libvpx: Pull from upstream

source/config/win/ia32/vpx_dsp_rtcd.h

Index: source/config/win/ia32/vpx_dsp_rtcd.h
diff --git a/source/config/win/ia32/vpx_dsp_rtcd.h b/source/config/win/ia32/vpx_dsp_rtcd.h
index 30709eec7e107ea40e6722a747408650c48466b0..8f0da9d1cecc7dab5f410ca5950a6d7f0724a53b 100644
--- a/source/config/win/ia32/vpx_dsp_rtcd.h
+++ b/source/config/win/ia32/vpx_dsp_rtcd.h
@@ -19,64 +19,6 @@
 extern "C" {
 #endif
 
-void vp9_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct4x4_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct8x8_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-void vp9_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-RTCD_EXTERN void (*vp9_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
-
-void vp9_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-#define vp9_iwht4x4_1_add vp9_iwht4x4_1_add_c
-
-void vp9_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void vp9_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-RTCD_EXTERN void (*vp9_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-
-void vp9_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-#define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
-
 void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
 #define vpx_comp_avg_pred vpx_comp_avg_pred_c
 
@@ -345,6 +287,57 @@ void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abov
 void vpx_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*vpx_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct4x4_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct8x8_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*vpx_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
 void vpx_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 void vpx_lpf_horizontal_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 void vpx_lpf_horizontal_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
@@ -408,6 +401,13 @@ unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint
 unsigned int vpx_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vpx_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*vpx_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
 unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int vpx_sad16x16_mmx(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int vpx_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -641,6 +641,25 @@ void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_p
 void vpx_sad8x8x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 RTCD_EXTERN void (*vpx_sad8x8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*vpx_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
 uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t vpx_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t vpx_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
@@ -902,32 +921,6 @@ static void setup_rtcd_internal(void)
 
     (void)flags;
 
-    vp9_idct16x16_10_add = vp9_idct16x16_10_add_c;
-    if (flags & HAS_SSE2) vp9_idct16x16_10_add = vp9_idct16x16_10_add_sse2;
-    vp9_idct16x16_1_add = vp9_idct16x16_1_add_c;
-    if (flags & HAS_SSE2) vp9_idct16x16_1_add = vp9_idct16x16_1_add_sse2;
-    vp9_idct16x16_256_add = vp9_idct16x16_256_add_c;
-    if (flags & HAS_SSE2) vp9_idct16x16_256_add = vp9_idct16x16_256_add_sse2;
-    vp9_idct32x32_1024_add = vp9_idct32x32_1024_add_c;
-    if (flags & HAS_SSE2) vp9_idct32x32_1024_add = vp9_idct32x32_1024_add_sse2;
-    vp9_idct32x32_1_add = vp9_idct32x32_1_add_c;
-    if (flags & HAS_SSE2) vp9_idct32x32_1_add = vp9_idct32x32_1_add_sse2;
-    vp9_idct32x32_34_add = vp9_idct32x32_34_add_c;
-    if (flags & HAS_SSE2) vp9_idct32x32_34_add = vp9_idct32x32_34_add_sse2;
-    vp9_idct4x4_16_add = vp9_idct4x4_16_add_c;
-    if (flags & HAS_SSE2) vp9_idct4x4_16_add = vp9_idct4x4_16_add_sse2;
-    vp9_idct4x4_1_add = vp9_idct4x4_1_add_c;
-    if (flags & HAS_SSE2) vp9_idct4x4_1_add = vp9_idct4x4_1_add_sse2;
-    vp9_idct8x8_12_add = vp9_idct8x8_12_add_c;
-    if (flags & HAS_SSE2) vp9_idct8x8_12_add = vp9_idct8x8_12_add_sse2;
-    vp9_idct8x8_1_add = vp9_idct8x8_1_add_c;
-    if (flags & HAS_SSE2) vp9_idct8x8_1_add = vp9_idct8x8_1_add_sse2;
-    vp9_idct8x8_64_add = vp9_idct8x8_64_add_c;
-    if (flags & HAS_SSE2) vp9_idct8x8_64_add = vp9_idct8x8_64_add_sse2;
-    vp9_iwht4x4_16_add = vp9_iwht4x4_16_add_c;
-    if (flags & HAS_SSE2) vp9_iwht4x4_16_add = vp9_iwht4x4_16_add_sse2;
-    vp9_quantize_b = vp9_quantize_b_c;
-    if (flags & HAS_SSE2) vp9_quantize_b = vp9_quantize_b_sse2;
     vpx_convolve8 = vpx_convolve8_c;
     if (flags & HAS_SSE2) vpx_convolve8 = vpx_convolve8_sse2;
     if (flags & HAS_SSSE3) vpx_convolve8 = vpx_convolve8_ssse3;
@@ -1054,6 +1047,30 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) vpx_h_predictor_4x4 = vpx_h_predictor_4x4_ssse3;
     vpx_h_predictor_8x8 = vpx_h_predictor_8x8_c;
     if (flags & HAS_SSSE3) vpx_h_predictor_8x8 = vpx_h_predictor_8x8_ssse3;
+    vpx_idct16x16_10_add = vpx_idct16x16_10_add_c;
+    if (flags & HAS_SSE2) vpx_idct16x16_10_add = vpx_idct16x16_10_add_sse2;
+    vpx_idct16x16_1_add = vpx_idct16x16_1_add_c;
+    if (flags & HAS_SSE2) vpx_idct16x16_1_add = vpx_idct16x16_1_add_sse2;
+    vpx_idct16x16_256_add = vpx_idct16x16_256_add_c;
+    if (flags & HAS_SSE2) vpx_idct16x16_256_add = vpx_idct16x16_256_add_sse2;
+    vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_c;
+    if (flags & HAS_SSE2) vpx_idct32x32_1024_add = vpx_idct32x32_1024_add_sse2;
+    vpx_idct32x32_1_add = vpx_idct32x32_1_add_c;
+    if (flags & HAS_SSE2) vpx_idct32x32_1_add = vpx_idct32x32_1_add_sse2;
+    vpx_idct32x32_34_add = vpx_idct32x32_34_add_c;
+    if (flags & HAS_SSE2) vpx_idct32x32_34_add = vpx_idct32x32_34_add_sse2;
+    vpx_idct4x4_16_add = vpx_idct4x4_16_add_c;
+    if (flags & HAS_SSE2) vpx_idct4x4_16_add = vpx_idct4x4_16_add_sse2;
+    vpx_idct4x4_1_add = vpx_idct4x4_1_add_c;
+    if (flags & HAS_SSE2) vpx_idct4x4_1_add = vpx_idct4x4_1_add_sse2;
+    vpx_idct8x8_12_add = vpx_idct8x8_12_add_c;
+    if (flags & HAS_SSE2) vpx_idct8x8_12_add = vpx_idct8x8_12_add_sse2;
+    vpx_idct8x8_1_add = vpx_idct8x8_1_add_c;
+    if (flags & HAS_SSE2) vpx_idct8x8_1_add = vpx_idct8x8_1_add_sse2;
+    vpx_idct8x8_64_add = vpx_idct8x8_64_add_c;
+    if (flags & HAS_SSE2) vpx_idct8x8_64_add = vpx_idct8x8_64_add_sse2;
+    vpx_iwht4x4_16_add = vpx_iwht4x4_16_add_c;
+    if (flags & HAS_SSE2) vpx_iwht4x4_16_add = vpx_iwht4x4_16_add_sse2;
     vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_c;
     if (flags & HAS_SSE2) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_sse2;
     if (flags & HAS_AVX2) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_avx2;
@@ -1087,6 +1104,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) vpx_mse8x16 = vpx_mse8x16_sse2;
     vpx_mse8x8 = vpx_mse8x8_c;
     if (flags & HAS_SSE2) vpx_mse8x8 = vpx_mse8x8_sse2;
+    vpx_quantize_b = vpx_quantize_b_c;
+    if (flags & HAS_SSE2) vpx_quantize_b = vpx_quantize_b_sse2;
     vpx_sad16x16 = vpx_sad16x16_c;
     if (flags & HAS_MMX) vpx_sad16x16 = vpx_sad16x16_mmx;
     if (flags & HAS_SSE2) vpx_sad16x16 = vpx_sad16x16_sse2;
@@ -1204,6 +1223,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) vpx_sad8x8x4d = vpx_sad8x8x4d_sse2;
     vpx_sad8x8x8 = vpx_sad8x8x8_c;
     if (flags & HAS_SSE4_1) vpx_sad8x8x8 = vpx_sad8x8x8_sse4_1;
+    vpx_scaled_2d = vpx_scaled_2d_c;
+    if (flags & HAS_SSSE3) vpx_scaled_2d = vpx_scaled_2d_ssse3;
     vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_c;
     if (flags & HAS_SSE2) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_sse2;
     if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_ssse3;