libvpx: Pull from upstream

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

 #ifndef VP9_RTCD_H_
 #define VP9_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
 #else
 #define RTCD_EXTERN extern
 #endif
 
 #ifdef __cplusplus
@@ skipping 199 matching lines @@
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct32x32_rd vp9_fdct32x32_rd_c
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4 vp9_fdct4x4_c
 
 void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fdct4x4_1 vp9_fdct4x4_1_c
 
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct8x8 vp9_fdct8x8_c
+void vp9_fdct8x8_neon(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8 vp9_fdct8x8_neon
 
 void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct8x8_1 vp9_fdct8x8_1_c
+void vp9_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8_1 vp9_fdct8x8_1_neon
 
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht16x16 vp9_fht16x16_c
 
 void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht4x4 vp9_fht4x4_c
 
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 #define vp9_fht8x8 vp9_fht8x8_c
 
 int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
 int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define vp9_full_search_sad vp9_full_search_sad_c
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
 #define vp9_fwht4x4 vp9_fwht4x4_c
 
 void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get16x16var vp9_get16x16var_c
+void vp9_get16x16var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get16x16var vp9_get16x16var_neon
 
 void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get8x8var vp9_get8x8var_c
+void vp9_get8x8var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get8x8var vp9_get8x8var_neon
 
 unsigned int vp9_get_mb_ss_c(const int16_t *);
 #define vp9_get_mb_ss vp9_get_mb_ss_c
 
 void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_h_predictor_16x16 vp9_h_predictor_16x16_neon
 
 void vp9_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ skipping 124 matching lines @@
 unsigned int vp9_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define vp9_mse8x8 vp9_mse8x8_c
 
 void vp9_quantize_b_c(const int16_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, int16_t *qcoeff_ptr, int16_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_32x32_c(const int16_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, int16_t *qcoeff_ptr, int16_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 int16_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, int16_t *qcoeff_ptr, int16_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_neon(const int16_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, int16_t *qcoeff_ptr, int16_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_neon
 
 void vp9_quantize_fp_32x32_c(const int16_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, int16_t *qcoeff_ptr, int16_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
 
 int vp9_refining_search_sad_c(const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define vp9_refining_search_sad vp9_refining_search_sad_c
 
 unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 unsigned int vp9_sad16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 #define vp9_sad16x16 vp9_sad16x16_neon
@@ skipping 141 matching lines @@
 unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x4_avg vp9_sad8x4_avg_c
 
 void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
 #define vp9_sad8x4x4d vp9_sad8x4x4d_c
 
 void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define vp9_sad8x4x8 vp9_sad8x4x8_c
 
 unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
-#define vp9_sad8x8 vp9_sad8x8_c
+unsigned int vp9_sad8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+#define vp9_sad8x8 vp9_sad8x8_neon
 
 unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 #define vp9_sad8x8_avg vp9_sad8x8_avg_c
 
 void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
 #define vp9_sad8x8x3 vp9_sad8x8x3_c
 
 void vp9_sad8x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 #define vp9_sad8x8x4d vp9_sad8x8x4d_c
 
@@ skipping 33 matching lines @@
 unsigned int vp9_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x16 vp9_sub_pixel_avg_variance8x16_c
 
 unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x4 vp9_sub_pixel_avg_variance8x4_c
 
 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 #define vp9_sub_pixel_avg_variance8x8 vp9_sub_pixel_avg_variance8x8_c
 
 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_c
+unsigned int vp9_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance16x16 vp9_sub_pixel_variance16x16_neon
 
 unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x32 vp9_sub_pixel_variance16x32_c
 
 unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance16x8 vp9_sub_pixel_variance16x8_c
 
 unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance32x16 vp9_sub_pixel_variance32x16_c
 
 unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_variance32x32 vp9_sub_pixel_variance32x32_c
+unsigned int vp9_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance32x32 vp9_sub_pixel_variance32x32_neon
 
 unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance32x64 vp9_sub_pixel_variance32x64_c
 
 unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance4x4 vp9_sub_pixel_variance4x4_c
 
 unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance4x8 vp9_sub_pixel_variance4x8_c
 
 unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance64x32 vp9_sub_pixel_variance64x32_c
 
 unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance64x64 vp9_sub_pixel_variance64x64_c
 
 unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance8x16 vp9_sub_pixel_variance8x16_c
 
 unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_sub_pixel_variance8x4 vp9_sub_pixel_variance8x4_c
 
 unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_sub_pixel_variance8x8 vp9_sub_pixel_variance8x8_c
+unsigned int vp9_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_variance8x8 vp9_sub_pixel_variance8x8_neon
 
 void vp9_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
-#define vp9_subtract_block vp9_subtract_block_c
+void vp9_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vp9_subtract_block vp9_subtract_block_neon
 
 void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
 
 void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_tm_predictor_16x16 vp9_tm_predictor_16x16_neon
 
 void vp9_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ skipping 17 matching lines @@
 
 void vp9_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_v_predictor_4x4 vp9_v_predictor_4x4_neon
 
 void vp9_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_v_predictor_8x8 vp9_v_predictor_8x8_neon
 
 unsigned int vp9_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance16x16 vp9_variance16x16_c
+unsigned int vp9_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x16 vp9_variance16x16_neon
 
 unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance16x32 vp9_variance16x32_c
 
 unsigned int vp9_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance16x8 vp9_variance16x8_c
 
 unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance32x16 vp9_variance32x16_c
 
 unsigned int vp9_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance32x32 vp9_variance32x32_c
+unsigned int vp9_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x32 vp9_variance32x32_neon
 
 unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance32x64 vp9_variance32x64_c
 
 unsigned int vp9_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance4x4 vp9_variance4x4_c
 
 unsigned int vp9_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance4x8 vp9_variance4x8_c
 
 unsigned int vp9_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance64x32 vp9_variance64x32_c
 
 unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance64x64 vp9_variance64x64_c
 
 unsigned int vp9_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x16 vp9_variance8x16_c
 
 unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x4 vp9_variance8x4_c
 
 unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance8x8 vp9_variance8x8_c
+unsigned int vp9_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x8 vp9_variance8x8_neon
 
 void vp9_rtcd(void);
 
 #include "vpx_config.h"
 
 #ifdef RTCD_C
 #include "vpx_ports/arm.h"
 static void setup_rtcd_internal(void)
 {
     int flags = arm_cpu_caps();
 
     (void)flags;
 
 }
 #endif
 
 #ifdef __cplusplus
 }  // extern "C"
 #endif
 
 #endif