Add mips msa sum_squares_2d_i16 function

average improvement ~4x-5x

Change-Id: I8d91b71d0677009be52b412e4f52b40b98573a53

diff --git a/test/sum_squares_test.cc b/test/sum_squares_test.cc
index 3aa43d2ce07c0113af372e9f3acc2b4e06b1e98e..9c407c649f41abee22b877db36efda1fc7472666 100644 (file)
--- a/test/sum_squares_test.cc
+++ b/test/sum_squares_test.cc
@@ -110,4 +110,10 @@ INSTANTIATE_TEST_CASE_P(
     ::testing::Values(make_tuple(&vpx_sum_squares_2d_i16_c,
                                  &vpx_sum_squares_2d_i16_sse2)));
 #endif  // HAVE_SSE2
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, SumSquaresTest, ::testing::Values(make_tuple(
+                                                 &vpx_sum_squares_2d_i16_c,
+                                                 &vpx_sum_squares_2d_i16_msa)));
+#endif  // HAVE_MSA
 }  // namespace

diff --git a/vpx_dsp/mips/sum_squares_msa.c b/vpx_dsp/mips/sum_squares_msa.c
new file mode 100644 (file)
index 0000000..d4563dc
--- /dev/null
+++ b/vpx_dsp/mips/sum_squares_msa.c
@@ -0,0 +1,129 @@
+/*
+ *  Copyright (c) 2017 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 "./vpx_dsp_rtcd.h"
+#include "./macros_msa.h"
+
+uint64_t vpx_sum_squares_2d_i16_msa(const int16_t *src, int src_stride,
+                                    int size) {
+  int row, col;
+  uint64_t ss_res = 0;
+  v4i32 mul0, mul1;
+  v2i64 res0 = { 0 };
+
+  if (4 == size) {
+    uint64_t src0, src1, src2, src3;
+    v8i16 diff0 = { 0 };
+    v8i16 diff1 = { 0 };
+
+    LD4(src, src_stride, src0, src1, src2, src3);
+    INSERT_D2_SH(src0, src1, diff0);
+    INSERT_D2_SH(src2, src3, diff1);
+    DOTP_SH2_SW(diff0, diff1, diff0, diff1, mul0, mul1);
+    mul0 += mul1;
+    res0 = __msa_hadd_s_d(mul0, mul0);
+    res0 += __msa_splati_d(res0, 1);
+    ss_res = (uint64_t)__msa_copy_s_d(res0, 0);
+  } else if (8 == size) {
+    v8i16 src0, src1, src2, src3, src4, src5, src6, src7;
+
+    LD_SH8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    DOTP_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+    DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+    DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+    DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+    mul0 += mul1;
+    res0 = __msa_hadd_s_d(mul0, mul0);
+    res0 += __msa_splati_d(res0, 1);
+    ss_res = (uint64_t)__msa_copy_s_d(res0, 0);
+  } else if (16 == size) {
+    v8i16 src0, src1, src2, src3, src4, src5, src6, src7;
+
+    LD_SH8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    DOTP_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+    DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+    DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+    DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+    LD_SH8(src + 8, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += 8 * src_stride;
+    DPADD_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+    DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+    DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+    DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+    LD_SH8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    DPADD_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+    DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+    DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+    DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+    LD_SH8(src + 8, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    DPADD_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+    DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+    DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+    DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+    mul0 += mul1;
+    res0 += __msa_hadd_s_d(mul0, mul0);
+
+    res0 += __msa_splati_d(res0, 1);
+    ss_res = (uint64_t)__msa_copy_s_d(res0, 0);
+  } else if (0 == (size % 16)) {
+    v8i16 src0, src1, src2, src3, src4, src5, src6, src7;
+
+    for (row = 0; row < (size >> 4); row++) {
+      for (col = 0; col < size; col += 16) {
+        const int16_t *src_ptr = src + col;
+        LD_SH8(src_ptr, src_stride, src0, src1, src2, src3, src4, src5, src6,
+               src7);
+        DOTP_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+        DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+        DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+        DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+        LD_SH8(src_ptr + 8, src_stride, src0, src1, src2, src3, src4, src5,
+               src6, src7);
+        src_ptr += 8 * src_stride;
+        DPADD_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+        DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+        DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+        DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+        LD_SH8(src_ptr, src_stride, src0, src1, src2, src3, src4, src5, src6,
+               src7);
+        DPADD_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+        DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+        DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+        DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+        LD_SH8(src_ptr + 8, src_stride, src0, src1, src2, src3, src4, src5,
+               src6, src7);
+        DPADD_SH2_SW(src0, src1, src0, src1, mul0, mul1);
+        DPADD_SH2_SW(src2, src3, src2, src3, mul0, mul1);
+        DPADD_SH2_SW(src4, src5, src4, src5, mul0, mul1);
+        DPADD_SH2_SW(src6, src7, src6, src7, mul0, mul1);
+        mul0 += mul1;
+        res0 += __msa_hadd_s_d(mul0, mul0);
+      }
+
+      src += 16 * src_stride;
+    }
+
+    res0 += __msa_splati_d(res0, 1);
+    ss_res = (uint64_t)__msa_copy_s_d(res0, 0);
+  } else {
+    int16_t val;
+
+    for (row = 0; row < size; row++) {
+      for (col = 0; col < size; col++) {
+        val = src[col];
+        ss_res += val * val;
+      }
+
+      src += src_stride;
+    }
+  }
+
+  return ss_res;
+}

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index bb20ea27421c8b2bc7d03c95b43ca57a151d8470..642c4260102759af058b63efb05db338e6fb8fa9 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -276,6 +276,7 @@ DSP_SRCS-yes            += sad.c
 DSP_SRCS-yes            += subtract.c
 DSP_SRCS-yes            += sum_squares.c
 DSP_SRCS-$(HAVE_SSE2)   += x86/sum_squares_sse2.c
+DSP_SRCS-$(HAVE_MSA)    += mips/sum_squares_msa.c
 
 DSP_SRCS-$(HAVE_NEON)   += arm/sad4d_neon.c
 DSP_SRCS-$(HAVE_NEON)   += arm/sad_neon.c

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index e7f4010deaba9c5e306630c50e2f8be5abf5c5df..535fab366b2876bae6392b382070ca4733d91544 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1039,7 +1039,7 @@ add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const
 specialize qw/vpx_sad4x4x4d msa sse2/;
 
 add_proto qw/uint64_t vpx_sum_squares_2d_i16/, "const int16_t *src, int stride, int size";
-specialize qw/vpx_sum_squares_2d_i16 sse2/;
+specialize qw/vpx_sum_squares_2d_i16 sse2 msa/;
 
 #
 # Structured Similarity (SSIM)