Add MSA optimized remaining scale row functions

source/scale_msa.cc

 /*
  *  Copyright 2016 The LibYuv 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.
  */
 
@@ skipping 151 matching lines @@
     reg4 = (v8u16)__msa_srari_h((v8i16)reg4, 2);
     reg5 = (v8u16)__msa_srari_h((v8i16)reg5, 2);
     dst0 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
     ST_UB(dst0, dst_argb);
     src_argb += stepx * 4;
     nxt_argb += stepx * 4;
     dst_argb += 16;
   }
 }
 
+void ScaleRowDown2_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Linear_MSA(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0, dst1;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    dst0 = __msa_aver_u_b(vec1, vec0);
+    dst1 = __msa_aver_u_b(vec3, vec2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  const uint8_t* nxt_ptr = src_ptr + src_stride;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)nxt_ptr, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)nxt_ptr, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)nxt_ptr, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)nxt_ptr, 48);
+    vec0 = __msa_hadd_u_h(src0, src0);
+    vec1 = __msa_hadd_u_h(src1, src1);
+    vec2 = __msa_hadd_u_h(src2, src2);
+    vec3 = __msa_hadd_u_h(src3, src3);
+    vec4 = __msa_hadd_u_h(src4, src4);
+    vec5 = __msa_hadd_u_h(src5, src5);
+    vec6 = __msa_hadd_u_h(src6, src6);
+    vec7 = __msa_hadd_u_h(src7, src7);
+    vec0 += vec4;
+    vec1 += vec5;
+    vec2 += vec6;
+    vec3 += vec7;
+    vec0 = (v8u16)__msa_srari_h((v8i16)vec0, 2);
+    vec1 = (v8u16)__msa_srari_h((v8i16)vec1, 2);
+    vec2 = (v8u16)__msa_srari_h((v8i16)vec2, 2);
+    vec3 = (v8u16)__msa_srari_h((v8i16)vec3, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    nxt_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown4_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, dst0;
+
+  for (x = 0; x < dst_width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst);
+    src_ptr += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown4Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  const uint8_t* nxt0_ptr = src_ptr + src_stride;
+  const uint8_t* nxt1_ptr = src_ptr + src_stride * 2;
+  const uint8_t* nxt2_ptr = src_ptr + src_stride * 3;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v4u32 reg0, reg1, reg2, reg3;
+
+  for (x = 0; x < dst_width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)nxt0_ptr, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)nxt0_ptr, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)nxt0_ptr, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)nxt0_ptr, 48);
+    vec0 = __msa_hadd_u_h(src0, src0);
+    vec1 = __msa_hadd_u_h(src1, src1);
+    vec2 = __msa_hadd_u_h(src2, src2);
+    vec3 = __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)nxt1_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)nxt1_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)nxt1_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)nxt1_ptr, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)nxt2_ptr, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)nxt2_ptr, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)nxt2_ptr, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)nxt2_ptr, 48);
+    vec0 += __msa_hadd_u_h(src0, src0);
+    vec1 += __msa_hadd_u_h(src1, src1);
+    vec2 += __msa_hadd_u_h(src2, src2);
+    vec3 += __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    reg0 = __msa_hadd_u_w(vec0, vec0);
+    reg1 = __msa_hadd_u_w(vec1, vec1);
+    reg2 = __msa_hadd_u_w(vec2, vec2);
+    reg3 = __msa_hadd_u_w(vec3, vec3);
+    reg0 = (v4u32)__msa_srari_w((v4i32)reg0, 4);
+    reg1 = (v4u32)__msa_srari_w((v4i32)reg1, 4);
+    reg2 = (v4u32)__msa_srari_w((v4i32)reg2, 4);
+    reg3 = (v4u32)__msa_srari_w((v4i32)reg3, 4);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst);
+    src_ptr += 64;
+    nxt0_ptr += 64;
+    nxt1_ptr += 64;
+    nxt2_ptr += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown38_MSA(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x, width;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, vec0;
+  v16i8 mask = {0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0};
+
+  assert(dst_width % 3 == 0);
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    vec0 = (v16u8)__msa_vshf_b(mask, (v16i8)src1, (v16i8)src0);
+    dst0 = __msa_copy_u_d((v2i64)vec0, 0);
+    dst1 = __msa_copy_u_w((v4i32)vec0, 2);
+    SD(dst0, dst);
+    SW(dst1, dst + 8);
+    src_ptr += 32;
+    dst += 12;
+  }
+}
+
+void ScaleRowDown38_2_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  const uint8_t* nxt_ptr = src_ptr + src_stride;
+  int x, width;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, src2, src3, out;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4u32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  v8i16 zero = {0};
+  v8i16 mask = {0, 1, 2, 8, 3, 4, 5, 9};
+  v16i8 dst_mask = {0, 2, 16, 4, 6, 18, 8, 10, 20, 12, 14, 22, 0, 0, 0, 0};
+  v4u32 const_10922 = (v4u32)__msa_fill_w(10922);
+  v4u32 const_16384 = (v4u32)__msa_fill_w(16384);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)nxt_ptr, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)nxt_ptr, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec0);
+    vec5 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec1);
+    vec6 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec2);
+    vec7 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec3);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    vec1 = (v8u16)__msa_pckod_w((v4i32)vec3, (v4i32)vec2);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    tmp0 = __msa_hadd_u_w(vec4, vec4);
+    tmp1 = __msa_hadd_u_w(vec5, vec5);
+    tmp2 = __msa_hadd_u_w(vec6, vec6);
+    tmp3 = __msa_hadd_u_w(vec7, vec7);
+    tmp4 = __msa_hadd_u_w(vec0, vec0);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    tmp0 = __msa_hadd_u_w(vec0, vec0);
+    tmp1 = __msa_hadd_u_w(vec1, vec1);
+    tmp0 *= const_10922;
+    tmp1 *= const_10922;
+    tmp4 *= const_16384;
+    tmp0 = (v4u32)__msa_srai_w((v4i32)tmp0, 16);
+    tmp1 = (v4u32)__msa_srai_w((v4i32)tmp1, 16);
+    tmp4 = (v4u32)__msa_srai_w((v4i32)tmp4, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp4, (v8i16)tmp4);
+    out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0);
+    dst0 = __msa_copy_u_d((v2i64)out, 0);
+    dst1 = __msa_copy_u_w((v4i32)out, 2);
+    SD(dst0, dst_ptr);
+    SW(dst1, dst_ptr + 8);
+    src_ptr += 32;
+    nxt_ptr += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleRowDown38_3_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, width;
+  const uint8_t* nxt0_ptr = src_ptr + src_stride;
+  const uint8_t* nxt1_ptr = src_ptr + src_stride * 2;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, src2, src3, src4, src5, out;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4u32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  v8u16 zero = {0};
+  v8i16 mask = {0, 1, 2, 8, 3, 4, 5, 9};
+  v16i8 dst_mask = {0, 2, 16, 4, 6, 18, 8, 10, 20, 12, 14, 22, 0, 0, 0, 0};
+  v4u32 const_7281 = (v4u32)__msa_fill_w(7281);
+  v4u32 const_10922 = (v4u32)__msa_fill_w(10922);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)nxt0_ptr, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)nxt0_ptr, 16);
+    src4 = (v16u8)__msa_ld_b((v16i8*)nxt1_ptr, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)nxt1_ptr, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src4);
+    vec5 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src4);
+    vec6 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src5);
+    vec7 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src5);
+    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = __msa_hadd_u_h((v16u8)vec4, (v16u8)vec4);
+    vec5 = __msa_hadd_u_h((v16u8)vec5, (v16u8)vec5);
+    vec6 = __msa_hadd_u_h((v16u8)vec6, (v16u8)vec6);
+    vec7 = __msa_hadd_u_h((v16u8)vec7, (v16u8)vec7);
+    vec0 += vec4;
+    vec1 += vec5;
+    vec2 += vec6;
+    vec3 += vec7;
+    vec4 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec0);
+    vec5 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec1);
+    vec6 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec2);
+    vec7 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec3);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    vec1 = (v8u16)__msa_pckod_w((v4i32)vec3, (v4i32)vec2);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    tmp0 = __msa_hadd_u_w(vec4, vec4);
+    tmp1 = __msa_hadd_u_w(vec5, vec5);
+    tmp2 = __msa_hadd_u_w(vec6, vec6);
+    tmp3 = __msa_hadd_u_w(vec7, vec7);
+    tmp4 = __msa_hadd_u_w(vec0, vec0);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    tmp0 = __msa_hadd_u_w(vec0, vec0);
+    tmp1 = __msa_hadd_u_w(vec1, vec1);
+    tmp0 *= const_7281;
+    tmp1 *= const_7281;
+    tmp4 *= const_10922;
+    tmp0 = (v4u32)__msa_srai_w((v4i32)tmp0, 16);
+    tmp1 = (v4u32)__msa_srai_w((v4i32)tmp1, 16);
+    tmp4 = (v4u32)__msa_srai_w((v4i32)tmp4, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp4, (v8i16)tmp4);
+    out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0);
+    dst0 = __msa_copy_u_d((v2i64)out, 0);
+    dst1 = __msa_copy_u_w((v4i32)out, 2);
+    SD(dst0, dst_ptr);
+    SW(dst1, dst_ptr + 8);
+    src_ptr += 32;
+    nxt0_ptr += 32;
+    nxt1_ptr += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleAddRow_MSA(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
+  int x;
+  v16u8 src0;
+  v8u16 vec0, vec1, dst0, dst1;
+  v16i8 zero = {0};
+
+  assert(src_width > 0);
+
+  for (x = 0; x < src_width; x += 16) {
+    src0 = LD_UB(src_ptr);
+    dst0 = (v8u16)__msa_ld_h((v8i16*)dst_ptr, 0);
+    dst1 = (v8u16)__msa_ld_h((v8i16*)dst_ptr, 16);
+    vec0 = (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b(zero, (v16i8)src0);
+    dst0 += vec0;
+    dst1 += vec1;
+    ST_UH2(dst0, dst1, dst_ptr, 8);
+    src_ptr += 16;
+    dst_ptr += 16;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
 #endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)