avcodec/mips: MSA (MIPS-SIMD-Arch) optimizations for me_cmp functions
authorShivraj Patil <shivraj.patil@imgtec.com>
Mon, 29 Jun 2015 15:27:14 +0000 (20:57 +0530)
committerMichael Niedermayer <michaelni@gmx.at>
Mon, 6 Jul 2015 16:25:14 +0000 (18:25 +0200)
This patch adds MSA (MIPS-SIMD-Arch) optimizations for me_cmp functions in new file me_cmp_msa.c

Signed-off-by: Shivraj Patil <shivraj.patil@imgtec.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
libavcodec/me_cmp.c
libavcodec/me_cmp.h
libavcodec/mips/Makefile
libavcodec/mips/me_cmp_init_mips.c [new file with mode: 0644]
libavcodec/mips/me_cmp_mips.h [new file with mode: 0644]
libavcodec/mips/me_cmp_msa.c [new file with mode: 0644]
libavutil/mips/generic_macros_msa.h

index d4213d2..dc76b07 100644 (file)
@@ -991,4 +991,6 @@ av_cold void ff_me_cmp_init(MECmpContext *c, AVCodecContext *avctx)
         ff_me_cmp_init_ppc(c, avctx);
     if (ARCH_X86)
         ff_me_cmp_init_x86(c, avctx);
+    if (ARCH_MIPS)
+        ff_me_cmp_init_mips(c, avctx);
 }
index 98ee53c..a3603ec 100644 (file)
@@ -87,6 +87,7 @@ void ff_me_cmp_init_alpha(MECmpContext *c, AVCodecContext *avctx);
 void ff_me_cmp_init_arm(MECmpContext *c, AVCodecContext *avctx);
 void ff_me_cmp_init_ppc(MECmpContext *c, AVCodecContext *avctx);
 void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx);
+void ff_me_cmp_init_mips(MECmpContext *c, AVCodecContext *avctx);
 
 void ff_set_cmp(MECmpContext *c, me_cmp_func *cmp, int type);
 
index 59c1f79..2993891 100644 (file)
@@ -31,6 +31,7 @@ OBJS-$(CONFIG_BLOCKDSP)                   += mips/blockdsp_init_mips.o
 OBJS-$(CONFIG_PIXBLOCKDSP)                += mips/pixblockdsp_init_mips.o
 OBJS-$(CONFIG_MPEGVIDEO)                  += mips/mpegvideo_init_mips.o
 OBJS-$(CONFIG_MPEGVIDEOENC)               += mips/mpegvideoencdsp_init_mips.o
+OBJS-$(CONFIG_ME_CMP)                     += mips/me_cmp_init_mips.o
 MSA-OBJS-$(CONFIG_HEVC_DECODER)           += mips/hevcdsp_msa.o            \
                                              mips/hevc_mc_uni_msa.o        \
                                              mips/hevc_mc_uniw_msa.o       \
@@ -51,5 +52,6 @@ MSA-OBJS-$(CONFIG_BLOCKDSP)               += mips/blockdsp_msa.o
 MSA-OBJS-$(CONFIG_PIXBLOCKDSP)            += mips/pixblockdsp_msa.o
 MSA-OBJS-$(CONFIG_MPEGVIDEO)              += mips/mpegvideo_msa.o
 MSA-OBJS-$(CONFIG_MPEGVIDEOENC)           += mips/mpegvideoencdsp_msa.o
+MSA-OBJS-$(CONFIG_ME_CMP)                 += mips/me_cmp_msa.o
 LOONGSON3-OBJS-$(CONFIG_H264DSP)          += mips/h264dsp_mmi.o
 LOONGSON3-OBJS-$(CONFIG_H264CHROMA)       += mips/h264chroma_mmi.o
diff --git a/libavcodec/mips/me_cmp_init_mips.c b/libavcodec/mips/me_cmp_init_mips.c
new file mode 100644 (file)
index 0000000..219a0dc
--- /dev/null
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2015 Parag Salasakar (Parag.Salasakar@imgtec.com)
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "me_cmp_mips.h"
+
+#if HAVE_MSA
+static av_cold void me_cmp_msa(MECmpContext *c, AVCodecContext *avctx)
+{
+#if BIT_DEPTH == 8
+    c->pix_abs[0][0] = ff_pix_abs16_msa;
+    c->pix_abs[0][1] = ff_pix_abs16_x2_msa;
+    c->pix_abs[0][2] = ff_pix_abs16_y2_msa;
+    c->pix_abs[0][3] = ff_pix_abs16_xy2_msa;
+    c->pix_abs[1][0] = ff_pix_abs8_msa;
+    c->pix_abs[1][1] = ff_pix_abs8_x2_msa;
+    c->pix_abs[1][2] = ff_pix_abs8_y2_msa;
+    c->pix_abs[1][3] = ff_pix_abs8_xy2_msa;
+
+    c->hadamard8_diff[0] = ff_hadamard8_diff16_msa;
+    c->hadamard8_diff[1] = ff_hadamard8_diff8x8_msa;
+
+    c->hadamard8_diff[4] = ff_hadamard8_intra16_msa;
+    c->hadamard8_diff[5] = ff_hadamard8_intra8x8_msa;
+
+    c->sad[0] = ff_pix_abs16_msa;
+    c->sad[1] = ff_pix_abs8_msa;
+    c->sse[0] = ff_sse16_msa;
+    c->sse[1] = ff_sse8_msa;
+    c->sse[2] = ff_sse4_msa;
+#endif
+}
+#endif  // #if HAVE_MSA
+
+av_cold void ff_me_cmp_init_mips(MECmpContext *c, AVCodecContext *avctx)
+{
+#if HAVE_MSA
+    me_cmp_msa(c, avctx);
+#endif  // #if HAVE_MSA
+}
diff --git a/libavcodec/mips/me_cmp_mips.h b/libavcodec/mips/me_cmp_mips.h
new file mode 100644 (file)
index 0000000..e0d0f51
--- /dev/null
@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) 2015 Parag Salasakar (Parag.Salasakar@imgtec.com)
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_MIPS_ME_CMP_MIPS_H
+#define AVCODEC_MIPS_ME_CMP_MIPS_H
+
+#include "../mpegvideo.h"
+#include "libavcodec/bit_depth_template.c"
+
+int ff_hadamard8_diff8x8_msa(MpegEncContext *s, uint8_t *dst, uint8_t *src,
+                             ptrdiff_t stride, int h);
+int ff_hadamard8_intra8x8_msa(MpegEncContext *s, uint8_t *dst, uint8_t *src,
+                              ptrdiff_t stride, int h);
+int ff_hadamard8_diff16_msa(MpegEncContext *s, uint8_t *dst, uint8_t *src,
+                            ptrdiff_t stride, int h);
+int ff_hadamard8_intra16_msa(MpegEncContext *s, uint8_t *dst, uint8_t *src,
+                             ptrdiff_t stride, int h);
+int ff_pix_abs16_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                     ptrdiff_t stride, int h);
+int ff_pix_abs16_x2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                        ptrdiff_t stride, int h);
+int ff_pix_abs16_y2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                        ptrdiff_t stride, int h);
+int ff_pix_abs16_xy2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                         ptrdiff_t stride, int h);
+int ff_pix_abs8_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                    ptrdiff_t stride, int h);
+int ff_pix_abs8_x2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                       ptrdiff_t stride, int h);
+int ff_pix_abs8_y2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                       ptrdiff_t stride, int h);
+int ff_pix_abs8_xy2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                        ptrdiff_t stride, int h);
+int ff_sse16_msa(MpegEncContext *v, uint8_t *pu8Src, uint8_t *pu8Ref,
+                 ptrdiff_t stride, int i32Height);
+int ff_sse8_msa(MpegEncContext *v, uint8_t *pu8Src, uint8_t *pu8Ref,
+                ptrdiff_t stride, int i32Height);
+int ff_sse4_msa(MpegEncContext *v, uint8_t *pu8Src, uint8_t *pu8Ref,
+                ptrdiff_t stride, int i32Height);
+void ff_add_pixels8_msa(uint8_t *av_restrict pixels, int16_t *block,
+                        ptrdiff_t stride);
+
+#endif  // #ifndef AVCODEC_MIPS_ME_CMP_MIPS_H
diff --git a/libavcodec/mips/me_cmp_msa.c b/libavcodec/mips/me_cmp_msa.c
new file mode 100644 (file)
index 0000000..0e3165c
--- /dev/null
@@ -0,0 +1,686 @@
+/*
+ * Copyright (c) 2015 Parag Salasakar (Parag.Salasakar@imgtec.com)
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/mips/generic_macros_msa.h"
+#include "me_cmp_mips.h"
+
+static uint32_t sad_8width_msa(uint8_t *src, int32_t src_stride,
+                               uint8_t *ref, int32_t ref_stride,
+                               int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 2); ht_cnt--;) {
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+        ref += (4 * ref_stride);
+
+        PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                    src0, src1, ref0, ref1);
+        sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_16width_msa(uint8_t *src, int32_t src_stride,
+                                uint8_t *ref, int32_t ref_stride,
+                                int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, ref0, ref1;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 2); ht_cnt--;) {
+        LD_UB2(src, src_stride, src0, src1);
+        src += (2 * src_stride);
+        LD_UB2(ref, ref_stride, ref0, ref1);
+        ref += (2 * ref_stride);
+        sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+        LD_UB2(src, src_stride, src0, src1);
+        src += (2 * src_stride);
+        LD_UB2(ref, ref_stride, ref0, ref1);
+        ref += (2 * ref_stride);
+        sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_horiz_bilinear_filter_8width_msa(uint8_t *src,
+                                                     int32_t src_stride,
+                                                     uint8_t *ref,
+                                                     int32_t ref_stride,
+                                                     int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, comp0, comp1;
+    v16u8 ref0, ref1, ref2, ref3, ref4, ref5;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 3); ht_cnt--;) {
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+        ref += (4 * ref_stride);
+
+        PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+        PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref4, ref5);
+        SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+        SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+        PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+        AVER_UB2_UB(ref4, ref0, ref5, ref1, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+        ref += (4 * ref_stride);
+
+        PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+        PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref4, ref5);
+        SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+        SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+        PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+        AVER_UB2_UB(ref4, ref0, ref5, ref1, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_horiz_bilinear_filter_16width_msa(uint8_t *src,
+                                                      int32_t src_stride,
+                                                      uint8_t *ref,
+                                                      int32_t ref_stride,
+                                                      int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, comp0, comp1;
+    v16u8 ref00, ref10, ref20, ref30, ref01, ref11, ref21, ref31;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 3); ht_cnt--;) {
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB4(ref, ref_stride, ref00, ref10, ref20, ref30);
+        LD_UB4(ref + 1, ref_stride, ref01, ref11, ref21, ref31);
+        ref += (4 * ref_stride);
+
+        AVER_UB2_UB(ref01, ref00, ref11, ref10, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+        AVER_UB2_UB(ref21, ref20, ref31, ref30, comp0, comp1);
+        sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB4(ref, ref_stride, ref00, ref10, ref20, ref30);
+        LD_UB4(ref + 1, ref_stride, ref01, ref11, ref21, ref31);
+        ref += (4 * ref_stride);
+
+        AVER_UB2_UB(ref01, ref00, ref11, ref10, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+        AVER_UB2_UB(ref21, ref20, ref31, ref30, comp0, comp1);
+        sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_vert_bilinear_filter_8width_msa(uint8_t *src,
+                                                    int32_t src_stride,
+                                                    uint8_t *ref,
+                                                    int32_t ref_stride,
+                                                    int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, comp0, comp1;
+    v16u8 ref0, ref1, ref2, ref3, ref4;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 3); ht_cnt--;) {
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB5(ref, ref_stride, ref0, ref1, ref2, ref3, ref4);
+        ref += (4 * ref_stride);
+
+        PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+        PCKEV_D2_UB(ref1, ref0, ref2, ref1, ref0, ref1);
+        PCKEV_D2_UB(ref3, ref2, ref4, ref3, ref2, ref3);
+        AVER_UB2_UB(ref1, ref0, ref3, ref2, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB5(ref, ref_stride, ref0, ref1, ref2, ref3, ref4);
+        ref += (4 * ref_stride);
+
+        PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+        PCKEV_D2_UB(ref1, ref0, ref2, ref1, ref0, ref1);
+        PCKEV_D2_UB(ref3, ref2, ref4, ref3, ref2, ref3);
+        AVER_UB2_UB(ref1, ref0, ref3, ref2, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_vert_bilinear_filter_16width_msa(uint8_t *src,
+                                                     int32_t src_stride,
+                                                     uint8_t *ref,
+                                                     int32_t ref_stride,
+                                                     int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, comp0, comp1;
+    v16u8 ref0, ref1, ref2, ref3, ref4;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 3); ht_cnt--;) {
+        LD_UB5(ref, ref_stride, ref4, ref0, ref1, ref2, ref3);
+        ref += (5 * ref_stride);
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+
+        AVER_UB2_UB(ref0, ref4, ref1, ref0, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+        AVER_UB2_UB(ref2, ref1, ref3, ref2, comp0, comp1);
+        sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+
+        ref4 = ref3;
+
+        LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+        ref += (3 * ref_stride);
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+
+        AVER_UB2_UB(ref0, ref4, ref1, ref0, comp0, comp1);
+        sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+        AVER_UB2_UB(ref2, ref1, ref3, ref2, comp0, comp1);
+        sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_hv_bilinear_filter_8width_msa(uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *ref,
+                                                  int32_t ref_stride,
+                                                  int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, temp0, temp1, diff;
+    v16u8 ref0, ref1, ref2, ref3, ref4;
+    v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+    v8u16 comp0, comp1, comp2, comp3;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 2); ht_cnt--;) {
+        LD_UB5(ref, ref_stride, ref4, ref0, ref1, ref2, ref3);
+        ref += (4 * ref_stride);
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+
+        PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+
+        VSHF_B2_UB(ref4, ref4, ref0, ref0, mask, mask, temp0, temp1);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp1 = __msa_hadd_u_h(temp1, temp1);
+        comp0 += comp1;
+        comp0 = (v8u16) __msa_srari_h((v8i16) comp0, 2);
+        comp0 = (v8u16) __msa_pckev_b((v16i8) comp0, (v16i8) comp0);
+
+        temp0 = (v16u8) __msa_vshf_b(mask, (v16i8) ref1, (v16i8) ref1);
+        comp2 = __msa_hadd_u_h(temp0, temp0);
+        comp1 += comp2;
+        comp1 = (v8u16) __msa_srari_h((v8i16) comp1, 2);
+        comp1 = (v8u16) __msa_pckev_b((v16i8) comp1, (v16i8) comp1);
+        comp1 = (v8u16) __msa_pckev_d((v2i64) comp1, (v2i64) comp0);
+        diff = (v16u8) __msa_asub_u_b(src0, (v16u8) comp1);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        temp1 = (v16u8) __msa_vshf_b(mask, (v16i8) ref2, (v16i8) ref2);
+        comp3 = __msa_hadd_u_h(temp1, temp1);
+        comp2 += comp3;
+        comp2 = (v8u16) __msa_srari_h((v8i16) comp2, 2);
+        comp2 = (v8u16) __msa_pckev_b((v16i8) comp2, (v16i8) comp2);
+
+        temp0 = (v16u8) __msa_vshf_b(mask, (v16i8) ref3, (v16i8) ref3);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp3 += comp0;
+        comp3 = (v8u16) __msa_srari_h((v8i16) comp3, 2);
+        comp3 = (v8u16) __msa_pckev_b((v16i8) comp3, (v16i8) comp3);
+        comp3 = (v8u16) __msa_pckev_d((v2i64) comp3, (v2i64) comp2);
+        diff = (v16u8) __msa_asub_u_b(src1, (v16u8) comp3);
+        sad += __msa_hadd_u_h(diff, diff);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+static uint32_t sad_hv_bilinear_filter_16width_msa(uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *ref,
+                                                   int32_t ref_stride,
+                                                   int32_t height)
+{
+    int32_t ht_cnt;
+    v16u8 src0, src1, src2, src3, comp, diff;
+    v16u8 temp0, temp1, temp2, temp3;
+    v16u8 ref00, ref01, ref02, ref03, ref04, ref10, ref11, ref12, ref13, ref14;
+    v8u16 comp0, comp1, comp2, comp3;
+    v8u16 sad = { 0 };
+
+    for (ht_cnt = (height >> 3); ht_cnt--;) {
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB5(ref, ref_stride, ref04, ref00, ref01, ref02, ref03);
+        LD_UB5(ref + 1, ref_stride, ref14, ref10, ref11, ref12, ref13);
+        ref += (5 * ref_stride);
+
+        ILVRL_B2_UB(ref14, ref04, temp0, temp1);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp1 = __msa_hadd_u_h(temp1, temp1);
+        ILVRL_B2_UB(ref10, ref00, temp2, temp3);
+        comp2 = __msa_hadd_u_h(temp2, temp2);
+        comp3 = __msa_hadd_u_h(temp3, temp3);
+        comp0 += comp2;
+        comp1 += comp3;
+        SRARI_H2_UH(comp0, comp1, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp1, (v16i8) comp0);
+        diff = __msa_asub_u_b(src0, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        ILVRL_B2_UB(ref11, ref01, temp0, temp1);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp1 = __msa_hadd_u_h(temp1, temp1);
+        comp2 += comp0;
+        comp3 += comp1;
+        SRARI_H2_UH(comp2, comp3, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp3, (v16i8) comp2);
+        diff = __msa_asub_u_b(src1, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        ILVRL_B2_UB(ref12, ref02, temp2, temp3);
+        comp2 = __msa_hadd_u_h(temp2, temp2);
+        comp3 = __msa_hadd_u_h(temp3, temp3);
+        comp0 += comp2;
+        comp1 += comp3;
+        SRARI_H2_UH(comp0, comp1, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp1, (v16i8) comp0);
+        diff = __msa_asub_u_b(src2, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        ILVRL_B2_UB(ref13, ref03, temp0, temp1);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp1 = __msa_hadd_u_h(temp1, temp1);
+        comp2 += comp0;
+        comp3 += comp1;
+        SRARI_H2_UH(comp2, comp3, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp3, (v16i8) comp2);
+        diff = __msa_asub_u_b(src3, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        LD_UB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+        LD_UB4(ref, ref_stride, ref00, ref01, ref02, ref03);
+        LD_UB4(ref + 1, ref_stride, ref10, ref11, ref12, ref13);
+        ref += (3 * ref_stride);
+
+        ILVRL_B2_UB(ref10, ref00, temp2, temp3);
+        comp2 = __msa_hadd_u_h(temp2, temp2);
+        comp3 = __msa_hadd_u_h(temp3, temp3);
+        comp0 += comp2;
+        comp1 += comp3;
+        SRARI_H2_UH(comp0, comp1, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp1, (v16i8) comp0);
+        diff = __msa_asub_u_b(src0, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        ILVRL_B2_UB(ref11, ref01, temp0, temp1);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp1 = __msa_hadd_u_h(temp1, temp1);
+        comp2 += comp0;
+        comp3 += comp1;
+        SRARI_H2_UH(comp2, comp3, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp3, (v16i8) comp2);
+        diff = __msa_asub_u_b(src1, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        ILVRL_B2_UB(ref12, ref02, temp2, temp3);
+        comp2 = __msa_hadd_u_h(temp2, temp2);
+        comp3 = __msa_hadd_u_h(temp3, temp3);
+        comp0 += comp2;
+        comp1 += comp3;
+        SRARI_H2_UH(comp0, comp1, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp1, (v16i8) comp0);
+        diff = __msa_asub_u_b(src2, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+
+        ILVRL_B2_UB(ref13, ref03, temp0, temp1);
+        comp0 = __msa_hadd_u_h(temp0, temp0);
+        comp1 = __msa_hadd_u_h(temp1, temp1);
+        comp2 += comp0;
+        comp3 += comp1;
+        SRARI_H2_UH(comp2, comp3, 2);
+        comp = (v16u8) __msa_pckev_b((v16i8) comp3, (v16i8) comp2);
+        diff = __msa_asub_u_b(src3, comp);
+        sad += __msa_hadd_u_h(diff, diff);
+    }
+
+    return (HADD_UH_U32(sad));
+}
+
+#define CALC_MSE_B(src, ref, var)                                    \
+{                                                                    \
+    v16u8 src_l0_m, src_l1_m;                                        \
+    v8i16 res_l0_m, res_l1_m;                                        \
+                                                                     \
+    ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                       \
+    HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);             \
+    DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var);  \
+}
+
+static uint32_t sse_4width_msa(uint8_t *src_ptr, int32_t src_stride,
+                               uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height)
+{
+    int32_t ht_cnt;
+    uint32_t sse;
+    uint32_t src0, src1, src2, src3;
+    uint32_t ref0, ref1, ref2, ref3;
+    v16u8 src = { 0 };
+    v16u8 ref = { 0 };
+    v4i32 var = { 0 };
+
+    for (ht_cnt = (height >> 2); ht_cnt--;) {
+        LW4(src_ptr, src_stride, src0, src1, src2, src3);
+        src_ptr += (4 * src_stride);
+        LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+        ref_ptr += (4 * ref_stride);
+
+        INSERT_W4_UB(src0, src1, src2, src3, src);
+        INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+        CALC_MSE_B(src, ref, var);
+    }
+
+    sse = HADD_SW_S32(var);
+
+    return sse;
+}
+
+static uint32_t sse_8width_msa(uint8_t *src_ptr, int32_t src_stride,
+                               uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height)
+{
+    int32_t ht_cnt;
+    uint32_t sse;
+    v16u8 src0, src1, src2, src3;
+    v16u8 ref0, ref1, ref2, ref3;
+    v4i32 var = { 0 };
+
+    for (ht_cnt = (height >> 2); ht_cnt--;) {
+        LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+        src_ptr += (4 * src_stride);
+        LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+        ref_ptr += (4 * ref_stride);
+
+        PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                    src0, src1, ref0, ref1);
+        CALC_MSE_B(src0, ref0, var);
+        CALC_MSE_B(src1, ref1, var);
+    }
+
+    sse = HADD_SW_S32(var);
+
+    return sse;
+}
+
+static uint32_t sse_16width_msa(uint8_t *src_ptr, int32_t src_stride,
+                                uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height)
+{
+    int32_t ht_cnt;
+    uint32_t sse;
+    v16u8 src, ref;
+    v4i32 var = { 0 };
+
+    for (ht_cnt = (height >> 2); ht_cnt--;) {
+        src = LD_UB(src_ptr);
+        src_ptr += src_stride;
+        ref = LD_UB(ref_ptr);
+        ref_ptr += ref_stride;
+        CALC_MSE_B(src, ref, var);
+
+        src = LD_UB(src_ptr);
+        src_ptr += src_stride;
+        ref = LD_UB(ref_ptr);
+        ref_ptr += ref_stride;
+        CALC_MSE_B(src, ref, var);
+
+        src = LD_UB(src_ptr);
+        src_ptr += src_stride;
+        ref = LD_UB(ref_ptr);
+        ref_ptr += ref_stride;
+        CALC_MSE_B(src, ref, var);
+
+        src = LD_UB(src_ptr);
+        src_ptr += src_stride;
+        ref = LD_UB(ref_ptr);
+        ref_ptr += ref_stride;
+        CALC_MSE_B(src, ref, var);
+    }
+
+    sse = HADD_SW_S32(var);
+
+    return sse;
+}
+
+static int32_t hadamard_diff_8x8_msa(uint8_t *src, int32_t src_stride,
+                                     uint8_t *ref, int32_t ref_stride)
+{
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+    v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+    v8u16 diff0, diff1, diff2, diff3, diff4, diff5, diff6, diff7;
+    v8u16 temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+    v8i16 sum = { 0 };
+    v8i16 zero = { 0 };
+
+    LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    LD_UB8(ref, ref_stride, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
+    ILVR_B8_UH(src0, ref0, src1, ref1, src2, ref2, src3, ref3,
+               src4, ref4, src5, ref5, src6, ref6, src7, ref7,
+               diff0, diff1, diff2, diff3, diff4, diff5, diff6, diff7);
+    HSUB_UB4_UH(diff0, diff1, diff2, diff3, diff0, diff1, diff2, diff3);
+    HSUB_UB4_UH(diff4, diff5, diff6, diff7, diff4, diff5, diff6, diff7);
+    TRANSPOSE8x8_UH_UH(diff0, diff1, diff2, diff3, diff4, diff5, diff6, diff7,
+                       diff0, diff1, diff2, diff3, diff4, diff5, diff6, diff7);
+    BUTTERFLY_8(diff0, diff2, diff4, diff6, diff7, diff5, diff3, diff1,
+                temp0, temp2, temp4, temp6, temp7, temp5, temp3, temp1);
+    BUTTERFLY_8(temp0, temp1, temp4, temp5, temp7, temp6, temp3, temp2,
+                diff0, diff1, diff4, diff5, diff7, diff6, diff3, diff2);
+    BUTTERFLY_8(diff0, diff1, diff2, diff3, diff7, diff6, diff5, diff4,
+                temp0, temp1, temp2, temp3, temp7, temp6, temp5, temp4);
+    TRANSPOSE8x8_UH_UH(temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7,
+                       temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7);
+    BUTTERFLY_8(temp0, temp2, temp4, temp6, temp7, temp5, temp3, temp1,
+                diff0, diff2, diff4, diff6, diff7, diff5, diff3, diff1);
+    BUTTERFLY_8(diff0, diff1, diff4, diff5, diff7, diff6, diff3, diff2,
+                temp0, temp1, temp4, temp5, temp7, temp6, temp3, temp2);
+    ADD4(temp0, temp4, temp1, temp5, temp2, temp6, temp3, temp7,
+         diff0, diff1, diff2, diff3);
+    sum = __msa_asub_s_h((v8i16) temp3, (v8i16) temp7);
+    sum += __msa_asub_s_h((v8i16) temp2, (v8i16) temp6);
+    sum += __msa_asub_s_h((v8i16) temp1, (v8i16) temp5);
+    sum += __msa_asub_s_h((v8i16) temp0, (v8i16) temp4);
+    sum += __msa_add_a_h((v8i16) diff0, zero);
+    sum += __msa_add_a_h((v8i16) diff1, zero);
+    sum += __msa_add_a_h((v8i16) diff2, zero);
+    sum += __msa_add_a_h((v8i16) diff3, zero);
+
+    return (HADD_UH_U32(sum));
+}
+
+static int32_t hadamard_intra_8x8_msa(uint8_t *src, int32_t src_stride,
+                                      uint8_t *ref, int32_t ref_stride)
+{
+    int32_t sum_res = 0;
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+    v8u16 diff0, diff1, diff2, diff3, diff4, diff5, diff6, diff7;
+    v8u16 temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+    v8i16 sum = { 0 };
+    v16i8 zero = { 0 };
+
+    LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    TRANSPOSE8x8_UB_UB(src0, src1, src2, src3, src4, src5, src6, src7,
+                       src0, src1, src2, src3, src4, src5, src6, src7);
+    ILVR_B8_UH(zero, src0, zero, src1, zero, src2, zero, src3,
+               zero, src4, zero, src5, zero, src6, zero, src7,
+               diff0, diff1, diff2, diff3, diff4, diff5, diff6, diff7);
+    BUTTERFLY_8(diff0, diff2, diff4, diff6, diff7, diff5, diff3, diff1,
+                temp0, temp2, temp4, temp6, temp7, temp5, temp3, temp1);
+    BUTTERFLY_8(temp0, temp1, temp4, temp5, temp7, temp6, temp3, temp2,
+                diff0, diff1, diff4, diff5, diff7, diff6, diff3, diff2);
+    BUTTERFLY_8(diff0, diff1, diff2, diff3, diff7, diff6, diff5, diff4,
+                temp0, temp1, temp2, temp3, temp7, temp6, temp5, temp4);
+    TRANSPOSE8x8_UH_UH(temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7,
+                       temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7);
+    BUTTERFLY_8(temp0, temp2, temp4, temp6, temp7, temp5, temp3, temp1,
+                diff0, diff2, diff4, diff6, diff7, diff5, diff3, diff1);
+    BUTTERFLY_8(diff0, diff1, diff4, diff5, diff7, diff6, diff3, diff2,
+                temp0, temp1, temp4, temp5, temp7, temp6, temp3, temp2);
+    ADD4(temp0, temp4, temp1, temp5, temp2, temp6, temp3, temp7,
+         diff0, diff1, diff2, diff3);
+    sum = __msa_asub_s_h((v8i16) temp3, (v8i16) temp7);
+    sum += __msa_asub_s_h((v8i16) temp2, (v8i16) temp6);
+    sum += __msa_asub_s_h((v8i16) temp1, (v8i16) temp5);
+    sum += __msa_asub_s_h((v8i16) temp0, (v8i16) temp4);
+    sum += __msa_add_a_h((v8i16) diff0, (v8i16) zero);
+    sum += __msa_add_a_h((v8i16) diff1, (v8i16) zero);
+    sum += __msa_add_a_h((v8i16) diff2, (v8i16) zero);
+    sum += __msa_add_a_h((v8i16) diff3, (v8i16) zero);
+    sum_res = (HADD_UH_U32(sum));
+    sum_res -= abs(temp0[0] + temp4[0]);
+
+    return sum_res;
+}
+
+int ff_pix_abs16_msa(MpegEncContext *v, uint8_t *src, uint8_t *ref,
+                     ptrdiff_t stride, int height)
+{
+    return sad_16width_msa(src, stride, ref, stride, height);
+}
+
+int ff_pix_abs8_msa(MpegEncContext *v, uint8_t *src, uint8_t *ref,
+                    ptrdiff_t stride, int height)
+{
+    return sad_8width_msa(src, stride, ref, stride, height);
+}
+
+int ff_pix_abs16_x2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                        ptrdiff_t stride, int h)
+{
+    return sad_horiz_bilinear_filter_16width_msa(pix1, stride, pix2, stride, h);
+}
+
+int ff_pix_abs16_y2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                        ptrdiff_t stride, int h)
+{
+    return sad_vert_bilinear_filter_16width_msa(pix1, stride, pix2, stride, h);
+}
+
+int ff_pix_abs16_xy2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                         ptrdiff_t stride, int h)
+{
+    return sad_hv_bilinear_filter_16width_msa(pix1, stride, pix2, stride, h);
+}
+
+int ff_pix_abs8_x2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                       ptrdiff_t stride, int h)
+{
+    return sad_horiz_bilinear_filter_8width_msa(pix1, stride, pix2, stride, h);
+}
+
+int ff_pix_abs8_y2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                       ptrdiff_t stride, int h)
+{
+    return sad_vert_bilinear_filter_8width_msa(pix1, stride, pix2, stride, h);
+}
+
+int ff_pix_abs8_xy2_msa(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                        ptrdiff_t stride, int h)
+{
+    return sad_hv_bilinear_filter_8width_msa(pix1, stride, pix2, stride, h);
+}
+
+int ff_sse16_msa(MpegEncContext *v, uint8_t *src, uint8_t *ref,
+                 ptrdiff_t stride, int height)
+{
+    return sse_16width_msa(src, stride, ref, stride, height);
+}
+
+int ff_sse8_msa(MpegEncContext *v, uint8_t *src, uint8_t *ref,
+                ptrdiff_t stride, int height)
+{
+    return sse_8width_msa(src, stride, ref, stride, height);
+}
+
+int ff_sse4_msa(MpegEncContext *v, uint8_t *src, uint8_t *ref,
+                ptrdiff_t stride, int height)
+{
+    return sse_4width_msa(src, stride, ref, stride, height);
+}
+
+int ff_hadamard8_diff8x8_msa(MpegEncContext *s, uint8_t *dst, uint8_t *src,
+                             ptrdiff_t stride, int h)
+{
+    return hadamard_diff_8x8_msa(src, stride, dst, stride);
+}
+
+int ff_hadamard8_intra8x8_msa(MpegEncContext *s, uint8_t *dst, uint8_t *src,
+                              ptrdiff_t stride, int h)
+{
+    return hadamard_intra_8x8_msa(src, stride, dst, stride);
+}
+
+/* Hadamard Transform functions */
+#define WRAPPER8_16_SQ(name8, name16)                      \
+int name16(MpegEncContext *s, uint8_t *dst, uint8_t *src,  \
+           ptrdiff_t stride, int h)                        \
+{                                                          \
+    int score = 0;                                         \
+    score += name8(s, dst, src, stride, 8);                \
+    score += name8(s, dst + 8, src + 8, stride, 8);        \
+    if(h == 16) {                                          \
+        dst += 8 * stride;                                 \
+        src += 8 * stride;                                 \
+        score +=name8(s, dst, src, stride, 8);             \
+        score +=name8(s, dst + 8, src + 8, stride, 8);     \
+    }                                                      \
+    return score;                                          \
+}
+
+WRAPPER8_16_SQ(ff_hadamard8_diff8x8_msa, ff_hadamard8_diff16_msa);
+WRAPPER8_16_SQ(ff_hadamard8_intra8x8_msa, ff_hadamard8_intra16_msa);
index b1e62b6..d6a2573 100644 (file)
 #define HSUB_UB4_UH(...) HSUB_UB4(v8u16, __VA_ARGS__)
 #define HSUB_UB4_SH(...) HSUB_UB4(v8i16, __VA_ARGS__)
 
+/* Description : SAD (Sum of Absolute Difference)
+   Arguments   : Inputs  - in0, in1, ref0, ref1  (unsigned byte src & ref)
+                 Outputs - sad_m                 (halfword vector with sad)
+                 Return Type - unsigned halfword
+   Details     : Absolute difference of all the byte elements from 'in0' with
+                 'ref0' is calculated and preserved in 'diff0'. From the 16
+                 unsigned absolute diff values, even-odd pairs are added
+                 together to generate 8 halfword results.
+*/
+#define SAD_UB2_UH(in0, in1, ref0, ref1)                        \
+( {                                                             \
+    v16u8 diff0_m, diff1_m;                                     \
+    v8u16 sad_m = { 0 };                                        \
+                                                                \
+    diff0_m = __msa_asub_u_b((v16u8) in0, (v16u8) ref0);        \
+    diff1_m = __msa_asub_u_b((v16u8) in1, (v16u8) ref1);        \
+                                                                \
+    sad_m += __msa_hadd_u_h((v16u8) diff0_m, (v16u8) diff0_m);  \
+    sad_m += __msa_hadd_u_h((v16u8) diff1_m, (v16u8) diff1_m);  \
+                                                                \
+    sad_m;                                                      \
+} )
+
 /* Description : Insert specified word elements from input vectors to 1
                  destination vector
    Arguments   : Inputs  - in0, in1, in2, in3 (4 input vectors)
 }
 #define TRANSPOSE8x8_UB_UB(...) TRANSPOSE8x8_UB(v16u8, __VA_ARGS__)
 #define TRANSPOSE8x8_UB_UH(...) TRANSPOSE8x8_UB(v8u16, __VA_ARGS__)
+
+/* Description : Transposes 16x4 block into 4x16 with byte elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7,
+                           in8, in9, in10, in11, in12, in13, in14, in15
+                 Outputs - out0, out1, out2, out3
+                 Return Type - unsigned byte
+   Details     :
+*/
+#define TRANSPOSE16x4_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7,        \
+                            in8, in9, in10, in11, in12, in13, in14, in15,  \
+                            out0, out1, out2, out3)                        \
+{                                                                          \
+    v2i64 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                  \
+                                                                           \
+    ILVEV_W2_SD(in0, in4, in8, in12, tmp0_m, tmp1_m);                      \
+    out1 = (v16u8) __msa_ilvev_d(tmp1_m, tmp0_m);                          \
+                                                                           \
+    ILVEV_W2_SD(in1, in5, in9, in13, tmp0_m, tmp1_m);                      \
+    out3 = (v16u8) __msa_ilvev_d(tmp1_m, tmp0_m);                          \
+                                                                           \
+    ILVEV_W2_SD(in2, in6, in10, in14, tmp0_m, tmp1_m);                     \
+                                                                           \
+    tmp2_m = __msa_ilvev_d(tmp1_m, tmp0_m);                                \
+    ILVEV_W2_SD(in3, in7, in11, in15, tmp0_m, tmp1_m);                     \
+                                                                           \
+    tmp3_m = __msa_ilvev_d(tmp1_m, tmp0_m);                                \
+    ILVEV_B2_SD(out1, out3, tmp2_m, tmp3_m, tmp0_m, tmp1_m);               \
+    out0 = (v16u8) __msa_ilvev_h((v8i16) tmp1_m, (v8i16) tmp0_m);          \
+    out2 = (v16u8) __msa_ilvod_h((v8i16) tmp1_m, (v8i16) tmp0_m);          \
+                                                                           \
+    tmp0_m = (v2i64) __msa_ilvod_b((v16i8) out3, (v16i8) out1);            \
+    tmp1_m = (v2i64) __msa_ilvod_b((v16i8) tmp3_m, (v16i8) tmp2_m);        \
+    out1 = (v16u8) __msa_ilvev_h((v8i16) tmp1_m, (v8i16) tmp0_m);          \
+    out3 = (v16u8) __msa_ilvod_h((v8i16) tmp1_m, (v8i16) tmp0_m);          \
+}
+
 /* Description : Transposes 16x8 block into 8x16 with byte elements in vectors
    Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7,
                            in8, in9, in10, in11, in12, in13, in14, in15