Roll src/third_party/libvpx/source/libvpx/ 373e08f92..742ae4b24 (19 commits)

https://chromium.googlesource.com/webm/libvpx.git/+log/373e08f921e5..742ae4b24d28

$ git log 373e08f92..742ae4b24 --date=short --no-merges --format='%ad %ae %s'
2018-01-23 jzern vpx_codec_enc_init_multi: fix segfault w/vp9
2018-01-23 johannkoenig update .clang-format for v5.0.0
2018-01-23 linfengz Add vp9_highbd_iht16x16_256_add_sse4_1()
2018-01-23 linfengz Add "vpx_" prefix to 2 idct x86 functions
2018-01-23 yinshiyou-hf vp8: [loongson] fix bug of type conflict.
2018-01-22 jianj Fix frame sizes in pkt to support spatial layers.
2018-01-17 jianj Fix crash invalid params for vp8 multres. Add test.
2018-01-18 linfengz Add vp9_highbd_iht8x8_64_add_sse4_1()
2018-01-12 johannkoenig clang-format v5.0.0 vp9/
2018-01-12 johannkoenig remove spurious comments
2018-01-12 johannkoenig clang-format v5.0.0 vp8/
2018-01-12 johannkoenig clang-format v5.0.0 vpx_dsp/
2018-01-16 marpan vp8: Fix to multi-res-encoder for skipping streams.
2018-01-11 johannkoenig clang-format v5.0.0 mem_ops.h
2018-01-11 johannkoenig clang-format v5.0.0 vpx_util/vpx_atomic.h
2018-01-11 johannkoenig clang-format v5.0.0 y4minput.c
2018-01-11 johannkoenig clang-format v5.0.0 vpxenc.c
2018-01-11 johannkoenig clang-format v5.0.0 examples/
2018-01-11 johannkoenig clang-format v5.0.0 test/

Created with:
  roll-dep src/third_party/libvpx/source/libvpx
R=johannkoenig@google.com

BUG=webm:1485

Change-Id: I8fa8e20b1d69a13ed07e9a5a6ac37a86650bba79
Reviewed-on: https://chromium-review.googlesource.com/883779Reviewed-by: Johann Koenig <johannkoenig@google.com>
Commit-Queue: Jerome Jiang <jianj@google.com>
Cr-Commit-Position: refs/heads/master@{#531728}

DEPS

@@ -481,7 +481,7 @@ deps = {
   },
 
   'src/third_party/libvpx/source/libvpx':
-    Var('chromium_git') + '/webm/libvpx.git' + '@' +  '373e08f921e5bfd5a96963fabbbbe16ec793d44e',
+    Var('chromium_git') + '/webm/libvpx.git' + '@' +  '742ae4b24d28342a1642c9749826ee8428b20d6e',
 
   'src/third_party/libwebm/source':
     Var('chromium_git') + '/webm/libwebm.git' + '@' + 'b03c65468b06d097f27235d93d76bfc45f490ede',

third_party/libvpx/README.chromium

@@ -5,9 +5,9 @@ License: BSD
 License File: source/libvpx/LICENSE
 Security Critical: yes
 
-Date: Tuesday January 16 2018
+Date: Wednesday January 24 2018
 Branch: master
-Commit: 373e08f921e5bfd5a96963fabbbbe16ec793d44e
+Commit: 742ae4b24d28342a1642c9749826ee8428b20d6e
 
 Description:
 Contains the sources used to compile libvpx binaries used by Google Chrome and

third_party/libvpx/libvpx_srcs.gni

@@ -450,7 +450,9 @@ libvpx_srcs_x86_ssse3 = [
 ]
 libvpx_srcs_x86_sse4_1 = [
   "//third_party/libvpx/source/libvpx/vp8/encoder/x86/quantize_sse4.c",
+  "//third_party/libvpx/source/libvpx/vp9/common/x86/vp9_highbd_iht16x16_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vp9/common/x86/vp9_highbd_iht4x4_add_sse4.c",
+  "//third_party/libvpx/source/libvpx/vp9/common/x86/vp9_highbd_iht8x8_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vpx_dsp/x86/highbd_idct16x16_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vpx_dsp/x86/highbd_idct32x32_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vpx_dsp/x86/highbd_idct4x4_add_sse4.c",
@@ -924,7 +926,9 @@ libvpx_srcs_x86_64_ssse3 = [
 ]
 libvpx_srcs_x86_64_sse4_1 = [
   "//third_party/libvpx/source/libvpx/vp8/encoder/x86/quantize_sse4.c",
+  "//third_party/libvpx/source/libvpx/vp9/common/x86/vp9_highbd_iht16x16_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vp9/common/x86/vp9_highbd_iht4x4_add_sse4.c",
+  "//third_party/libvpx/source/libvpx/vp9/common/x86/vp9_highbd_iht8x8_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vpx_dsp/x86/highbd_idct16x16_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vpx_dsp/x86/highbd_idct32x32_add_sse4.c",
   "//third_party/libvpx/source/libvpx/vpx_dsp/x86/highbd_idct4x4_add_sse4.c",

third_party/libvpx/source/config/ios/arm-neon/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/ios/arm64/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/arm-neon-cpu-detect/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/arm-neon/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/arm/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/arm64/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/generic/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/ia32/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 
@@ -277,7 +277,16 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t* input,
                                    int pitch,
                                    int tx_type,
                                    int bd);
-#define vp9_highbd_iht16x16_256_add vp9_highbd_iht16x16_256_add_c
+void vp9_highbd_iht16x16_256_add_sse4_1(const tran_low_t* input,
+                                        uint16_t* output,
+                                        int pitch,
+                                        int tx_type,
+                                        int bd);
+RTCD_EXTERN void (*vp9_highbd_iht16x16_256_add)(const tran_low_t* input,
+                                                uint16_t* output,
+                                                int pitch,
+                                                int tx_type,
+                                                int bd);
 
 void vp9_highbd_iht4x4_16_add_c(const tran_low_t* input,
                                 uint16_t* dest,
@@ -300,7 +309,16 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t* input,
                                 int stride,
                                 int tx_type,
                                 int bd);
-#define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_c
+void vp9_highbd_iht8x8_64_add_sse4_1(const tran_low_t* input,
+                                     uint16_t* dest,
+                                     int stride,
+                                     int tx_type,
+                                     int bd);
+RTCD_EXTERN void (*vp9_highbd_iht8x8_64_add)(const tran_low_t* input,
+                                             uint16_t* dest,
+                                             int stride,
+                                             int tx_type,
+                                             int bd);
 
 void vp9_highbd_mbpost_proc_across_ip_c(uint16_t* src,
                                         int pitch,
@@ -512,9 +530,15 @@ static void setup_rtcd_internal(void) {
   vp9_highbd_block_error = vp9_highbd_block_error_c;
   if (flags & HAS_SSE2)
     vp9_highbd_block_error = vp9_highbd_block_error_sse2;
+  vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_sse4_1;
   vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_c;
   if (flags & HAS_SSE4_1)
     vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_sse4_1;
+  vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_sse4_1;
   vp9_iht16x16_256_add = vp9_iht16x16_256_add_c;
   if (flags & HAS_SSE2)
     vp9_iht16x16_256_add = vp9_iht16x16_256_add_sse2;

third_party/libvpx/source/config/linux/mips64el/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/mipsel/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/linux/x64/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 
@@ -246,7 +246,16 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t* input,
                                    int pitch,
                                    int tx_type,
                                    int bd);
-#define vp9_highbd_iht16x16_256_add vp9_highbd_iht16x16_256_add_c
+void vp9_highbd_iht16x16_256_add_sse4_1(const tran_low_t* input,
+                                        uint16_t* output,
+                                        int pitch,
+                                        int tx_type,
+                                        int bd);
+RTCD_EXTERN void (*vp9_highbd_iht16x16_256_add)(const tran_low_t* input,
+                                                uint16_t* output,
+                                                int pitch,
+                                                int tx_type,
+                                                int bd);
 
 void vp9_highbd_iht4x4_16_add_c(const tran_low_t* input,
                                 uint16_t* dest,
@@ -269,7 +278,16 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t* input,
                                 int stride,
                                 int tx_type,
                                 int bd);
-#define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_c
+void vp9_highbd_iht8x8_64_add_sse4_1(const tran_low_t* input,
+                                     uint16_t* dest,
+                                     int stride,
+                                     int tx_type,
+                                     int bd);
+RTCD_EXTERN void (*vp9_highbd_iht8x8_64_add)(const tran_low_t* input,
+                                             uint16_t* dest,
+                                             int stride,
+                                             int tx_type,
+                                             int bd);
 
 void vp9_highbd_mbpost_proc_across_ip_c(uint16_t* src,
                                         int pitch,
@@ -476,9 +494,15 @@ static void setup_rtcd_internal(void) {
   vp9_fdct8x8_quant = vp9_fdct8x8_quant_c;
   if (flags & HAS_SSSE3)
     vp9_fdct8x8_quant = vp9_fdct8x8_quant_ssse3;
+  vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_sse4_1;
   vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_c;
   if (flags & HAS_SSE4_1)
     vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_sse4_1;
+  vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_sse4_1;
   vp9_quantize_fp = vp9_quantize_fp_sse2;
   if (flags & HAS_SSSE3)
     vp9_quantize_fp = vp9_quantize_fp_ssse3;

third_party/libvpx/source/config/mac/ia32/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 
@@ -277,7 +277,16 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t* input,
                                    int pitch,
                                    int tx_type,
                                    int bd);
-#define vp9_highbd_iht16x16_256_add vp9_highbd_iht16x16_256_add_c
+void vp9_highbd_iht16x16_256_add_sse4_1(const tran_low_t* input,
+                                        uint16_t* output,
+                                        int pitch,
+                                        int tx_type,
+                                        int bd);
+RTCD_EXTERN void (*vp9_highbd_iht16x16_256_add)(const tran_low_t* input,
+                                                uint16_t* output,
+                                                int pitch,
+                                                int tx_type,
+                                                int bd);
 
 void vp9_highbd_iht4x4_16_add_c(const tran_low_t* input,
                                 uint16_t* dest,
@@ -300,7 +309,16 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t* input,
                                 int stride,
                                 int tx_type,
                                 int bd);
-#define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_c
+void vp9_highbd_iht8x8_64_add_sse4_1(const tran_low_t* input,
+                                     uint16_t* dest,
+                                     int stride,
+                                     int tx_type,
+                                     int bd);
+RTCD_EXTERN void (*vp9_highbd_iht8x8_64_add)(const tran_low_t* input,
+                                             uint16_t* dest,
+                                             int stride,
+                                             int tx_type,
+                                             int bd);
 
 void vp9_highbd_mbpost_proc_across_ip_c(uint16_t* src,
                                         int pitch,
@@ -512,9 +530,15 @@ static void setup_rtcd_internal(void) {
   vp9_highbd_block_error = vp9_highbd_block_error_c;
   if (flags & HAS_SSE2)
     vp9_highbd_block_error = vp9_highbd_block_error_sse2;
+  vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_sse4_1;
   vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_c;
   if (flags & HAS_SSE4_1)
     vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_sse4_1;
+  vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_sse4_1;
   vp9_iht16x16_256_add = vp9_iht16x16_256_add_c;
   if (flags & HAS_SSE2)
     vp9_iht16x16_256_add = vp9_iht16x16_256_add_sse2;

third_party/libvpx/source/config/mac/x64/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 
@@ -246,7 +246,16 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t* input,
                                    int pitch,
                                    int tx_type,
                                    int bd);
-#define vp9_highbd_iht16x16_256_add vp9_highbd_iht16x16_256_add_c
+void vp9_highbd_iht16x16_256_add_sse4_1(const tran_low_t* input,
+                                        uint16_t* output,
+                                        int pitch,
+                                        int tx_type,
+                                        int bd);
+RTCD_EXTERN void (*vp9_highbd_iht16x16_256_add)(const tran_low_t* input,
+                                                uint16_t* output,
+                                                int pitch,
+                                                int tx_type,
+                                                int bd);
 
 void vp9_highbd_iht4x4_16_add_c(const tran_low_t* input,
                                 uint16_t* dest,
@@ -269,7 +278,16 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t* input,
                                 int stride,
                                 int tx_type,
                                 int bd);
-#define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_c
+void vp9_highbd_iht8x8_64_add_sse4_1(const tran_low_t* input,
+                                     uint16_t* dest,
+                                     int stride,
+                                     int tx_type,
+                                     int bd);
+RTCD_EXTERN void (*vp9_highbd_iht8x8_64_add)(const tran_low_t* input,
+                                             uint16_t* dest,
+                                             int stride,
+                                             int tx_type,
+                                             int bd);
 
 void vp9_highbd_mbpost_proc_across_ip_c(uint16_t* src,
                                         int pitch,
@@ -476,9 +494,15 @@ static void setup_rtcd_internal(void) {
   vp9_fdct8x8_quant = vp9_fdct8x8_quant_c;
   if (flags & HAS_SSSE3)
     vp9_fdct8x8_quant = vp9_fdct8x8_quant_ssse3;
+  vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_sse4_1;
   vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_c;
   if (flags & HAS_SSE4_1)
     vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_sse4_1;
+  vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_sse4_1;
   vp9_quantize_fp = vp9_quantize_fp_sse2;
   if (flags & HAS_SSSE3)
     vp9_quantize_fp = vp9_quantize_fp_ssse3;

third_party/libvpx/source/config/nacl/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 

third_party/libvpx/source/config/vpx_version.h

@@ -2,7 +2,7 @@
 #define VERSION_MAJOR  1
 #define VERSION_MINOR  6
 #define VERSION_PATCH  1
-#define VERSION_EXTRA  "1494-g373e08f92"
+#define VERSION_EXTRA  "1521-g742ae4b24"
 #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
-#define VERSION_STRING_NOSP "v1.6.1-1494-g373e08f92"
-#define VERSION_STRING      " v1.6.1-1494-g373e08f92"
+#define VERSION_STRING_NOSP "v1.6.1-1521-g742ae4b24"
+#define VERSION_STRING      " v1.6.1-1521-g742ae4b24"

third_party/libvpx/source/config/win/ia32/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 
@@ -277,7 +277,16 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t* input,
                                    int pitch,
                                    int tx_type,
                                    int bd);
-#define vp9_highbd_iht16x16_256_add vp9_highbd_iht16x16_256_add_c
+void vp9_highbd_iht16x16_256_add_sse4_1(const tran_low_t* input,
+                                        uint16_t* output,
+                                        int pitch,
+                                        int tx_type,
+                                        int bd);
+RTCD_EXTERN void (*vp9_highbd_iht16x16_256_add)(const tran_low_t* input,
+                                                uint16_t* output,
+                                                int pitch,
+                                                int tx_type,
+                                                int bd);
 
 void vp9_highbd_iht4x4_16_add_c(const tran_low_t* input,
                                 uint16_t* dest,
@@ -300,7 +309,16 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t* input,
                                 int stride,
                                 int tx_type,
                                 int bd);
-#define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_c
+void vp9_highbd_iht8x8_64_add_sse4_1(const tran_low_t* input,
+                                     uint16_t* dest,
+                                     int stride,
+                                     int tx_type,
+                                     int bd);
+RTCD_EXTERN void (*vp9_highbd_iht8x8_64_add)(const tran_low_t* input,
+                                             uint16_t* dest,
+                                             int stride,
+                                             int tx_type,
+                                             int bd);
 
 void vp9_highbd_mbpost_proc_across_ip_c(uint16_t* src,
                                         int pitch,
@@ -512,9 +530,15 @@ static void setup_rtcd_internal(void) {
   vp9_highbd_block_error = vp9_highbd_block_error_c;
   if (flags & HAS_SSE2)
     vp9_highbd_block_error = vp9_highbd_block_error_sse2;
+  vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_sse4_1;
   vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_c;
   if (flags & HAS_SSE4_1)
     vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_sse4_1;
+  vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_sse4_1;
   vp9_iht16x16_256_add = vp9_iht16x16_256_add_c;
   if (flags & HAS_SSE2)
     vp9_iht16x16_256_add = vp9_iht16x16_256_add_sse2;

third_party/libvpx/source/config/win/x64/vp9_rtcd.h

@@ -12,10 +12,10 @@
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
 #include "vp9/common/vp9_filter.h"
-#include "vpx/vpx_integer.h"
 
 struct macroblockd;
 
@@ -246,7 +246,16 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t* input,
                                    int pitch,
                                    int tx_type,
                                    int bd);
-#define vp9_highbd_iht16x16_256_add vp9_highbd_iht16x16_256_add_c
+void vp9_highbd_iht16x16_256_add_sse4_1(const tran_low_t* input,
+                                        uint16_t* output,
+                                        int pitch,
+                                        int tx_type,
+                                        int bd);
+RTCD_EXTERN void (*vp9_highbd_iht16x16_256_add)(const tran_low_t* input,
+                                                uint16_t* output,
+                                                int pitch,
+                                                int tx_type,
+                                                int bd);
 
 void vp9_highbd_iht4x4_16_add_c(const tran_low_t* input,
                                 uint16_t* dest,
@@ -269,7 +278,16 @@ void vp9_highbd_iht8x8_64_add_c(const tran_low_t* input,
                                 int stride,
                                 int tx_type,
                                 int bd);
-#define vp9_highbd_iht8x8_64_add vp9_highbd_iht8x8_64_add_c
+void vp9_highbd_iht8x8_64_add_sse4_1(const tran_low_t* input,
+                                     uint16_t* dest,
+                                     int stride,
+                                     int tx_type,
+                                     int bd);
+RTCD_EXTERN void (*vp9_highbd_iht8x8_64_add)(const tran_low_t* input,
+                                             uint16_t* dest,
+                                             int stride,
+                                             int tx_type,
+                                             int bd);
 
 void vp9_highbd_mbpost_proc_across_ip_c(uint16_t* src,
                                         int pitch,
@@ -476,9 +494,15 @@ static void setup_rtcd_internal(void) {
   vp9_fdct8x8_quant = vp9_fdct8x8_quant_c;
   if (flags & HAS_SSSE3)
     vp9_fdct8x8_quant = vp9_fdct8x8_quant_ssse3;
+  vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht16x16_256_add = vp9_highbd_iht16x16_256_add_sse4_1;
   vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_c;
   if (flags & HAS_SSE4_1)
     vp9_highbd_iht4x4_16_add = vp9_highbd_iht4x4_16_add_sse4_1;
+  vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_c;
+  if (flags & HAS_SSE4_1)
+    vp9_highbd_iht8x8_64_add = vp9_highbd_iht8x8_64_add_sse4_1;
   vp9_quantize_fp = vp9_quantize_fp_sse2;
   if (flags & HAS_SSSE3)
     vp9_quantize_fp = vp9_quantize_fp_ssse3;