Merge remote branch 'internal/upstream' into HEAD

Conflicts:
	vp8/vp8_cx_iface.c

Change-Id: Iecfd4532ab1c722d10ecce8a5ec473e96093cf3b

8 files changed, 117 insertions, 53 deletions

diff --git a/vp8/encoder/arm/arm_csystemdependent.c b/vp8/encoder/arm/arm_csystemdependent.c
index 5852afddb..73007d414 100644
--- a/vp8/encoder/arm/arm_csystemdependent.c
+++ b/vp8/encoder/arm/arm_csystemdependent.c
@@ -46,6 +46,9 @@ void vp8_arch_arm_encoder_init(VP8_COMP *cpi)
         cpi->rtcd.variance.subpixvar8x16         = vp8_sub_pixel_variance8x16_c;
         cpi->rtcd.variance.subpixvar16x8         = vp8_sub_pixel_variance16x8_c;*/
         cpi->rtcd.variance.subpixvar16x16        = vp8_sub_pixel_variance16x16_armv6;
+        cpi->rtcd.variance.halfpixvar16x16_h     = vp8_variance_halfpixvar16x16_h_armv6;
+        cpi->rtcd.variance.halfpixvar16x16_v     = vp8_variance_halfpixvar16x16_v_armv6;
+        cpi->rtcd.variance.halfpixvar16x16_hv    = vp8_variance_halfpixvar16x16_hv_armv6;
 
         /*cpi->rtcd.variance.mse16x16              = vp8_mse16x16_c;
         cpi->rtcd.variance.getmbss               = vp8_get_mb_ss_c;*/
diff --git a/vp8/encoder/arm/variance_arm.c b/vp8/encoder/arm/variance_arm.c
index 1cf9fd8f0..64d76bcf8 100644
--- a/vp8/encoder/arm/variance_arm.c
+++ b/vp8/encoder/arm/variance_arm.c
@@ -43,7 +43,41 @@ unsigned int vp8_sub_pixel_variance16x16_armv6
                                    dst_pixels_per_line, sse);
 }
 
-#endif
+unsigned int vp8_variance_halfpixvar16x16_h_armv6(
+    const unsigned char *src_ptr,
+    int  source_stride,
+    const unsigned char *ref_ptr,
+    int  recon_stride,
+    unsigned int *sse)
+{
+    return vp8_sub_pixel_variance16x16_armv6(src_ptr, source_stride, 4, 0,
+                                         ref_ptr, recon_stride, sse);
+}
+
+unsigned int vp8_variance_halfpixvar16x16_v_armv6(
+    const unsigned char *src_ptr,
+    int  source_stride,
+    const unsigned char *ref_ptr,
+    int  recon_stride,
+    unsigned int *sse)
+{
+    return vp8_sub_pixel_variance16x16_armv6(src_ptr, source_stride, 0, 4,
+                                         ref_ptr, recon_stride, sse);
+}
+
+unsigned int vp8_variance_halfpixvar16x16_hv_armv6(
+    const unsigned char *src_ptr,
+    int  source_stride,
+    const unsigned char *ref_ptr,
+    int  recon_stride,
+    unsigned int *sse)
+{
+    return vp8_sub_pixel_variance16x16_armv6(src_ptr, source_stride, 4, 4,
+                                         ref_ptr, recon_stride, sse);
+}
+
+#endif /* HAVE_ARMV6 */
+
 
 #if HAVE_ARMV7
 
diff --git a/vp8/encoder/arm/variance_arm.h b/vp8/encoder/arm/variance_arm.h
index c807e29c0..7ac0ac08e 100644
--- a/vp8/encoder/arm/variance_arm.h
+++ b/vp8/encoder/arm/variance_arm.h
@@ -17,6 +17,9 @@
 extern prototype_sad(vp8_sad16x16_armv6);
 extern prototype_variance(vp8_variance16x16_armv6);
 extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_armv6);
+extern prototype_variance(vp8_variance_halfpixvar16x16_h_armv6);
+extern prototype_variance(vp8_variance_halfpixvar16x16_v_armv6);
+extern prototype_variance(vp8_variance_halfpixvar16x16_hv_armv6);
 
 #if !CONFIG_RUNTIME_CPU_DETECT
 
@@ -29,10 +32,20 @@ extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_armv6);
 #undef  vp8_variance_var16x16
 #define vp8_variance_var16x16 vp8_variance16x16_armv6
 
+#undef  vp8_variance_halfpixvar16x16_h
+#define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_armv6
+
+#undef  vp8_variance_halfpixvar16x16_v
+#define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_armv6
+
+#undef  vp8_variance_halfpixvar16x16_hv
+#define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_armv6
+
 #endif /* !CONFIG_RUNTIME_CPU_DETECT */
 
 #endif /* HAVE_ARMV6 */
 
+
 #if HAVE_ARMV7
 extern prototype_sad(vp8_sad4x4_neon);
 extern prototype_sad(vp8_sad8x8_neon);
diff --git a/vp8/encoder/generic/csystemdependent.c b/vp8/encoder/generic/csystemdependent.c
index f067b3937..07151fc4c 100644
--- a/vp8/encoder/generic/csystemdependent.c
+++ b/vp8/encoder/generic/csystemdependent.c
@@ -95,9 +95,10 @@ void vp8_cmachine_specific_config(VP8_COMP *cpi)
     cpi->rtcd.search.full_search             = vp8_full_search_sad;
 #endif
     cpi->rtcd.search.diamond_search          = vp8_diamond_search_sad;
-
+#if !(CONFIG_REALTIME_ONLY)
     cpi->rtcd.temporal.apply                 = vp8_temporal_filter_apply_c;
 #endif
+#endif
 
     // Pure C:
     vp8_yv12_copy_partial_frame_ptr = vp8_yv12_copy_partial_frame;
diff --git a/vp8/encoder/onyx_if.c b/vp8/encoder/onyx_if.c
index 3f0ee8e91..0d471cd28 100644
--- a/vp8/encoder/onyx_if.c
+++ b/vp8/encoder/onyx_if.c
@@ -330,35 +330,31 @@ static void setup_features(VP8_COMP *cpi)
 
 void vp8_dealloc_compressor_data(VP8_COMP *cpi)
 {
-        vpx_free(cpi->tplist);
+    vpx_free(cpi->tplist);
     cpi->tplist = NULL;
 
     // Delete last frame MV storage buffers
-        vpx_free(cpi->lfmv);
-
+    vpx_free(cpi->lfmv);
     cpi->lfmv = 0;
 
-        vpx_free(cpi->lf_ref_frame_sign_bias);
-
+    vpx_free(cpi->lf_ref_frame_sign_bias);
     cpi->lf_ref_frame_sign_bias = 0;
 
-        vpx_free(cpi->lf_ref_frame);
-
+    vpx_free(cpi->lf_ref_frame);
     cpi->lf_ref_frame = 0;
 
     // Delete sementation map
-        vpx_free(cpi->segmentation_map);
-
+    vpx_free(cpi->segmentation_map);
     cpi->segmentation_map = 0;
 
-        vpx_free(cpi->active_map);
-
+    vpx_free(cpi->active_map);
     cpi->active_map = 0;
 
+#if !(CONFIG_REALTIME_ONLY)
     // Delete first pass motion map
-        vpx_free(cpi->fp_motion_map);
-
+    vpx_free(cpi->fp_motion_map);
     cpi->fp_motion_map = 0;
+#endif
 
     vp8_de_alloc_frame_buffers(&cpi->common);
 
@@ -380,21 +376,17 @@ void vp8_dealloc_compressor_data(VP8_COMP *cpi)
     cpi->tok = 0;
 
     // Structure used to monitor GF usage
-        vpx_free(cpi->gf_active_flags);
-
+    vpx_free(cpi->gf_active_flags);
     cpi->gf_active_flags = 0;
 
-        vpx_free(cpi->mb.pip);
-
+    vpx_free(cpi->mb.pip);
     cpi->mb.pip = 0;
 
 #if !(CONFIG_REALTIME_ONLY)
-        vpx_free(cpi->total_stats);
-
+    vpx_free(cpi->total_stats);
     cpi->total_stats = 0;
 
-        vpx_free(cpi->this_frame_stats);
-
+    vpx_free(cpi->this_frame_stats);
     cpi->this_frame_stats = 0;
 #endif
 }
@@ -2227,9 +2219,11 @@ VP8_PTR vp8_create_compressor(VP8_CONFIG *oxcf)
     vpx_memset(cpi->active_map , 1, (cpi->common.mb_rows * cpi->common.mb_cols));
     cpi->active_map_enabled = 0;
 
+#if !(CONFIG_REALTIME_ONLY)
     // Create the first pass motion map structure and set to 0
     // Allocate space for maximum of 15 buffers
     CHECK_MEM_ERROR(cpi->fp_motion_map, vpx_calloc(15*cpi->common.MBs, 1));
+#endif
 
 #if 0
     // Experimental code for lagged and one pass
@@ -3195,8 +3189,11 @@ static void update_golden_frame_and_stats(VP8_COMP *cpi)
     // Update the Golden frame reconstruction buffer if signalled and the GF usage counts.
     if (cm->refresh_golden_frame)
     {
-        // Update the golden frame buffer
-        vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cm->yv12_fb[cm->gld_fb_idx]);
+        if (cm->frame_type != KEY_FRAME)
+        {
+            // Update the golden frame buffer
+            vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cm->yv12_fb[cm->gld_fb_idx]);
+        }
 
         // Select an interval before next GF
         if (!cpi->auto_gold)
@@ -4833,16 +4830,19 @@ static void encode_frame_to_data_rate
 
     if (cpi->oxcf.error_resilient_mode)
     {
-        // Is this an alternate reference update
-        if (cpi->common.refresh_alt_ref_frame)
-            vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cm->yv12_fb[cm->alt_fb_idx]);
+        if (cm->frame_type != KEY_FRAME)
+        {
+            // Is this an alternate reference update
+            if (cm->refresh_alt_ref_frame)
+                vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cm->yv12_fb[cm->alt_fb_idx]);
 
-        if (cpi->common.refresh_golden_frame)
-            vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cm->yv12_fb[cm->gld_fb_idx]);
+            if (cm->refresh_golden_frame)
+                vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cm->yv12_fb[cm->gld_fb_idx]);
+        }
     }
     else
     {
-        if (cpi->oxcf.play_alternate && cpi->common.refresh_alt_ref_frame && (cpi->common.frame_type != KEY_FRAME))
+        if (cpi->oxcf.play_alternate && cm->refresh_alt_ref_frame && (cm->frame_type != KEY_FRAME))
             // Update the alternate reference frame and stats as appropriate.
             update_alt_ref_frame_and_stats(cpi);
         else
@@ -5317,23 +5317,6 @@ int vp8_get_compressed_data(VP8_PTR ptr, unsigned int *frame_flags, unsigned lon
     if (cpi->compressor_speed == 2)
     {
         vp8_check_gf_quality(cpi);
-    }
-
-    if (!cpi)
-    {
-#if HAVE_ARMV7
-#if CONFIG_RUNTIME_CPU_DETECT
-        if (cm->rtcd.flags & HAS_NEON)
-#endif
-        {
-            vp8_pop_neon(store_reg);
-        }
-#endif
-        return 0;
-    }
-
-    if (cpi->compressor_speed == 2)
-    {
         vpx_usec_timer_start(&tsctimer);
         vpx_usec_timer_start(&ticktimer);
     }
diff --git a/vp8/encoder/onyx_int.h b/vp8/encoder/onyx_int.h
index 0d33b42a6..5a4b3c185 100644
--- a/vp8/encoder/onyx_int.h
+++ b/vp8/encoder/onyx_int.h
@@ -51,7 +51,9 @@
 #define MV_ZBIN_BOOST        4
 #define ZBIN_OQ_MAX 192
 
+#if !(CONFIG_REALTIME_ONLY)
 #define VP8_TEMPORAL_ALT_REF 1
+#endif
 
 typedef struct
 {
@@ -496,9 +498,11 @@ typedef struct
     FIRSTPASS_STATS *stats_in, *stats_in_end;
     struct vpx_codec_pkt_list  *output_pkt_list;
     int                          first_pass_done;
-    unsigned char *fp_motion_map;
 
+#if !(CONFIG_REALTIME_ONLY)
+    unsigned char *fp_motion_map;
     unsigned char *fp_motion_map_stats, *fp_motion_map_stats_save;
+#endif
 
 #if 0
     // Experimental code for lagged and one pass
diff --git a/vp8/encoder/x86/variance_impl_sse2.asm b/vp8/encoder/x86/variance_impl_sse2.asm
index 7178e7e31..6cdc47bc9 100644
--- a/vp8/encoder/x86/variance_impl_sse2.asm
+++ b/vp8/encoder/x86/variance_impl_sse2.asm
@@ -85,10 +85,9 @@ sym(vp8_get16x16var_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 6
-    GET_GOT     rbx
+    push rbx
     push rsi
     push rdi
-    sub         rsp, 16
     ; end prolog
 
         mov         rsi,            arg(0) ;[src_ptr]
@@ -97,6 +96,29 @@ sym(vp8_get16x16var_sse2):
         movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
         movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
 
+        ; Prefetch data
+        lea             rcx,    [rax+rax*2]
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+rax]
+        prefetcht0      [rsi+rax*2]
+        prefetcht0      [rsi+rcx]
+        lea             rbx,    [rsi+rax*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax*2]
+        prefetcht0      [rbx+rcx]
+
+        lea             rcx,    [rdx+rdx*2]
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+rdx]
+        prefetcht0      [rdi+rdx*2]
+        prefetcht0      [rdi+rcx]
+        lea             rbx,    [rdi+rdx*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx*2]
+        prefetcht0      [rbx+rcx]
+
         pxor        xmm0,           xmm0                        ; clear xmm0 for unpack
         pxor        xmm7,           xmm7                        ; clear xmm7 for accumulating diffs
 
@@ -107,6 +129,9 @@ var16loop:
         movdqu      xmm1,           XMMWORD PTR [rsi]
         movdqu      xmm2,           XMMWORD PTR [rdi]
 
+        prefetcht0      [rsi+rax*8]
+        prefetcht0      [rdi+rdx*8]
+
         movdqa      xmm3,           xmm1
         movdqa      xmm4,           xmm2
 
@@ -178,10 +203,9 @@ var16loop:
 
 
     ; begin epilog
-    add rsp, 16
     pop rdi
     pop rsi
-    RESTORE_GOT
+    pop rbx
     UNSHADOW_ARGS
     pop         rbp
     ret
diff --git a/vp8/encoder/x86/x86_csystemdependent.c b/vp8/encoder/x86/x86_csystemdependent.c
index 4e3ce070d..61c603229 100644
--- a/vp8/encoder/x86/x86_csystemdependent.c
+++ b/vp8/encoder/x86/x86_csystemdependent.c
@@ -301,7 +301,9 @@ void vp8_arch_x86_encoder_init(VP8_COMP *cpi)
 #endif
         cpi->rtcd.quantize.fastquantb            = vp8_fast_quantize_b_sse2;
 
+#if !(CONFIG_REALTIME_ONLY)
         cpi->rtcd.temporal.apply                 = vp8_temporal_filter_apply_sse2;
+#endif
     }
 #endif