libvpx: Pull from upstream

source/config/linux/ia32/vp9_rtcd.h

 #ifndef VP9_RTCD_H_
 #define VP9_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
 #else
 #define RTCD_EXTERN extern
 #endif
 
 /*
  * VP9
  */
 
 #include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_enums.h"
-#include "vp9/common/vp9_idct.h"
 
 struct macroblockd;
 
 /* Encoder forward decls */
 struct macroblock;
 struct vp9_variance_vtable;
 struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+unsigned int vp9_avg_8x8_c(const uint8_t *, int p);
+unsigned int vp9_avg_8x8_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*vp9_avg_8x8)(const uint8_t *, int p);
+
 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 RTCD_EXTERN int64_t (*vp9_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 
 void vp9_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 vp9_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 vp9_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);
 RTCD_EXTERN void (*vp9_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 vp9_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 377 matching lines @@
 
 unsigned int vp9_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int vp9_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int vp9_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int vp9_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 (*vp9_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-#define vp9_quantize_b vp9_quantize_b_c
+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, int zbin_oq_value, 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, int zbin_oq_value, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
 
 void vp9_quantize_fp_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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_fp vp9_quantize_fp_c
 
 void vp9_quantize_fp_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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
 
@@ skipping 435 matching lines @@
 void vp9_rtcd(void);
 
 #ifdef RTCD_C
 #include "vpx_ports/x86.h"
 static void setup_rtcd_internal(void)
 {
     int flags = x86_simd_caps();
 
     (void)flags;
 
+    vp9_avg_8x8 = vp9_avg_8x8_c;
+    if (flags & HAS_SSE2) vp9_avg_8x8 = vp9_avg_8x8_sse2;
     vp9_block_error = vp9_block_error_c;
     if (flags & HAS_SSE2) vp9_block_error = vp9_block_error_sse2;
     vp9_convolve8 = vp9_convolve8_c;
     if (flags & HAS_SSE2) vp9_convolve8 = vp9_convolve8_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8 = vp9_convolve8_ssse3;
     vp9_convolve8_avg = vp9_convolve8_avg_c;
     if (flags & HAS_SSE2) vp9_convolve8_avg = vp9_convolve8_avg_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8_avg = vp9_convolve8_avg_ssse3;
     vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_c;
     if (flags & HAS_SSE2) vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_sse2;
@@ skipping 145 matching lines @@
     vp9_lpf_vertical_8_dual = vp9_lpf_vertical_8_dual_c;
     if (flags & HAS_SSE2) vp9_lpf_vertical_8_dual = vp9_lpf_vertical_8_dual_sse2;
     vp9_mse16x16 = vp9_mse16x16_c;
     if (flags & HAS_SSE2) vp9_mse16x16 = vp9_mse16x16_sse2;
     vp9_mse16x8 = vp9_mse16x8_c;
     if (flags & HAS_SSE2) vp9_mse16x8 = vp9_mse16x8_sse2;
     vp9_mse8x16 = vp9_mse8x16_c;
     if (flags & HAS_SSE2) vp9_mse8x16 = vp9_mse8x16_sse2;
     vp9_mse8x8 = vp9_mse8x8_c;
     if (flags & HAS_SSE2) vp9_mse8x8 = vp9_mse8x8_sse2;
+    vp9_quantize_b = vp9_quantize_b_c;
+    if (flags & HAS_SSE2) vp9_quantize_b = vp9_quantize_b_sse2;
     vp9_sad16x16 = vp9_sad16x16_c;
     if (flags & HAS_SSE2) vp9_sad16x16 = vp9_sad16x16_sse2;
     vp9_sad16x16_avg = vp9_sad16x16_avg_c;
     if (flags & HAS_SSE2) vp9_sad16x16_avg = vp9_sad16x16_avg_sse2;
     vp9_sad16x16x3 = vp9_sad16x16x3_c;
     if (flags & HAS_SSE3) vp9_sad16x16x3 = vp9_sad16x16x3_sse3;
     if (flags & HAS_SSSE3) vp9_sad16x16x3 = vp9_sad16x16x3_ssse3;
     vp9_sad16x16x4d = vp9_sad16x16x4d_c;
     if (flags & HAS_SSE2) vp9_sad16x16x4d = vp9_sad16x16x4d_sse2;
     vp9_sad16x32 = vp9_sad16x32_c;
@@ skipping 200 matching lines @@
     vp9_variance8x8 = vp9_variance8x8_c;
     if (flags & HAS_SSE2) vp9_variance8x8 = vp9_variance8x8_sse2;
 }
 #endif
 
 #ifdef __cplusplus
 }  // extern "C"
 #endif
 
 #endif