[libvpx] Update to v1.12.0

libs/libvpx/vp9/encoder/arm/neon/vp9_denoiser_neon.c

@@ -21,6 +21,9 @@
 
 // Compute the sum of all pixel differences of this MB.
 static INLINE int horizontal_add_s8x16(const int8x16_t v_sum_diff_total) {
+#if defined(__aarch64__)
+  return vaddlvq_s8(v_sum_diff_total);
+#else
   const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff_total);
   const int32x4_t fedc_ba98_7654_3210 = vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
   const int64x2_t fedcba98_76543210 = vpaddlq_s32(fedc_ba98_7654_3210);
@@ -28,6 +31,7 @@ static INLINE int horizontal_add_s8x16(const int8x16_t v_sum_diff_total) {
                                 vget_low_s64(fedcba98_76543210));
   const int sum_diff = vget_lane_s32(vreinterpret_s32_s64(x), 0);
   return sum_diff;
+#endif
 }
 
 // Denoise a 16x1 vector.

libs/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c

@@ -26,12 +26,27 @@
 #include "vpx_dsp/arm/mem_neon.h"
 #include "vpx_dsp/vpx_dsp_common.h"
 
+static INLINE void calculate_dqcoeff_and_store(const int16x8_t qcoeff,
+                                               const int16x8_t dequant,
+                                               tran_low_t *dqcoeff) {
+  const int32x4_t dqcoeff_0 =
+      vmull_s16(vget_low_s16(qcoeff), vget_low_s16(dequant));
+  const int32x4_t dqcoeff_1 =
+      vmull_s16(vget_high_s16(qcoeff), vget_high_s16(dequant));
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  vst1q_s32(dqcoeff, dqcoeff_0);
+  vst1q_s32(dqcoeff + 4, dqcoeff_1);
+#else
+  vst1q_s16(dqcoeff, vcombine_s16(vmovn_s32(dqcoeff_0), vmovn_s32(dqcoeff_1)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
 void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
-                          int skip_block, const int16_t *round_ptr,
-                          const int16_t *quant_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) {
+                          const int16_t *round_ptr, const int16_t *quant_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) {
   // Quantization pass: All coefficients with index >= zero_flag are
   // skippable. Note: zero_flag can be zero.
   int i;
@@ -43,8 +58,6 @@ void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
   int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
 
   (void)scan;
-  (void)skip_block;
-  assert(!skip_block);
 
   // adjust for dc
   v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
@@ -55,7 +68,8 @@ void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
     const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
     const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr);
     const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
-    const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+    const int16x8_t v_abs = vabsq_s16(v_coeff);
+    const int16x8_t v_tmp = vqaddq_s16(v_abs, v_round);
     const int32x4_t v_tmp_lo =
         vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
     const int32x4_t v_tmp_hi =
@@ -67,10 +81,9 @@ void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
     const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
     const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
     const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
-    const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+    calculate_dqcoeff_and_store(v_qcoeff, v_dequant, dqcoeff_ptr);
     v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
     store_s16q_to_tran_low(qcoeff_ptr, v_qcoeff);
-    store_s16q_to_tran_low(dqcoeff_ptr, v_dqcoeff);
     v_round = vmovq_n_s16(round_ptr[1]);
     v_quant = vmovq_n_s16(quant_ptr[1]);
     v_dequant = vmovq_n_s16(dequant_ptr[1]);
@@ -80,7 +93,8 @@ void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
     const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
     const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr + i);
     const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
-    const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+    const int16x8_t v_abs = vabsq_s16(v_coeff);
+    const int16x8_t v_tmp = vqaddq_s16(v_abs, v_round);
     const int32x4_t v_tmp_lo =
         vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
     const int32x4_t v_tmp_hi =
@@ -92,10 +106,9 @@ void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
     const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
     const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
     const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
-    const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+    calculate_dqcoeff_and_store(v_qcoeff, v_dequant, dqcoeff_ptr + i);
     v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
     store_s16q_to_tran_low(qcoeff_ptr + i, v_qcoeff);
-    store_s16q_to_tran_low(dqcoeff_ptr + i, v_dqcoeff);
   }
 #ifdef __aarch64__
   *eob_ptr = vmaxvq_s16(v_eobmax_76543210);
@@ -122,7 +135,7 @@ static INLINE int32x4_t extract_sign_bit(int32x4_t a) {
 }
 
 void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count,
-                                int skip_block, const int16_t *round_ptr,
+                                const int16_t *round_ptr,
                                 const int16_t *quant_ptr,
                                 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
                                 const int16_t *dequant_ptr, uint16_t *eob_ptr,
@@ -146,14 +159,11 @@ void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count,
   const int16x8_t dequant_mask =
       vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, dequant_thresh));
 
-  int16x8_t qcoeff = vaddq_s16(coeff_abs, round);
+  int16x8_t qcoeff = vqaddq_s16(coeff_abs, round);
   int32x4_t dqcoeff_0, dqcoeff_1;
-  int16x8_t dqcoeff;
   uint16x8_t eob_max;
   (void)scan;
   (void)count;
-  (void)skip_block;
-  assert(!skip_block);
 
   // coeff * quant_ptr[]) >> 15
   qcoeff = vqdmulhq_s16(qcoeff, quant);
@@ -170,13 +180,17 @@ void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count,
   // Add 1 if negative to round towards zero because the C uses division.
   dqcoeff_0 = vaddq_s32(dqcoeff_0, extract_sign_bit(dqcoeff_0));
   dqcoeff_1 = vaddq_s32(dqcoeff_1, extract_sign_bit(dqcoeff_1));
-
-  dqcoeff = vcombine_s16(vshrn_n_s32(dqcoeff_0, 1), vshrn_n_s32(dqcoeff_1, 1));
+#if CONFIG_VP9_HIGHBITDEPTH
+  vst1q_s32(dqcoeff_ptr, vshrq_n_s32(dqcoeff_0, 1));
+  vst1q_s32(dqcoeff_ptr + 4, vshrq_n_s32(dqcoeff_1, 1));
+#else
+  store_s16q_to_tran_low(dqcoeff_ptr, vcombine_s16(vshrn_n_s32(dqcoeff_0, 1),
+                                                   vshrn_n_s32(dqcoeff_1, 1)));
+#endif
 
   eob_max = vandq_u16(vtstq_s16(qcoeff, neg_one), v_iscan);
 
   store_s16q_to_tran_low(qcoeff_ptr, qcoeff);
-  store_s16q_to_tran_low(dqcoeff_ptr, dqcoeff);
 
   iscan += 8;
   coeff_ptr += 8;
@@ -200,9 +214,8 @@ void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count,
       const int16x8_t dequant_mask =
           vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, dequant_thresh));
 
-      int16x8_t qcoeff = vaddq_s16(coeff_abs, round);
+      int16x8_t qcoeff = vqaddq_s16(coeff_abs, round);
       int32x4_t dqcoeff_0, dqcoeff_1;
-      int16x8_t dqcoeff;
 
       qcoeff = vqdmulhq_s16(qcoeff, quant);
       qcoeff = veorq_s16(qcoeff, coeff_sign);
@@ -215,14 +228,19 @@ void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count,
       dqcoeff_0 = vaddq_s32(dqcoeff_0, extract_sign_bit(dqcoeff_0));
       dqcoeff_1 = vaddq_s32(dqcoeff_1, extract_sign_bit(dqcoeff_1));
 
-      dqcoeff =
-          vcombine_s16(vshrn_n_s32(dqcoeff_0, 1), vshrn_n_s32(dqcoeff_1, 1));
+#if CONFIG_VP9_HIGHBITDEPTH
+      vst1q_s32(dqcoeff_ptr, vshrq_n_s32(dqcoeff_0, 1));
+      vst1q_s32(dqcoeff_ptr + 4, vshrq_n_s32(dqcoeff_1, 1));
+#else
+      store_s16q_to_tran_low(
+          dqcoeff_ptr,
+          vcombine_s16(vshrn_n_s32(dqcoeff_0, 1), vshrn_n_s32(dqcoeff_1, 1)));
+#endif
 
       eob_max =
           vmaxq_u16(eob_max, vandq_u16(vtstq_s16(qcoeff, neg_one), v_iscan));
 
       store_s16q_to_tran_low(qcoeff_ptr, qcoeff);
-      store_s16q_to_tran_low(dqcoeff_ptr, dqcoeff);
 
       iscan += 8;
       coeff_ptr += 8;