Add MSA optimized NV12/21 To RGB row functions

source/row_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 2005 matching lines @@
     res1 = __msa_copy_u_d((v2i64)dst0, 1);
     SD(res0, dst_u);
     SD(res1, dst_v);
     t += 48;
     s += 48;
     dst_u += 8;
     dst_v += 8;
   }
 }
 
+void NV12ToARGBRow_MSA(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  uint64 val0, val1;
+  v16u8 src0, src1, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3;
+  v4i32 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v4i32 vec_ub, vec_vr, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg;
+  v16u8 zero = {0};
+  v16u8 const_255 = (v16u8)__msa_ldi_b(255);
+  v4i32 max = __msa_ldi_w(255);
+
+  vec_ub = __msa_fill_w(yuvconstants->kUVToB[0]);
+  vec_vr = __msa_fill_w(yuvconstants->kUVToR[1]);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg =
+      __msa_fill_w((yuvconstants->kUVToG[0]) | (yuvconstants->kUVToG[1] << 16));
+  vec_bb = __msa_fill_w(yuvconstants->kUVBiasB[0]);
+  vec_bg = __msa_fill_w(yuvconstants->kUVBiasG[0]);
+  vec_br = __msa_fill_w(yuvconstants->kUVBiasR[0]);
+  vec_yg = __msa_fill_w(yuvconstants->kYToRgb[0]);
+
+  for (x = 0; x < width; x += 8) {
+    val0 = LD(src_y);
+    val1 = LD(src_uv);
+    src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
+    src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
+    reg0 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
+    reg1 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
+    reg2 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec1);
+    reg3 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec1);

[fbarchard1, 2017/01/13 18:46:01]

consider macros to read nv12, do yuvtorgb, and write argb.
these yuv to rgb functions have many variations, so it will reduce code size /
ease maintenance to share code.

[manojkumar.bhosale, 2017/01/17 10:29:45]

Done.

+    reg0 *= vec_yg;
+    reg1 *= vec_yg;
+    reg2 *= vec_ubvr;
+    reg3 *= vec_ubvr;
+    reg0 = __msa_srai_w(reg0, 16);
+    reg1 = __msa_srai_w(reg1, 16);
+    reg4 = __msa_dotp_s_w((v8i16)vec1, (v8i16)vec_ugvg);
+    reg5 = __msa_ilvev_w(reg2, reg2);
+    reg6 = __msa_ilvev_w(reg3, reg3);
+    reg7 = __msa_ilvr_w(reg4, reg4);
+    reg2 = __msa_ilvod_w(reg2, reg2);
+    reg3 = __msa_ilvod_w(reg3, reg3);
+    reg4 = __msa_ilvl_w(reg4, reg4);
+    reg5 = reg0 - reg5;
+    reg6 = reg1 - reg6;
+    reg2 = reg0 - reg2;
+    reg3 = reg1 - reg3;
+    reg7 = reg0 - reg7;
+    reg4 = reg1 - reg4;
+    reg5 += vec_bb;
+    reg6 += vec_bb;
+    reg7 += vec_bg;
+    reg4 += vec_bg;
+    reg2 += vec_br;
+    reg3 += vec_br;
+    reg5 = __msa_srai_w(reg5, 6);
+    reg6 = __msa_srai_w(reg6, 6);
+    reg7 = __msa_srai_w(reg7, 6);
+    reg4 = __msa_srai_w(reg4, 6);
+    reg2 = __msa_srai_w(reg2, 6);
+    reg3 = __msa_srai_w(reg3, 6);
+    reg5 = __msa_maxi_s_w(reg5, 0);

[fbarchard1, 2017/01/13 18:46:01]

no free way to pack words to bytes with saturation?

[manojkumar.bhosale, 2017/01/17 10:29:45]

No

+    reg6 = __msa_maxi_s_w(reg6, 0);
+    reg7 = __msa_maxi_s_w(reg7, 0);
+    reg4 = __msa_maxi_s_w(reg4, 0);
+    reg2 = __msa_maxi_s_w(reg2, 0);
+    reg3 = __msa_maxi_s_w(reg3, 0);
+    reg5 = __msa_min_s_w(max, reg5);
+    reg6 = __msa_min_s_w(max, reg6);
+    reg7 = __msa_min_s_w(max, reg7);
+    reg4 = __msa_min_s_w(max, reg4);
+    reg2 = __msa_min_s_w(max, reg2);
+    reg3 = __msa_min_s_w(max, reg3);
+    vec0 = (v8u16)__msa_ilvev_b((v16i8)reg7, (v16i8)reg5);
+    vec1 = (v8u16)__msa_ilvev_b((v16i8)const_255, (v16i8)reg2);
+    vec2 = (v8u16)__msa_ilvev_b((v16i8)reg4, (v16i8)reg6);
+    vec3 = (v8u16)__msa_ilvev_b((v16i8)const_255, (v16i8)reg3);
+    dst0 = (v16u8)__msa_ilvev_h((v8i16)vec1, (v8i16)vec0);
+    dst1 = (v16u8)__msa_ilvev_h((v8i16)vec3, (v8i16)vec2);
+    ST_UB2(dst0, dst1, rgb_buf, 16);
+    src_y += 8;
+    src_uv += 8;
+    rgb_buf += 32;
+  }
+}
+
+void NV12ToRGB565Row_MSA(const uint8* src_y,
+                         const uint8* src_uv,
+                         uint8* rgb_buf,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  int x;
+  uint64 val0, val1;
+  v16u8 src0, src1, dst0;
+  v8u16 vec0, vec1, vec2;
+  v4i32 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v4i32 vec_ub, vec_vr, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg;
+  v16u8 zero = {0};
+  v4i32 max = __msa_ldi_w(255);
+
+  vec_ub = __msa_fill_w(yuvconstants->kUVToB[0]);
+  vec_vr = __msa_fill_w(yuvconstants->kUVToR[1]);
+  vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
+  vec_ugvg =
+      __msa_fill_w((yuvconstants->kUVToG[0]) | (yuvconstants->kUVToG[1] << 16));
+  vec_bb = __msa_fill_w(yuvconstants->kUVBiasB[0]);
+  vec_bg = __msa_fill_w(yuvconstants->kUVBiasG[0]);
+  vec_br = __msa_fill_w(yuvconstants->kUVBiasR[0]);
+  vec_yg = __msa_fill_w(yuvconstants->kYToRgb[0]);
+
+  for (x = 0; x < width; x += 8) {
+    val0 = LD(src_y);
+    val1 = LD(src_uv);
+    src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
+    src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
+    reg0 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
+    reg1 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
+    reg2 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec1);
+    reg3 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec1);
+    reg0 *= vec_yg;
+    reg1 *= vec_yg;
+    reg2 *= vec_ubvr;
+    reg3 *= vec_ubvr;
+    reg0 = __msa_srai_w(reg0, 16);
+    reg1 = __msa_srai_w(reg1, 16);
+    reg4 = __msa_dotp_s_w((v8i16)vec1, (v8i16)vec_ugvg);
+    reg5 = __msa_ilvev_w(reg2, reg2);
+    reg6 = __msa_ilvev_w(reg3, reg3);
+    reg7 = __msa_ilvr_w(reg4, reg4);
+    reg2 = __msa_ilvod_w(reg2, reg2);
+    reg3 = __msa_ilvod_w(reg3, reg3);
+    reg4 = __msa_ilvl_w(reg4, reg4);
+    reg5 = reg0 - reg5;
+    reg6 = reg1 - reg6;
+    reg2 = reg0 - reg2;
+    reg3 = reg1 - reg3;
+    reg7 = reg0 - reg7;
+    reg4 = reg1 - reg4;
+    reg5 += vec_bb;
+    reg6 += vec_bb;
+    reg7 += vec_bg;
+    reg4 += vec_bg;
+    reg2 += vec_br;
+    reg3 += vec_br;
+    reg5 = __msa_srai_w(reg5, 6);
+    reg6 = __msa_srai_w(reg6, 6);
+    reg7 = __msa_srai_w(reg7, 6);
+    reg4 = __msa_srai_w(reg4, 6);
+    reg2 = __msa_srai_w(reg2, 6);
+    reg3 = __msa_srai_w(reg3, 6);
+    reg5 = __msa_maxi_s_w(reg5, 0);
+    reg6 = __msa_maxi_s_w(reg6, 0);
+    reg7 = __msa_maxi_s_w(reg7, 0);
+    reg4 = __msa_maxi_s_w(reg4, 0);
+    reg2 = __msa_maxi_s_w(reg2, 0);
+    reg3 = __msa_maxi_s_w(reg3, 0);
+    reg5 = __msa_min_s_w(max, reg5);
+    reg6 = __msa_min_s_w(max, reg6);
+    reg7 = __msa_min_s_w(max, reg7);
+    reg4 = __msa_min_s_w(max, reg4);
+    reg2 = __msa_min_s_w(max, reg2);
+    reg3 = __msa_min_s_w(max, reg3);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg6, (v8i16)reg5);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg4, (v8i16)reg7);
+    vec2 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    vec0 = vec0 >> 3;
+    vec1 = (vec1 >> 2) << 5;
+    vec2 = (vec2 >> 3) << 11;
+    dst0 = (v16u8)(vec0 | vec1 | vec2);
+    ST_UB(dst0, rgb_buf);
+    src_y += 8;
+    src_uv += 8;
+    rgb_buf += 16;
+  }
+}
+
+void NV21ToARGBRow_MSA(const uint8* src_y,
+                       const uint8* src_vu,
+                       uint8* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  uint64 val0, val1;
+  v16u8 src0, src1, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3;
+  v4i32 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v4i32 vec_ub, vec_vr, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg;
+  v16u8 zero = {0};
+  v16u8 const_255 = (v16u8)__msa_ldi_b(255);
+  v4i32 max = __msa_ldi_w(255);
+
+  vec_ub = __msa_fill_w(yuvconstants->kUVToB[0]);
+  vec_vr = __msa_fill_w(yuvconstants->kUVToR[1]);
+  vec_ubvr = __msa_ilvr_w(vec_ub, vec_vr);
+  vec_ugvg =
+      __msa_fill_w((yuvconstants->kUVToG[1] | (yuvconstants->kUVToG[0] << 16)));
+  vec_bb = __msa_fill_w(yuvconstants->kUVBiasB[0]);
+  vec_bg = __msa_fill_w(yuvconstants->kUVBiasG[0]);
+  vec_br = __msa_fill_w(yuvconstants->kUVBiasR[0]);
+  vec_yg = __msa_fill_w(yuvconstants->kYToRgb[0]);
+
+  for (x = 0; x < width; x += 8) {
+    val0 = LD(src_y);
+    val1 = LD(src_vu);
+    src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
+    src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
+    reg0 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
+    reg1 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
+    reg2 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec1);
+    reg3 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec1);
+    reg0 *= vec_yg;
+    reg1 *= vec_yg;
+    reg2 *= vec_ubvr;
+    reg3 *= vec_ubvr;
+    reg0 = __msa_srai_w(reg0, 16);
+    reg1 = __msa_srai_w(reg1, 16);
+    reg4 = __msa_dotp_s_w((v8i16)vec1, (v8i16)vec_ugvg);
+    reg5 = __msa_ilvod_w(reg2, reg2);
+    reg6 = __msa_ilvod_w(reg3, reg3);
+    reg7 = __msa_ilvr_w(reg4, reg4);
+    reg2 = __msa_ilvev_w(reg2, reg2);
+    reg3 = __msa_ilvev_w(reg3, reg3);
+    reg4 = __msa_ilvl_w(reg4, reg4);
+    reg5 = reg0 - reg5;
+    reg6 = reg1 - reg6;
+    reg2 = reg0 - reg2;
+    reg3 = reg1 - reg3;
+    reg7 = reg0 - reg7;
+    reg4 = reg1 - reg4;
+    reg5 += vec_bb;
+    reg6 += vec_bb;
+    reg7 += vec_bg;
+    reg4 += vec_bg;
+    reg2 += vec_br;
+    reg3 += vec_br;
+    reg5 = __msa_srai_w(reg5, 6);
+    reg6 = __msa_srai_w(reg6, 6);
+    reg7 = __msa_srai_w(reg7, 6);
+    reg4 = __msa_srai_w(reg4, 6);
+    reg2 = __msa_srai_w(reg2, 6);
+    reg3 = __msa_srai_w(reg3, 6);
+    reg5 = __msa_maxi_s_w(reg5, 0);
+    reg6 = __msa_maxi_s_w(reg6, 0);
+    reg7 = __msa_maxi_s_w(reg7, 0);
+    reg4 = __msa_maxi_s_w(reg4, 0);
+    reg2 = __msa_maxi_s_w(reg2, 0);
+    reg3 = __msa_maxi_s_w(reg3, 0);
+    reg5 = __msa_min_s_w(max, reg5);
+    reg6 = __msa_min_s_w(max, reg6);
+    reg7 = __msa_min_s_w(max, reg7);
+    reg4 = __msa_min_s_w(max, reg4);
+    reg2 = __msa_min_s_w(max, reg2);
+    reg3 = __msa_min_s_w(max, reg3);
+    vec0 = (v8u16)__msa_ilvev_b((v16i8)reg7, (v16i8)reg5);
+    vec1 = (v8u16)__msa_ilvev_b((v16i8)const_255, (v16i8)reg2);
+    vec2 = (v8u16)__msa_ilvev_b((v16i8)reg4, (v16i8)reg6);
+    vec3 = (v8u16)__msa_ilvev_b((v16i8)const_255, (v16i8)reg3);
+    dst0 = (v16u8)__msa_ilvev_h((v8i16)vec1, (v8i16)vec0);
+    dst1 = (v16u8)__msa_ilvev_h((v8i16)vec3, (v8i16)vec2);
+    ST_UB2(dst0, dst1, rgb_buf, 16);
+    src_y += 8;
+    src_vu += 8;
+    rgb_buf += 32;
+  }
+}
+
+void SobelRow_MSA(const uint8* src_sobelx,
+                  const uint8* src_sobely,
+                  uint8* dst_argb,
+                  int width) {
+  int x;
+  v16u8 src0, src1, vec0, dst0, dst1, dst2, dst3;
+  v16i8 mask0 = {0, 0, 0, 16, 1, 1, 1, 16, 2, 2, 2, 16, 3, 3, 3, 16};
+  v16i8 const_0x4 = __msa_ldi_b(0x4);
+  v16i8 mask1 = mask0 + const_0x4;
+  v16i8 mask2 = mask1 + const_0x4;
+  v16i8 mask3 = mask2 + const_0x4;
+  v16u8 const_0xFF = (v16u8)__msa_ldi_b(0xFF);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 0);
+    vec0 = __msa_adds_u_b(src0, src1);
+    dst0 = (v16u8)__msa_vshf_b(mask0, (v16i8)const_0xFF, (v16i8)vec0);
+    dst1 = (v16u8)__msa_vshf_b(mask1, (v16i8)const_0xFF, (v16i8)vec0);
+    dst2 = (v16u8)__msa_vshf_b(mask2, (v16i8)const_0xFF, (v16i8)vec0);
+    dst3 = (v16u8)__msa_vshf_b(mask3, (v16i8)const_0xFF, (v16i8)vec0);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void SobelToPlaneRow_MSA(const uint8* src_sobelx,
+                         const uint8* src_sobely,
+                         uint8* dst_y,
+                         int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 16);
+    dst0 = __msa_adds_u_b(src0, src2);
+    dst1 = __msa_adds_u_b(src1, src3);
+    ST_UB2(dst0, dst1, dst_y, 16);
+    src_sobelx += 32;
+    src_sobely += 32;
+    dst_y += 32;
+  }
+}
+
+void SobelXYRow_MSA(const uint8* src_sobelx,
+                    const uint8* src_sobely,
+                    uint8* dst_argb,
+                    int width) {
+  int x;
+  v16u8 src0, src1, vec0, vec1, vec2;
+  v16u8 reg0, reg1, dst0, dst1, dst2, dst3;
+  v16u8 const_0xFF = (v16u8)__msa_ldi_b(0xFF);
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 0);
+    vec0 = __msa_adds_u_b(src0, src1);
+    vec1 = (v16u8)__msa_ilvr_b((v16i8)src0, (v16i8)src1);
+    vec2 = (v16u8)__msa_ilvl_b((v16i8)src0, (v16i8)src1);
+    reg0 = (v16u8)__msa_ilvr_b((v16i8)const_0xFF, (v16i8)vec0);
+    reg1 = (v16u8)__msa_ilvl_b((v16i8)const_0xFF, (v16i8)vec0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)reg0, (v16i8)vec1);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)reg0, (v16i8)vec1);
+    dst2 = (v16u8)__msa_ilvr_b((v16i8)reg1, (v16i8)vec2);
+    dst3 = (v16u8)__msa_ilvl_b((v16i8)reg1, (v16i8)vec2);
+    ST_UB4(dst0, dst1, dst2, dst3, dst_argb, 16);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
 #endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)