mips msa vp9 loopfilter 4, 8 optimization

average improvement ~3x-4x

Change-Id: I59279293ce4b2a1e99bd10579ac97740e943643f

test/lpf_8_test.cc

@@ -694,9 +694,23 @@ INSTANTIATE_TEST_CASE_P(
 INSTANTIATE_TEST_CASE_P(
     MSA, Loop8Test6Param,
     ::testing::Values(
+        make_tuple(&vp9_lpf_horizontal_8_msa, &vp9_lpf_horizontal_8_c, 8, 1),
         make_tuple(&vp9_lpf_horizontal_16_msa, &vp9_lpf_horizontal_16_c, 8, 1),
         make_tuple(&vp9_lpf_horizontal_16_msa, &vp9_lpf_horizontal_16_c, 8, 2),
+        make_tuple(&vp9_lpf_vertical_8_msa, &vp9_lpf_vertical_8_c, 8, 1),
         make_tuple(&wrapper_vertical_16_msa, &wrapper_vertical_16_c, 8, 1)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, Loop8Test9Param,
+    ::testing::Values(
+        make_tuple(&vp9_lpf_horizontal_4_dual_msa,
+                   &vp9_lpf_horizontal_4_dual_c, 8),
+        make_tuple(&vp9_lpf_horizontal_8_dual_msa,
+                   &vp9_lpf_horizontal_8_dual_c, 8),
+        make_tuple(&vp9_lpf_vertical_4_dual_msa,
+                   &vp9_lpf_vertical_4_dual_c, 8),
+        make_tuple(&vp9_lpf_vertical_8_dual_msa,
+                   &vp9_lpf_vertical_8_dual_c, 8)));
 #endif  // HAVE_MSA && (!CONFIG_VP9_HIGHBITDEPTH)
 
 }  // namespace

vp9/common/mips/msa/vp9_loopfilter_4_msa.c

+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/mips/msa/vp9_loopfilter_msa.h"
+
+void vp9_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch,
+                              const uint8_t *b_limit_ptr,
+                              const uint8_t *limit_ptr,
+                              const uint8_t *thresh_ptr,
+                              int32_t count) {
+  uint64_t p1_d, p0_d, q0_d, q1_d;
+  v16u8 mask, hev, flat, thresh, b_limit, limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p1_out, p0_out, q0_out, q1_out;
+
+  (void)count;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+  p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+  q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+  q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+  SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+}
+
+void vp9_lpf_horizontal_4_dual_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit0_ptr,
+                                   const uint8_t *limit0_ptr,
+                                   const uint8_t *thresh0_ptr,
+                                   const uint8_t *b_limit1_ptr,
+                                   const uint8_t *limit1_ptr,
+                                   const uint8_t *thresh1_ptr) {
+  v16u8 mask, hev, flat, thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+  thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+  thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+  b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+  b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+  b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+  limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+  limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+  limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+
+  ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch);
+}
+
+void vp9_lpf_vertical_4_msa(uint8_t *src, int32_t pitch,
+                            const uint8_t *b_limit_ptr,
+                            const uint8_t *limit_ptr,
+                            const uint8_t *thresh_ptr,
+                            int32_t count) {
+  v16u8 mask, hev, flat, limit, thresh, b_limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v8i16 vec0, vec1, vec2, vec3;
+
+  (void)count;
+
+  LD_UB8((src - 4), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
+                     p3, p2, p1, p0, q0, q1, q2, q3);
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+  ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1);
+  ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+  src -= 2;
+  ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+  src += 4 * pitch;
+  ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+}
+
+void vp9_lpf_vertical_4_dual_msa(uint8_t *src, int32_t pitch,
+                                 const uint8_t *b_limit0_ptr,
+                                 const uint8_t *limit0_ptr,
+                                 const uint8_t *thresh0_ptr,
+                                 const uint8_t *b_limit1_ptr,
+                                 const uint8_t *limit1_ptr,
+                                 const uint8_t *thresh1_ptr) {
+  v16u8 mask, hev, flat;
+  v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+  LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+  LD_UB8(src - 4 + (8 * pitch), pitch,
+         row8, row9, row10, row11, row12, row13, row14, row15);
+
+  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                      row8, row9, row10, row11, row12, row13, row14, row15,
+                      p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+  thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+  thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+  b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+  b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+  b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+  limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+  limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+  limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+  ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+  ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
+  ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+  ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
+
+  src -= 2;
+
+  ST4x8_UB(tmp2, tmp3, src, pitch);
+  src += (8 * pitch);
+  ST4x8_UB(tmp4, tmp5, src, pitch);
+}

vp9/common/mips/msa/vp9_loopfilter_8_msa.c

+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/mips/msa/vp9_loopfilter_msa.h"
+
+void vp9_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch,
+                              const uint8_t *b_limit_ptr,
+                              const uint8_t *limit_ptr,
+                              const uint8_t *thresh_ptr,
+                              int32_t count) {
+  uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+  v16u8 mask, hev, flat, thresh, b_limit, limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v8i16 p2_filter8, p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+  v16i8 zero = { 0 };
+
+  (void)count;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+    SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+               q2_r, q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+                p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
+                zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
+                q0_filter8);
+    PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+    p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+    p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+    q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+    src -= 3 * pitch;
+
+    SD4(p2_d, p1_d, p0_d, q0_d, src, pitch);
+    src += (4 * pitch);
+    SD(q1_d, src);
+    src += pitch;
+    SD(q2_d, src);
+  }
+}
+
+void vp9_lpf_horizontal_8_dual_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit0,
+                                   const uint8_t *limit0,
+                                   const uint8_t *thresh0,
+                                   const uint8_t *b_limit1,
+                                   const uint8_t *limit1,
+                                   const uint8_t *thresh1) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, tmp, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh0);
+  tmp = (v16u8)__msa_fill_b(*thresh1);
+  thresh = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)thresh);
+
+  b_limit = (v16u8)__msa_fill_b(*b_limit0);
+  tmp = (v16u8)__msa_fill_b(*b_limit1);
+  b_limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)b_limit);
+
+  limit = (v16u8)__msa_fill_b(*limit0);
+  tmp = (v16u8)__msa_fill_b(*limit1);
+  limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)limit);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+               q2_r, q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    src -= 3 * pitch;
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, src, pitch);
+    src += (4 * pitch);
+    ST_UB2(q1_out, q2_out, src, pitch);
+    src += (2 * pitch);
+  }
+}
+
+void vp9_lpf_vertical_8_msa(uint8_t *src, int32_t pitch,
+                            const uint8_t *b_limit_ptr,
+                            const uint8_t *limit_ptr,
+                            const uint8_t *thresh_ptr,
+                            int32_t count) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p1_out, p0_out, q0_out, q1_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v16u8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4;
+
+  (void)count;
+
+  /* load vector elements */
+  LD_UB8(src - 4, pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
+                     p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  /* flat4 */
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    /* Store 4 pixels p1-_q1 */
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+    src -= 2;
+    ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+    src += 4 * pitch;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+               q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_r, p2_filt8_r, p1_filt8_r, p1_filt8_r, p0_filt8_r,
+                p0_filt8_r, q0_filt8_r, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_r, q1_filt8_r, q2_filt8_r, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    /* Store 6 pixels p2-_q2 */
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    vec4 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+    src -= 3;
+    ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec4, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec4, 4, src + 4, pitch);
+  }
+}
+
+void vp9_lpf_vertical_8_dual_msa(uint8_t *src, int32_t pitch,
+                                 const uint8_t *b_limit0,
+                                 const uint8_t *limit0,
+                                 const uint8_t *thresh0,
+                                 const uint8_t *b_limit1,
+                                 const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  uint8_t *temp_src;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p1_out, p0_out, q0_out, q1_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v16u8 row4, row5, row6, row7, row12, row13, row14, row15;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+  temp_src = src - 4;
+
+  LD_UB8(temp_src, pitch, p0, p1, p2, p3, row4, row5, row6, row7);
+  temp_src += (8 * pitch);
+  LD_UB8(temp_src, pitch, q3, q2, q1, q0, row12, row13, row14, row15);
+
+  /* transpose 16x8 matrix into 8x16 */
+  TRANSPOSE16x8_UB_UB(p0, p1, p2, p3, row4, row5, row6, row7,
+                      q3, q2, q1, q0, row12, row13, row14, row15,
+                      p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh0);
+  vec0 = (v8i16)__msa_fill_b(*thresh1);
+  thresh = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)thresh);
+
+  b_limit = (v16u8)__msa_fill_b(*b_limit0);
+  vec0 = (v8i16)__msa_fill_b(*b_limit1);
+  b_limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)b_limit);
+
+  limit = (v16u8)__msa_fill_b(*limit0);
+  vec0 = (v8i16)__msa_fill_b(*limit1);
+  limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)limit);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  /* flat4 */
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+    src -= 2;
+    ST4x8_UB(vec2, vec3, src, pitch);
+    src += 8 * pitch;
+    ST4x8_UB(vec4, vec5, src, pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+               q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+
+    /* filter8 */
+    VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+    ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+    src -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec2, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec2, 4, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec5, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec5, 4, src + 4, pitch);
+  }
+}

vp9/common/vp9_rtcd_defs.pl

@@ -224,36 +224,36 @@ specialize qw/vp9_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
 $vp9_lpf_vertical_16_dual_neon_asm=vp9_lpf_vertical_16_dual_neon;
 
 add_proto qw/void vp9_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_vertical_8 sse2 neon_asm dspr2/;
+specialize qw/vp9_lpf_vertical_8 sse2 neon_asm dspr2 msa/;
 $vp9_lpf_vertical_8_neon_asm=vp9_lpf_vertical_8_neon;
 
 add_proto qw/void vp9_lpf_vertical_8_dual/, "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";
-specialize qw/vp9_lpf_vertical_8_dual sse2 neon_asm dspr2/;
+specialize qw/vp9_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
 $vp9_lpf_vertical_8_dual_neon_asm=vp9_lpf_vertical_8_dual_neon;
 
 add_proto qw/void vp9_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_vertical_4 mmx neon dspr2/;
+specialize qw/vp9_lpf_vertical_4 mmx neon dspr2 msa/;
 
 add_proto qw/void vp9_lpf_vertical_4_dual/, "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";
-specialize qw/vp9_lpf_vertical_4_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_vertical_4_dual sse2 neon dspr2 msa/;
 
 add_proto qw/void vp9_lpf_horizontal_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
 specialize qw/vp9_lpf_horizontal_16 sse2 avx2 neon_asm dspr2 msa/;
 $vp9_lpf_horizontal_16_neon_asm=vp9_lpf_horizontal_16_neon;
 
 add_proto qw/void vp9_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_8 sse2 neon_asm dspr2/;
+specialize qw/vp9_lpf_horizontal_8 sse2 neon_asm dspr2 msa/;
 $vp9_lpf_horizontal_8_neon_asm=vp9_lpf_horizontal_8_neon;
 
 add_proto qw/void vp9_lpf_horizontal_8_dual/, "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";
-specialize qw/vp9_lpf_horizontal_8_dual sse2 neon_asm dspr2/;
+specialize qw/vp9_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
 $vp9_lpf_horizontal_8_dual_neon_asm=vp9_lpf_horizontal_8_dual_neon;
 
 add_proto qw/void vp9_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_4 mmx neon dspr2/;
+specialize qw/vp9_lpf_horizontal_4 mmx neon dspr2 msa/;
 
 add_proto qw/void vp9_lpf_horizontal_4_dual/, "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";
-specialize qw/vp9_lpf_horizontal_4_dual sse2 neon dspr2/;
+specialize qw/vp9_lpf_horizontal_4_dual sse2 neon dspr2 msa/;