Add mips msa vpx Integer projection row/col functions

average improvement ~4x-5x

Change-Id: I17c41383250282b39f5ecae0197ef1df7de20801

diff --git a/test/avg_test.cc b/test/avg_test.cc
index 6e416eccaaf061bd86d2ea05246cc80d6c371f5a..bf88d83dbd0013b70b8539f65a761fa37ae553b2 100644 (file)
--- a/test/avg_test.cc
+++ b/test/avg_test.cc
@@ -393,6 +393,20 @@ INSTANTIATE_TEST_CASE_P(
                       make_tuple(16, 16, 5, 4, &vpx_avg_4x4_msa),
                       make_tuple(32, 32, 15, 4, &vpx_avg_4x4_msa)));
 
+INSTANTIATE_TEST_CASE_P(
+    MSA, IntProRowTest,
+    ::testing::Values(make_tuple(16, &vpx_int_pro_row_msa, &vpx_int_pro_row_c),
+                      make_tuple(32, &vpx_int_pro_row_msa, &vpx_int_pro_row_c),
+                      make_tuple(64, &vpx_int_pro_row_msa,
+                                 &vpx_int_pro_row_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, IntProColTest,
+    ::testing::Values(make_tuple(16, &vpx_int_pro_col_msa, &vpx_int_pro_col_c),
+                      make_tuple(32, &vpx_int_pro_col_msa, &vpx_int_pro_col_c),
+                      make_tuple(64, &vpx_int_pro_col_msa,
+                                 &vpx_int_pro_col_c)));
+
 INSTANTIATE_TEST_CASE_P(MSA, SatdTest,
                         ::testing::Values(make_tuple(16, &vpx_satd_msa),
                                           make_tuple(64, &vpx_satd_msa),

diff --git a/vpx_dsp/mips/avg_msa.c b/vpx_dsp/mips/avg_msa.c
index 9e786b69edb88aeaef48a334f88043cb963e3c08..3023d4377d13e8cc591ba8af76529163002e0da0 100644 (file)
--- a/vpx_dsp/mips/avg_msa.c
+++ b/vpx_dsp/mips/avg_msa.c
@@ -389,3 +389,175 @@ int vpx_satd_msa(const int16_t *data, int length) {
 
   return satd;
 }
+
+void vpx_int_pro_row_msa(int16_t hbuf[16], const uint8_t *ref,
+                         const int ref_stride, const int height) {
+  int i;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v8i16 hbuf_r = { 0 };
+  v8i16 hbuf_l = { 0 };
+  v8i16 ref0_r, ref0_l, ref1_r, ref1_l, ref2_r, ref2_l, ref3_r, ref3_l;
+  v8i16 ref4_r, ref4_l, ref5_r, ref5_l, ref6_r, ref6_l, ref7_r, ref7_l;
+
+  if (16 == height) {
+    for (i = 2; i--;) {
+      LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
+      ref += 8 * ref_stride;
+      UNPCK_UB_SH(ref0, ref0_r, ref0_l);
+      UNPCK_UB_SH(ref1, ref1_r, ref1_l);
+      UNPCK_UB_SH(ref2, ref2_r, ref2_l);
+      UNPCK_UB_SH(ref3, ref3_r, ref3_l);
+      UNPCK_UB_SH(ref4, ref4_r, ref4_l);
+      UNPCK_UB_SH(ref5, ref5_r, ref5_l);
+      UNPCK_UB_SH(ref6, ref6_r, ref6_l);
+      UNPCK_UB_SH(ref7, ref7_r, ref7_l);
+      ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+    }
+
+    SRA_2V(hbuf_r, hbuf_l, 3);
+    ST_SH2(hbuf_r, hbuf_l, hbuf, 8);
+  } else if (32 == height) {
+    for (i = 2; i--;) {
+      LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
+      ref += 8 * ref_stride;
+      UNPCK_UB_SH(ref0, ref0_r, ref0_l);
+      UNPCK_UB_SH(ref1, ref1_r, ref1_l);
+      UNPCK_UB_SH(ref2, ref2_r, ref2_l);
+      UNPCK_UB_SH(ref3, ref3_r, ref3_l);
+      UNPCK_UB_SH(ref4, ref4_r, ref4_l);
+      UNPCK_UB_SH(ref5, ref5_r, ref5_l);
+      UNPCK_UB_SH(ref6, ref6_r, ref6_l);
+      UNPCK_UB_SH(ref7, ref7_r, ref7_l);
+      ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
+      ref += 8 * ref_stride;
+      UNPCK_UB_SH(ref0, ref0_r, ref0_l);
+      UNPCK_UB_SH(ref1, ref1_r, ref1_l);
+      UNPCK_UB_SH(ref2, ref2_r, ref2_l);
+      UNPCK_UB_SH(ref3, ref3_r, ref3_l);
+      UNPCK_UB_SH(ref4, ref4_r, ref4_l);
+      UNPCK_UB_SH(ref5, ref5_r, ref5_l);
+      UNPCK_UB_SH(ref6, ref6_r, ref6_l);
+      UNPCK_UB_SH(ref7, ref7_r, ref7_l);
+      ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+    }
+
+    SRA_2V(hbuf_r, hbuf_l, 4);
+    ST_SH2(hbuf_r, hbuf_l, hbuf, 8);
+  } else if (64 == height) {
+    for (i = 4; i--;) {
+      LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
+      ref += 8 * ref_stride;
+      UNPCK_UB_SH(ref0, ref0_r, ref0_l);
+      UNPCK_UB_SH(ref1, ref1_r, ref1_l);
+      UNPCK_UB_SH(ref2, ref2_r, ref2_l);
+      UNPCK_UB_SH(ref3, ref3_r, ref3_l);
+      UNPCK_UB_SH(ref4, ref4_r, ref4_l);
+      UNPCK_UB_SH(ref5, ref5_r, ref5_l);
+      UNPCK_UB_SH(ref6, ref6_r, ref6_l);
+      UNPCK_UB_SH(ref7, ref7_r, ref7_l);
+      ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
+      ref += 8 * ref_stride;
+      UNPCK_UB_SH(ref0, ref0_r, ref0_l);
+      UNPCK_UB_SH(ref1, ref1_r, ref1_l);
+      UNPCK_UB_SH(ref2, ref2_r, ref2_l);
+      UNPCK_UB_SH(ref3, ref3_r, ref3_l);
+      UNPCK_UB_SH(ref4, ref4_r, ref4_l);
+      UNPCK_UB_SH(ref5, ref5_r, ref5_l);
+      UNPCK_UB_SH(ref6, ref6_r, ref6_l);
+      UNPCK_UB_SH(ref7, ref7_r, ref7_l);
+      ADD4(hbuf_r, ref0_r, hbuf_l, ref0_l, hbuf_r, ref1_r, hbuf_l, ref1_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref2_r, hbuf_l, ref2_l, hbuf_r, ref3_r, hbuf_l, ref3_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref4_r, hbuf_l, ref4_l, hbuf_r, ref5_r, hbuf_l, ref5_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+      ADD4(hbuf_r, ref6_r, hbuf_l, ref6_l, hbuf_r, ref7_r, hbuf_l, ref7_l,
+           hbuf_r, hbuf_l, hbuf_r, hbuf_l);
+    }
+
+    SRA_2V(hbuf_r, hbuf_l, 5);
+    ST_SH2(hbuf_r, hbuf_l, hbuf, 8);
+  } else {
+    const int norm_factor = height >> 1;
+    int cnt;
+
+    for (cnt = 0; cnt < 16; cnt++) {
+      hbuf[cnt] = 0;
+    }
+
+    for (i = 0; i < height; ++i) {
+      for (cnt = 0; cnt < 16; cnt++) {
+        hbuf[cnt] += ref[cnt];
+      }
+
+      ref += ref_stride;
+    }
+
+    for (cnt = 0; cnt < 16; cnt++) {
+      hbuf[cnt] /= norm_factor;
+    }
+  }
+}
+
+int16_t vpx_int_pro_col_msa(const uint8_t *ref, const int width) {
+  int16_t sum;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 ref0_h;
+
+  if (16 == width) {
+    ref0 = LD_UB(ref);
+    ref0_h = __msa_hadd_u_h(ref0, ref0);
+    sum = HADD_UH_U32(ref0_h);
+  } else if (32 == width) {
+    LD_UB2(ref, 16, ref0, ref1);
+    ref0_h = __msa_hadd_u_h(ref0, ref0);
+    ref0_h += __msa_hadd_u_h(ref1, ref1);
+    sum = HADD_UH_U32(ref0_h);
+  } else if (64 == width) {
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref0_h = __msa_hadd_u_h(ref0, ref0);
+    ref0_h += __msa_hadd_u_h(ref1, ref1);
+    ref0_h += __msa_hadd_u_h(ref2, ref2);
+    ref0_h += __msa_hadd_u_h(ref3, ref3);
+    sum = HADD_UH_U32(ref0_h);
+  } else {
+    int idx;
+
+    sum = 0;
+    for (idx = 0; idx < width; ++idx) {
+      sum += ref[idx];
+    }
+  }
+
+  return sum;
+}

diff --git a/vpx_dsp/mips/macros_msa.h b/vpx_dsp/mips/macros_msa.h
index 002e574aa8f888a9d3bad90f2ca9682449e26e74..27b38865a4274485b15237dd5c31c1ef4dad9721 100644 (file)
--- a/vpx_dsp/mips/macros_msa.h
+++ b/vpx_dsp/mips/macros_msa.h
@@ -1049,6 +1049,7 @@
   }
 #define INSERT_D2_UB(...) INSERT_D2(v16u8, __VA_ARGS__)
 #define INSERT_D2_SB(...) INSERT_D2(v16i8, __VA_ARGS__)
+#define INSERT_D2_SH(...) INSERT_D2(v8i16, __VA_ARGS__)
 
 /* Description : Interleave even byte elements from vectors
    Arguments   : Inputs  - in0, in1, in2, in3
@@ -1559,6 +1560,12 @@
    Details     : Each element of vector 'in0' is right shifted by 'shift' and
                  the result is written in-place. 'shift' is a GP variable.
 */
+#define SRA_2V(in0, in1, shift) \
+  {                             \
+    in0 = in0 >> shift;         \
+    in1 = in1 >> shift;         \
+  }
+
 #define SRA_4V(in0, in1, in2, in3, shift) \
   {                                       \
     in0 = in0 >> shift;                   \

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 5805d1a8fa2b0fdfa662a523d481cc55cbe091f4..a3558fb2aa5d02d79349f43936bb891addc8d926 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -895,10 +895,10 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
   specialize qw/vpx_satd sse2 neon msa/;
 
   add_proto qw/void vpx_int_pro_row/, "int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height";
-  specialize qw/vpx_int_pro_row sse2 neon/;
+  specialize qw/vpx_int_pro_row sse2 neon msa/;
 
   add_proto qw/int16_t vpx_int_pro_col/, "const uint8_t *ref, const int width";
-  specialize qw/vpx_int_pro_col sse2 neon/;
+  specialize qw/vpx_int_pro_col sse2 neon msa/;
 
   add_proto qw/int vpx_vector_var/, "const int16_t *ref, const int16_t *src, const int bwl";
   specialize qw/vpx_vector_var neon sse2/;