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..7d517fa46d9876fe9d69f5070bed05fc61f8b74e
--- /dev/null
+++ b/source/row_msa.cc
@@ -0,0 +1,181 @@
+/*
+ *  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"
+#include "libyuv/macros_msa.h"

[fbarchard1, 2016/08/31 01:20:42]

This header wont compile on intel/arm as is.  May want to move the
defined(__mips_msa) around more of this file and/or put guards in
libyuv/macros_msa.h

+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_msa)
+
+void MirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width)
+{
+    int32_t count;
+    uint8_t val0, val1, val2, val3;
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7;

[fbarchard1, 2016/08/31 01:20:43]

prefer uvec8 be defined as v16u8 in row.h if that works.

+    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 };
+    int32_t cnt64, cnt32, cnt16;

[fbarchard1, 2016/08/31 01:20:43]

prefer int32 defined in libyuv/basic_types.h
or if storage size is not critical, use int.

+
+    src += width;
+
+    for (count = 0; count < (width >> 7); count++)
+    {
+        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);
+        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 >> 4) & 0x7)

[fbarchard1, 2016/08/31 01:20:42]

its not necessary to hand odd widths in row functions.
Define the most efficient version here, and a MirrorRow_Any_MSA() in row_any,h

+    {
+        cnt64 = (width >> 6) & 0x1;
+        cnt32 = (width >> 5) & 0x1;
+        cnt16 = (width >> 4) & 0x1;
+
+        if (cnt16)
+        {
+            src -= 16;
+            src0 = LD_UB(src);
+
+            if (cnt32)
+            {
+                src -= 32;
+                LD_UB2(src, 16, src2, src1);
+
+                if (cnt64)
+                {
+                    src -= 64;
+                    LD_UB4(src, 16, src6, src5, src4, src3);
+
+                    dst0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0,
+                                                (v16i8) src0);
+                    VSHF_B2_UB(src2, src2, src1, src1, mask, mask, dst2, dst1);
+                    VSHF_B2_UB(src6, src6, src5, src5, mask, mask, dst6, dst5);
+                    VSHF_B2_UB(src4, src4, src3, src3, mask, mask, dst4, dst3);
+
+                    ST_UB(dst0, dst);
+                    dst += 16;
+                    ST_UB2(dst1, dst2, dst, 16);
+                    dst += 32;
+                    ST_UB4(dst3, dst4, dst5, dst6, dst, 16);
+                    dst += 64;
+                }
+                else
+                {
+                    dst0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0,
+                                                (v16i8) src0);
+                    VSHF_B2_UB(src2, src2, src1, src1, mask, mask, dst2, dst1);
+
+                    ST_UB(dst0, dst);
+                    dst += 16;
+                    ST_UB2(dst1, dst2, dst, 16);
+                    dst += 32;
+                }
+            }
+            else if (cnt64)
+            {
+                src -= 64;
+                LD_UB4(src, 16, src6, src5, src4, src3);
+
+                dst0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0,
+                                            (v16i8) src0);
+                VSHF_B2_UB(src6, src6, src5, src5, mask, mask, dst4, dst3);
+                VSHF_B2_UB(src4, src4, src3, src3, mask, mask, dst2, dst1);
+
+                ST_UB(dst0, dst);
+                dst += 16;
+                ST_UB4(dst1, dst2, dst3, dst4, dst, 16);
+                dst += 64;
+            }
+            else
+            {
+                dst0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0,
+                                            (v16i8) src0);
+
+                ST_UB(dst0, dst);
+                dst += 16;
+            }
+        }
+        else if (cnt32)
+        {
+            src -= 32;
+            LD_UB2(src, 16, src2, src1);
+
+            if (cnt64)
+            {
+                src -= 64;
+                LD_UB4(src, 16, src6, src5, src4, src3);
+
+                VSHF_B2_UB(src2, src2, src1, src1, mask, mask, dst1, dst0);
+                VSHF_B2_UB(src6, src6, src5, src5, mask, mask, dst5, dst4);
+                VSHF_B2_UB(src4, src4, src3, src3, mask, mask, dst3, dst2);
+
+                ST_UB2(dst0, dst1, dst, 16);
+                dst += 32;
+                ST_UB4(dst2, dst3, dst4, dst5, dst, 16);
+                dst += 64;
+            }
+            else
+            {
+                VSHF_B2_UB(src2, src2, src1, src1, mask, mask, dst1, dst0);
+
+                ST_UB2(dst0, dst1, dst, 16);
+                dst += 32;
+            }
+        }
+        else if (cnt64)
+        {
+            src -= 64;
+            LD_UB4(src, 16, src6, src5, src4, src3);
+
+            VSHF_B2_UB(src6, src6, src5, src5, mask, mask, dst3, dst2);
+            VSHF_B2_UB(src4, src4, src3, src3, mask, mask, dst1, dst0);
+
+            ST_UB4(dst0, dst1, dst2, dst3, dst, 16);
+            dst += 64;
+        }
+    }
+
+    for (count = 0; count < ((width >> 2) & 0x3); count++)
+    {
+        src -= 4;
+        val0 = src[3];
+        val1 = src[2];
+        val2 = src[1];
+        val3 = src[0];
+        dst[0] = val0;
+        dst[1] = val1;
+        dst[2] = val2;
+        dst[3] = val3;
+        dst += 4;
+    }
+
+    for (count = 0; count < (width & 0x3); count++)
+    {
+        val0 = src[-1];
+        src--;
+        dst[0] = val0;
+        dst++;
+    }
+}
+#endif  // !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_msa)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif