shithub: libvpx

Info  •  Files  •  Log  •  Branches

Fix vpx_plane_add_noise_msa functionality bit-mismatch

Change-Id: I04961afb592ae6a67fdcfd8c9066e920dd4b30e7

--- a/AUTHORS

+++ b/AUTHORS

@@ -75,6 +75,7 @@

 Julia Robson <juliamrobson@gmail.com>

 Justin Clift <justin@salasaga.org>

 Justin Lebar <justin.lebar@gmail.com>

+Kaustubh Raste <kaustubh.raste@imgtec.com>

 KO Myung-Hun <komh@chollian.net>

 Lawrence Velázquez <larryv@macports.org>

 Linfeng Zhang <linfengz@google.com>

--- a/vpx_dsp/mips/add_noise_msa.c

+++ b/vpx_dsp/mips/add_noise_msa.c

@@ -15,7 +15,14 @@

                              int blackclamp, int whiteclamp, int width,

                              int height, int32_t pitch) {

   int i, j;

+  v16u8 pos0, pos1, ref0, ref1;

+  v16i8 black_clamp, white_clamp, both_clamp;

+  black_clamp = __msa_fill_b(blackclamp);

+  white_clamp = __msa_fill_b(whiteclamp);

+  both_clamp = black_clamp + white_clamp;

+  both_clamp = -both_clamp;

+

   for (i = 0; i < height / 2; ++i) {

     uint8_t *pos0_ptr = start_ptr + (2 * i) * pitch;

     const int8_t *ref0_ptr = noise + (rand() & 0xff);

@@ -22,29 +29,16 @@

     uint8_t *pos1_ptr = start_ptr + (2 * i + 1) * pitch;

     const int8_t *ref1_ptr = noise + (rand() & 0xff);

     for (j = width / 16; j--;) {

-      v16i8 temp00_s, temp01_s;

-      v16u8 temp00, temp01, black_clamp, white_clamp;

-      v16u8 pos0, ref0, pos1, ref1;

-      v16i8 const127 = __msa_ldi_b(127);

-

       pos0 = LD_UB(pos0_ptr);

       ref0 = LD_UB(ref0_ptr);

       pos1 = LD_UB(pos1_ptr);

       ref1 = LD_UB(ref1_ptr);

-      black_clamp = (v16u8)__msa_fill_b(blackclamp);

-      white_clamp = (v16u8)__msa_fill_b(whiteclamp);

-      temp00 = (pos0 < black_clamp);

-      pos0 = __msa_bmnz_v(pos0, black_clamp, temp00);

-      temp01 = (pos1 < black_clamp);

-      pos1 = __msa_bmnz_v(pos1, black_clamp, temp01);

-      XORI_B2_128_UB(pos0, pos1);

-      temp00_s = __msa_adds_s_b((v16i8)white_clamp, const127);

-      temp00 = (v16u8)(temp00_s < pos0);

-      pos0 = (v16u8)__msa_bmnz_v((v16u8)pos0, (v16u8)temp00_s, temp00);

-      temp01_s = __msa_adds_s_b((v16i8)white_clamp, const127);

-      temp01 = (temp01_s < pos1);

-      pos1 = (v16u8)__msa_bmnz_v((v16u8)pos1, (v16u8)temp01_s, temp01);

-      XORI_B2_128_UB(pos0, pos1);

+      pos0 = __msa_subsus_u_b(pos0, black_clamp);

+      pos1 = __msa_subsus_u_b(pos1, black_clamp);

+      pos0 = __msa_subsus_u_b(pos0, both_clamp);

+      pos1 = __msa_subsus_u_b(pos1, both_clamp);

+      pos0 = __msa_subsus_u_b(pos0, white_clamp);

+      pos1 = __msa_subsus_u_b(pos1, white_clamp);

       pos0 += ref0;

       ST_UB(pos0, pos0_ptr);

       pos1 += ref1;