libvpx: Pull from upstream

source/config/win/x64/vpx_dsp_rtcd.h

 #ifndef VPX_DSP_RTCD_H_
 #define VPX_DSP_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
 #else
 #define RTCD_EXTERN extern
 #endif
 
 /*
  * DSP
  */
 
 #include "vpx/vpx_integer.h"
 #include "vpx_dsp/vpx_dsp_common.h"
 
 
 #ifdef __cplusplus
 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);
-#define vp9_idct16x16_10_add vp9_idct16x16_10_add_sse2
-
-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);
-#define vp9_idct16x16_1_add vp9_idct16x16_1_add_sse2
-
-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);
-#define vp9_idct16x16_256_add vp9_idct16x16_256_add_sse2
-
-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);
-#define vp9_idct32x32_1024_add vp9_idct32x32_1024_add_sse2
-
-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);
-#define vp9_idct32x32_1_add vp9_idct32x32_1_add_sse2
-
-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);
-#define vp9_idct32x32_34_add vp9_idct32x32_34_add_sse2
-
-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);
-#define vp9_idct4x4_16_add vp9_idct4x4_16_add_sse2
-
-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);
-#define vp9_idct4x4_1_add vp9_idct4x4_1_add_sse2
-
-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);
-void vp9_idct8x8_12_add_ssse3(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);
-#define vp9_idct8x8_1_add vp9_idct8x8_1_add_sse2
-
-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);
-void vp9_idct8x8_64_add_ssse3(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);
-#define vp9_iwht4x4_16_add vp9_iwht4x4_16_add_sse2
-
-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);
-void vp9_quantize_b_ssse3(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);
-void vp9_quantize_b_32x32_ssse3(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_32x32)(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_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
 
 void vpx_convolve8_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_convolve8_sse2(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_convolve8_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);
 void vpx_convolve8_avx2(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_convolve8)(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_convolve8_avg_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);
@@ skipping 249 matching lines @@
 RTCD_EXTERN void (*vpx_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vpx_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*vpx_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 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);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_sse2
+
+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);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_sse2
+
+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);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_sse2
+
+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);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_sse2
+
+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);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_sse2
+
+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);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_sse2
+
+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);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_sse2
+
+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);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_sse2
+
+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);
+void vpx_idct8x8_12_add_ssse3(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);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_sse2
+
+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);
+void vpx_idct8x8_64_add_ssse3(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);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_sse2
+
+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);
 RTCD_EXTERN void (*vpx_lpf_horizontal_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 
 void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 void vpx_lpf_horizontal_4_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 #define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_mmx
 
 void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
@@ skipping 43 matching lines @@
 #define vpx_mse16x8 vpx_mse16x8_sse2
 
 unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int vpx_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define vpx_mse8x16 vpx_mse8x16_sse2
 
 unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int vpx_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define vpx_mse8x8 vpx_mse8x8_sse2
 
+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);
+void vpx_quantize_b_ssse3(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);
+void vpx_quantize_b_32x32_ssse3(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_32x32)(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);
+
 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);
 #define vpx_sad16x16 vpx_sad16x16_sse2
 
 unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 unsigned int vpx_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vpx_sad16x16_avg vpx_sad16x16_avg_sse2
 
 void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
@@ skipping 213 matching lines @@
 RTCD_EXTERN void (*vpx_sad8x8x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 
 void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 void vpx_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define vpx_sad8x8x4d vpx_sad8x8x4d_sse2
 
 void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 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);
 RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x16)(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_variance16x32_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_variance16x32_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_variance16x32_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);
 RTCD_EXTERN uint32_t (*vpx_sub_pixel_avg_variance16x32)(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);
 
@@ skipping 242 matching lines @@
 void vpx_dsp_rtcd(void);
 
 #ifdef RTCD_C
 #include "vpx_ports/x86.h"
 static void setup_rtcd_internal(void)
 {
     int flags = x86_simd_caps();
 
     (void)flags;
 
-    vp9_idct8x8_12_add = vp9_idct8x8_12_add_sse2;
-    if (flags & HAS_SSSE3) vp9_idct8x8_12_add = vp9_idct8x8_12_add_ssse3;
-    vp9_idct8x8_64_add = vp9_idct8x8_64_add_sse2;
-    if (flags & HAS_SSSE3) vp9_idct8x8_64_add = vp9_idct8x8_64_add_ssse3;
-    vp9_quantize_b = vp9_quantize_b_sse2;
-    if (flags & HAS_SSSE3) vp9_quantize_b = vp9_quantize_b_ssse3;
-    vp9_quantize_b_32x32 = vp9_quantize_b_32x32_c;
-    if (flags & HAS_SSSE3) vp9_quantize_b_32x32 = vp9_quantize_b_32x32_ssse3;
     vpx_convolve8 = vpx_convolve8_sse2;
     if (flags & HAS_SSSE3) vpx_convolve8 = vpx_convolve8_ssse3;
     if (flags & HAS_AVX2) vpx_convolve8 = vpx_convolve8_avx2;
     vpx_convolve8_avg = vpx_convolve8_avg_sse2;
     if (flags & HAS_SSSE3) vpx_convolve8_avg = vpx_convolve8_avg_ssse3;
     vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_sse2;
     if (flags & HAS_SSSE3) vpx_convolve8_avg_horiz = vpx_convolve8_avg_horiz_ssse3;
     vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_sse2;
     if (flags & HAS_SSSE3) vpx_convolve8_avg_vert = vpx_convolve8_avg_vert_ssse3;
     vpx_convolve8_horiz = vpx_convolve8_horiz_sse2;
@@ skipping 43 matching lines @@
     vpx_get16x16var = vpx_get16x16var_sse2;
     if (flags & HAS_AVX2) vpx_get16x16var = vpx_get16x16var_avx2;
     vpx_h_predictor_16x16 = vpx_h_predictor_16x16_c;
     if (flags & HAS_SSSE3) vpx_h_predictor_16x16 = vpx_h_predictor_16x16_ssse3;
     vpx_h_predictor_32x32 = vpx_h_predictor_32x32_c;
     if (flags & HAS_SSSE3) vpx_h_predictor_32x32 = vpx_h_predictor_32x32_ssse3;
     vpx_h_predictor_4x4 = vpx_h_predictor_4x4_c;
     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_idct8x8_12_add = vpx_idct8x8_12_add_sse2;
+    if (flags & HAS_SSSE3) vpx_idct8x8_12_add = vpx_idct8x8_12_add_ssse3;
+    vpx_idct8x8_64_add = vpx_idct8x8_64_add_sse2;
+    if (flags & HAS_SSSE3) vpx_idct8x8_64_add = vpx_idct8x8_64_add_ssse3;
     vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_sse2;
     if (flags & HAS_AVX2) vpx_lpf_horizontal_16 = vpx_lpf_horizontal_16_avx2;
     vpx_mse16x16 = vpx_mse16x16_sse2;
     if (flags & HAS_AVX2) vpx_mse16x16 = vpx_mse16x16_avx2;
+    vpx_quantize_b = vpx_quantize_b_sse2;
+    if (flags & HAS_SSSE3) vpx_quantize_b = vpx_quantize_b_ssse3;
+    vpx_quantize_b_32x32 = vpx_quantize_b_32x32_c;
+    if (flags & HAS_SSSE3) vpx_quantize_b_32x32 = vpx_quantize_b_32x32_ssse3;
     vpx_sad16x16x3 = vpx_sad16x16x3_c;
     if (flags & HAS_SSE3) vpx_sad16x16x3 = vpx_sad16x16x3_sse3;
     if (flags & HAS_SSSE3) vpx_sad16x16x3 = vpx_sad16x16x3_ssse3;
     vpx_sad16x16x8 = vpx_sad16x16x8_c;
     if (flags & HAS_SSE4_1) vpx_sad16x16x8 = vpx_sad16x16x8_sse4_1;
     vpx_sad16x8x3 = vpx_sad16x8x3_c;
     if (flags & HAS_SSE3) vpx_sad16x8x3 = vpx_sad16x8x3_sse3;
     if (flags & HAS_SSSE3) vpx_sad16x8x3 = vpx_sad16x8x3_ssse3;
     vpx_sad16x8x8 = vpx_sad16x8x8_c;
     if (flags & HAS_SSE4_1) vpx_sad16x8x8 = vpx_sad16x8x8_sse4_1;
@@ skipping 26 matching lines @@
     vpx_sad64x64x4d = vpx_sad64x64x4d_sse2;
     if (flags & HAS_AVX2) vpx_sad64x64x4d = vpx_sad64x64x4d_avx2;
     vpx_sad8x16x3 = vpx_sad8x16x3_c;
     if (flags & HAS_SSE3) vpx_sad8x16x3 = vpx_sad8x16x3_sse3;
     vpx_sad8x16x8 = vpx_sad8x16x8_c;
     if (flags & HAS_SSE4_1) vpx_sad8x16x8 = vpx_sad8x16x8_sse4_1;
     vpx_sad8x8x3 = vpx_sad8x8x3_c;
     if (flags & HAS_SSE3) vpx_sad8x8x3 = vpx_sad8x8x3_sse3;
     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_sse2;
     if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x16 = vpx_sub_pixel_avg_variance16x16_ssse3;
     vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_sse2;
     if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x32 = vpx_sub_pixel_avg_variance16x32_ssse3;
     vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_sse2;
     if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance16x8 = vpx_sub_pixel_avg_variance16x8_ssse3;
     vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_sse2;
     if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x16 = vpx_sub_pixel_avg_variance32x16_ssse3;
     vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_sse2;
     if (flags & HAS_SSSE3) vpx_sub_pixel_avg_variance32x32 = vpx_sub_pixel_avg_variance32x32_ssse3;
@@ skipping 54 matching lines @@
     vpx_variance64x64 = vpx_variance64x64_sse2;
     if (flags & HAS_AVX2) vpx_variance64x64 = vpx_variance64x64_avx2;
 }
 #endif
 
 #ifdef __cplusplus
 }  // extern "C"
 #endif
 
 #endif