient is int32_t in high depth, we need to reserve twice the space.
     DECLARE_ALIGNED(16, int16_t, mb_luma_dc)[3][16 * 2];
     int16_t mb_padding[256 * 2];        ///< as mb is addressed by scantable[i] and scantable is uint8_t we can either check that i is not too large or ensure that there is some unused stuff after mb
diff --git a/libavcodec/h264_cabac.c b/libavcodec/h264_cabac.c
index 8fc6b953cc..beb7d36b7c 100644
--- a/libavcodec/h264_cabac.c
+++ b/libavcodec/h264_cabac.c
@@ -2004,7 +2004,8 @@ decode_intra_mb:
         // The pixels are stored in the same order as levels in h->mb array.
         if ((int) (h->cabac.bytestream_end - ptr) < mb_size)
             return -1;
-        memcpy(h->mb, ptr, mb_size); ptr+=mb_size;
+        h->intra_pcm_ptr = ptr;
+        ptr += mb_size;
 
         ff_init_cabac_decoder(&h->cabac, ptr, h->cabac.bytestream_end - ptr);
 
diff --git a/libavcodec/h264_cavlc.c b/libavcodec/h264_cavlc.c
index 9585b77830..4b44cbefa9 100644
--- a/libavcodec/h264_cavlc.c
+++ b/libavcodec/h264_cavlc.c
@@ -761,17 +761,12 @@ decode_intra_mb:
     h->slice_table[ mb_xy ]= h->slice_num;
 
     if(IS_INTRA_PCM(mb_type)){
-        unsigned int x;
         const int mb_size = ff_h264_mb_sizes[h->sps.chroma_format_idc] *
-                            h->sps.bit_depth_luma >> 3;
+                            h->sps.bit_depth_luma;
 
         // We assume these blocks are very rare so we do not optimize it.
-        align_get_bits(&h->gb);
-
-        // The pixels are stored in the same order as levels in h->mb array.
-        for(x=0; x < mb_size; x++){
-            ((uint8_t*)h->mb)[x]= get_bits(&h->gb, 8);
-        }
+        h->intra_pcm_ptr = align_get_bits(&h->gb);
+        skip_bits_long(&h->gb, mb_size);
 
         // In deblocking, the quantizer is 0
         h->cur_pic.f.qscale_table[mb_xy] = 0;
diff --git a/libavcodec/h264_mb_template.c b/libavcodec/h264_mb_template.c
index fdefed4304..1c3b32db17 100644
--- a/libavcodec/h264_mb_template.c
+++ b/libavcodec/h264_mb_template.c
@@ -102,7 +102,7 @@ static av_noinline void FUNC(hl_decode_mb)(H264Context *h)
             const int bit_depth = h->sps.bit_depth_luma;
             int j;
             GetBitContext gb;
-            init_get_bits(&gb, (uint8_t *)h->mb,
+            init_get_bits(&gb, (uint8_t *)h->intra_pcm_ptr,
                           ff_h264_mb_sizes[h->sps.chroma_format_idc] * bit_depth);
 
             for (i = 0; i < 16; i++) {
@@ -137,7 +137,7 @@ static av_noinline void FUNC(hl_decode_mb)(H264Context *h)
             }
         } else {
             for (i = 0; i < 16; i++)
-                memcpy(dest_y + i * linesize, (uint8_t *)h->mb + i * 16, 16);
+                memcpy(dest_y + i * linesize, (uint8_t *)h->intra_pcm_ptr + i * 16, 16);
             if (SIMPLE || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
                 if (!h->sps.chroma_format_idc) {
                     for (i = 0; i < block_h; i++) {
@@ -145,8 +145,8 @@ static av_noinline void FUNC(hl_decode_mb)(H264Context *h)
                         memset(dest_cr + i * uvlinesize, 128, 8);
                     }
                 } else {
-                    uint8_t *src_cb = (uint8_t *)h->mb + 256;
-                    uint8_t *src_cr = (uint8_t *)h->mb + 256 + block_h * 8;
+                    uint8_t *src_cb = (uint8_t *)h->intra_pcm_ptr + 256;
+                    uint8_t *src_cr = (uint8_t *)h->intra_pcm_ptr + 256 + block_h * 8;
                     for (i = 0; i < block_h; i++) {
                         memcpy(dest_cb + i * uvlinesize, src_cb + i * 8, 8);
                         memcpy(dest_cr + i * uvlinesize, src_cr + i * 8, 8);
@@ -261,10 +261,10 @@ static av_noinline void FUNC(hl_decode_mb)(H264Context *h)
                 }
             }
         }
-    }
-    if (h->cbp || IS_INTRA(mb_type)) {
-        h->dsp.clear_blocks(h->mb);
-        h->dsp.clear_blocks(h->mb + (24 * 16 << PIXEL_SHIFT));
+        if (h->cbp || IS_INTRA(mb_type)) {
+            h->dsp.clear_blocks(h->mb);
+            h->dsp.clear_blocks(h->mb + (24 * 16 << PIXEL_SHIFT));
+        }
     }
 }
 
@@ -328,7 +328,7 @@ static av_noinline void FUNC(hl_decode_mb_444)(H264Context *h)
         if (PIXEL_SHIFT) {
             const int bit_depth = h->sps.bit_depth_luma;
             GetBitContext gb;
-            init_get_bits(&gb, (uint8_t *)h->mb, 768 * bit_depth);
+            init_get_bits(&gb, (uint8_t *)h->intra_pcm_ptr, 768 * bit_depth);
 
             for (p = 0; p < plane_count; p++)
                 for (i = 0; i < 16; i++) {
@@ -340,7 +340,7 @@ static av_noinline void FUNC(hl_decode_mb_444)(H264Context *h)
             for (p = 0; p < plane_count; p++)
                 for (i = 0; i < 16; i++)
                     memcpy(dest[p] + i * linesize,
-                           (uint8_t *)h->mb + p * 256 + i * 16, 16);
+                           (uint8_t *)h->intra_pcm_ptr + p * 256 + i * 16, 16);
         }
     } else {
         if (IS_INTRA(mb_type)) {
@@ -368,10 +368,11 @@ static av_noinline void FUNC(hl_decode_mb_444)(H264Context *h)
             hl_decode_mb_idct_luma(h, mb_type, 1, SIMPLE, transform_bypass,
                                    PIXEL_SHIFT, block_offset, linesize,
                                    dest[p], p);
-    }
-    if (h->cbp || IS_INTRA(mb_type)) {
-        h->dsp.clear_blocks(h->mb);
-        h->dsp.clear_blocks(h->mb + (24 * 16 << PIXEL_SHIFT));
+
+        if (h->cbp || IS_INTRA(mb_type)) {
+            h->dsp.clear_blocks(h->mb);
+            h->dsp.clear_blocks(h->mb + (24 * 16 << PIXEL_SHIFT));
+        }
     }
 }