Add MIPS SIMD Arch (MSA) optimized MirrorRow function

source/row_msa.cc

Index: source/row_msa.cc
diff --git a/source/row_msa.cc b/source/row_msa.cc
new file mode 100644
index 0000000000000000000000000000000000000000..4221a26dc4357a2ad8e08e4a68b97af3f8a2ef42
--- /dev/null
+++ b/source/row_msa.cc
@@ -0,0 +1,139 @@
+/*
+ *  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.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_msa)
+#include "libyuv/cpu/mips/macros_msa.h"
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_msa)
+void MirrorRow_MSA(const uint8* src, uint8* dst, int width)
+{

[fbarchard1, 2016/09/14 01:48:05]

{ should be on same row as void MirrorRow_MSA()

[manojkumar.bhosale, 2016/09/14 12:45:30]

Done.

+    int count;
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+    v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+    v16i8 mask = { 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 };
+
+    src += width;
+
+    for (count = 0; count < (width >> 7); count++)
+    {

[fbarchard1, 2016/09/14 01:48:05]

{ should be on same line as for ()

[manojkumar.bhosale, 2016/09/14 12:45:30]

Done.

+        src -= 128;
+        LD_UB8(src, 16, src7, src6, src5, src4, src3, src2, src1, src0);
+
+        VSHF_B2_UB(src7, src7, src6, src6, mask, mask, dst7, dst6);

[fbarchard1, 2016/09/14 01:48:04]

I wouldnt normally unroll this much.  2 or 4 vectors would normally provide a
good balance of size/performance?

[manojkumar.bhosale, 2016/09/14 12:45:30]

Done with 4 vectors

+        VSHF_B2_UB(src5, src5, src4, src4, mask, mask, dst5, dst4);
+        VSHF_B2_UB(src3, src3, src2, src2, mask, mask, dst3, dst2);
+        VSHF_B2_UB(src1, src1, src0, src0, mask, mask, dst1, dst0);
+
+        ST_UB8(dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst, 16);
+        dst += 128;
+    }
+
+    if (width & 0x40)

[fbarchard1, 2016/09/14 01:48:04]

suggest removing everything from here down to make the code simple and fastest
in the common case, which is camera and screen sizes.  e.g. 1280x720.
The MirrorRow_Any_MSA will handle the edge cases.
The current _Any implementations will call MirrorRow_MSA once in a temporary
buffer, which in this case would be 128 pixels wide.

[manojkumar.bhosale, 2016/09/14 12:45:30]

Done.

+    {
+        if (width & 0x20)
+        {
+            if (width & 0x10)
+            {
+                src -= 112;
+                LD_UB7(src, 16, src6, src5, src4, src3, src2, src1, src0);
+
+                VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
+                VSHF_B2_UB(src2, src2, src3, src3, mask, mask, dst2, dst3);
+                VSHF_B2_UB(src4, src4, src5, src5, mask, mask, dst4, dst5);
+                dst6 = (v16u8) __msa_vshf_b(mask, (v16i8) src6, (v16i8) src6);
+
+                ST_UB7(dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst, 16);
+                dst += 112;
+            }
+            else
+            {
+                src -= 96;
+                LD_UB6(src, 16, src5, src4, src3, src2, src1, src0);
+
+                VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
+                VSHF_B2_UB(src2, src2, src3, src3, mask, mask, dst2, dst3);
+                VSHF_B2_UB(src4, src4, src5, src5, mask, mask, dst4, dst5);
+
+                ST_UB6(dst0, dst1, dst2, dst3, dst4, dst5, dst, 16);
+                dst += 96;
+            }
+        }
+        else if (width & 0x10)
+        {
+            src -= 80;
+            LD_UB5(src, 16, src4, src3, src2, src1, src0);
+
+            VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
+            VSHF_B2_UB(src2, src2, src3, src3, mask, mask, dst2, dst3);
+            dst4 = (v16u8) __msa_vshf_b(mask, (v16i8) src4, (v16i8) src4);
+
+            ST_UB5(dst0, dst1, dst2, dst3, dst4, dst, 16);
+            dst += 80;
+        }
+        else
+        {
+            src -= 64;
+            LD_UB4(src, 16, src3, src2, src1, src0);
+
+            VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
+            VSHF_B2_UB(src2, src2, src3, src3, mask, mask, dst2, dst3);
+
+            ST_UB4(dst0, dst1, dst2, dst3, dst, 16);
+            dst += 64;
+        }
+    }
+    else if (width & 0x20)
+    {
+        if (width & 0x10)
+        {
+            src -= 48;
+            LD_UB3(src, 16, src2, src1, src0);
+
+            VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
+            dst2 = (v16u8) __msa_vshf_b(mask, (v16i8) src2, (v16i8) src2);
+
+            ST_UB3(dst0, dst1, dst2, dst, 16);
+            dst += 48;
+        }
+        else
+        {
+            src -= 32;
+            LD_UB2(src, 16, src1, src0);
+
+            VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
+
+            ST_UB2(dst0, dst1, dst, 16);
+            dst += 32;
+        }
+    }
+    else if (width & 0x10)
+    {
+        src -= 16;
+        src0 = LD_UB(src);
+
+        dst0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0, (v16i8) src0);
+        ST_UB(dst0, dst);
+        dst += 16;
+    }
+}
+#endif  // !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_msa)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif