Add MSA optimized HammingDistance and SumSquareError functions

TBR=kjellander@chromium.org
R=fbarchard@google.com

Bug:libyuv:634
Change-Id: Id0126ba5aff38817525b1efa6044f1dc2cfa1a36
Reviewed-on: https://chromium-review.googlesource.com/625739Reviewed-by: Frank Barchard <fbarchard@google.com>

Android.mk

@@ -9,6 +9,7 @@ LOCAL_SRC_FILES := \
     source/compare.cc           \
     source/compare_common.cc    \
     source/compare_gcc.cc       \
+    source/compare_msa.cc       \
     source/compare_neon.cc      \
     source/compare_neon64.cc    \
     source/convert.cc           \

BUILD.gn

@@ -199,6 +199,7 @@ if (libyuv_use_msa) {
   static_library("libyuv_msa") {
     sources = [
       # MSA Source Files
+      "source/compare_msa.cc",
       "source/rotate_msa.cc",
       "source/row_msa.cc",
       "source/scale_msa.cc",

include/libyuv/compare_row.h

@@ -42,6 +42,11 @@ extern "C" {
 #endif  // clang >= 3.4
 #endif  // __clang__
 
+// clang 6 mips issue https://bugs.chromium.org/p/libyuv/issues/detail?id=715
+#if defined(__clang__)
+#define DISABLE_CLANG_MSA 1
+#endif
+
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
     (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
 #define HAS_HASHDJB2_AVX2
@@ -69,14 +74,24 @@ extern "C" {
 #define HAS_HAMMINGDISTANCE_NEON
 #endif
 
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#define HAS_HAMMINGDISTANCE_MSA
+
+#ifndef DISABLE_CLANG_MSA
+#define HAS_SUMSQUAREERROR_MSA
+#endif
+#endif
+
 uint32 HammingDistance_C(const uint8* src_a, const uint8* src_b, int count);
 uint32 HammingDistance_X86(const uint8* src_a, const uint8* src_b, int count);
 uint32 HammingDistance_NEON(const uint8* src_a, const uint8* src_b, int count);
+uint32 HammingDistance_MSA(const uint8* src_a, const uint8* src_b, int count);
 
 uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
 uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
 uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
 uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
+uint32 SumSquareError_MSA(const uint8* src_a, const uint8* src_b, int count);
 
 uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
 uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);

libyuv.gypi

@@ -35,6 +35,7 @@
       'source/compare.cc',
       'source/compare_common.cc',
       'source/compare_gcc.cc',
+      'source/compare_msa.cc',
       'source/compare_neon.cc',
       'source/compare_neon64.cc',
       'source/compare_win.cc',

source/compare.cc

@@ -135,6 +135,11 @@ uint64 ComputeHammingDistance(const uint8* src_a,
     HammingDistance = HammingDistance_AVX2;
   }
 #endif
+#if defined(HAS_HAMMINGDISTANCE_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    HammingDistance = HammingDistance_MSA;
+  }
+#endif
 #ifdef _OPENMP
 #pragma omp parallel for reduction(+ : diff)
 #endif
@@ -186,6 +191,11 @@ uint64 ComputeSumSquareError(const uint8* src_a,
     SumSquareError = SumSquareError_AVX2;
   }
 #endif
+#if defined(HAS_SUMSQUAREERROR_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SumSquareError = SumSquareError_MSA;
+  }
+#endif
 #ifdef _OPENMP
 #pragma omp parallel for reduction(+ : sse)
 #endif

source/compare_msa.cc

+/*
+ *  Copyright 2017 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.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+#include "libyuv/row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+uint32 HammingDistance_MSA(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 diff = 0u;
+  int i;
+  v16u8 src0, src1, src2, src3;
+  v2i64 vec0 = {0}, vec1 = {0};
+
+  for (i = 0; i < count; i += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_a, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_a, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_b, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_b, 16);
+    src0 ^= src2;
+    src1 ^= src3;
+    vec0 += __msa_pcnt_d((v2i64)src0);
+    vec1 += __msa_pcnt_d((v2i64)src1);
+    src_a += 32;
+    src_b += 32;
+  }
+
+  vec0 += vec1;
+  diff = (uint32)__msa_copy_u_w((v4i32)vec0, 0);
+  diff += (uint32)__msa_copy_u_w((v4i32)vec0, 2);
+  return diff;
+}
+
+#ifndef DISABLE_CLANG_MSA
+uint32 SumSquareError_MSA(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse = 0u;
+  int i;
+  v16u8 src0, src1, src2, src3;
+  v8i16 vec0, vec1, vec2, vec3;
+  v4i32 reg0 = {0}, reg1 = {0}, reg2 = {0}, reg3 = {0};
+  v2i64 tmp0;
+
+  for (i = 0; i < count; i += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_a, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_a, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_b, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_b, 16);
+    vec0 = (v8i16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8i16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8i16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8i16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = __msa_hsub_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hsub_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hsub_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hsub_u_h((v16u8)vec3, (v16u8)vec3);
+    reg0 = __msa_dpadd_s_w(reg0, vec0, vec0);
+    reg1 = __msa_dpadd_s_w(reg1, vec1, vec1);
+    reg2 = __msa_dpadd_s_w(reg2, vec2, vec2);
+    reg3 = __msa_dpadd_s_w(reg3, vec3, vec3);
+    src_a += 32;
+    src_b += 32;
+  }
+
+  reg0 += reg1;
+  reg2 += reg3;
+  reg0 += reg2;
+  tmp0 = __msa_hadd_s_d(reg0, reg0);
+  sse = (uint32)__msa_copy_u_w((v4i32)tmp0, 0);
+  sse += (uint32)__msa_copy_u_w((v4i32)tmp0, 2);
+  return sse;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)