avcodec/mips: MSA (MIPS-SIMD-Arch) optimizations for HEVC bi mc functions

This patch adds MSA (MIPS-SIMD-Arch) optimizations for HEVC bi mc functions (qpel as well as epel) in new file hevc_mc_bi_msa.c
Adds new generic macros (needed for this patch) in libavutil/mips/generic_macros_msa.h
Adds HEVC specific macros (needed for this patch) in libavcodec/mips/hevc_macros_msa.h

Signed-off-by: Shivraj Patil <shivraj.patil@imgtec.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/mips/Makefile

@@ -21,6 +21,7 @@ MIPSFPU-OBJS-$(CONFIG_AAC_ENCODER)        += mips/iirfilter_mips.o
 OBJS-$(CONFIG_HEVC_DECODER)               += mips/hevcdsp_init_mips.o
 OBJS-$(CONFIG_H264DSP)                    += mips/h264dsp_init_mips.o
 MSA-OBJS-$(CONFIG_HEVC_DECODER)           += mips/hevcdsp_msa.o            \
-                                             mips/hevc_mc_uni_msa.o
+                                             mips/hevc_mc_uni_msa.o        \
+                                             mips/hevc_mc_bi_msa.o
 MSA-OBJS-$(CONFIG_H264DSP)                += mips/h264dsp_msa.o
 LOONGSON3-OBJS-$(CONFIG_H264DSP)          += mips/h264dsp_mmi.o

libavcodec/mips/hevc_macros_msa.h

@@ -37,6 +37,27 @@
     out = (v4i32) __msa_pckev_b((v16i8) tmp1_m, (v16i8) tmp0_m);  \
 }
 
+#define HEVC_PCK_SW_SB8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1)  \
+{                                                                            \
+    v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                    \
+                                                                             \
+    PCKEV_H4_SH(in0, in1, in2, in3, in4, in5, in6, in7,                      \
+                tmp0_m, tmp1_m, tmp2_m, tmp3_m);                             \
+    PCKEV_B2_SW(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out1);                 \
+}
+
+#define HEVC_PCK_SW_SB12(in0, in1, in2, in3, in4, in5, in6, in7,   \
+                         in8, in9, in10, in11, out0, out1, out2)   \
+{                                                                  \
+    v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m, tmp4_m, tmp5_m;          \
+                                                                   \
+    PCKEV_H4_SH(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                tmp0_m, tmp1_m, tmp2_m, tmp3_m);                   \
+    PCKEV_H2_SH(in8, in9, in10, in11, tmp4_m, tmp5_m);             \
+    PCKEV_B2_SW(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out1);       \
+    out2 = (v4i32) __msa_pckev_b((v16i8) tmp5_m, (v16i8) tmp4_m);  \
+}
+
 #define HEVC_FILT_8TAP(in0, in1, in2, in3,                       \
                        filt0, filt1, filt2, filt3)               \
 ( {                                                              \
@@ -48,4 +69,13 @@
     out_m;                                                       \
 } )
 
+#define HEVC_FILT_4TAP(in0, in1, filt0, filt1)           \
+( {                                                      \
+    v4i32 out_m;                                         \
+                                                         \
+    out_m = __msa_dotp_s_w(in0, (v8i16) filt0);          \
+    out_m = __msa_dpadd_s_w(out_m, in1, (v8i16) filt1);  \
+    out_m;                                               \
+} )
+
 #endif  /* AVCODEC_MIPS_HEVC_MACROS_MSA_H */

libavcodec/mips/hevcdsp_init_mips.c

@@ -96,6 +96,74 @@ static av_cold void hevc_dsp_init_msa(HEVCDSPContext *c,
         c->put_hevc_qpel_uni[7][1][1] = ff_hevc_put_hevc_uni_qpel_hv32_8_msa;
         c->put_hevc_qpel_uni[8][1][1] = ff_hevc_put_hevc_uni_qpel_hv48_8_msa;
         c->put_hevc_qpel_uni[9][1][1] = ff_hevc_put_hevc_uni_qpel_hv64_8_msa;
+
+        c->put_hevc_qpel_bi[1][0][0] = ff_hevc_put_hevc_bi_pel_pixels4_8_msa;
+        c->put_hevc_qpel_bi[3][0][0] = ff_hevc_put_hevc_bi_pel_pixels8_8_msa;
+        c->put_hevc_qpel_bi[4][0][0] = ff_hevc_put_hevc_bi_pel_pixels12_8_msa;
+        c->put_hevc_qpel_bi[5][0][0] = ff_hevc_put_hevc_bi_pel_pixels16_8_msa;
+        c->put_hevc_qpel_bi[6][0][0] = ff_hevc_put_hevc_bi_pel_pixels24_8_msa;
+        c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_msa;
+        c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_msa;
+        c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_msa;
+
+        c->put_hevc_qpel_bi[1][0][1] = ff_hevc_put_hevc_bi_qpel_h4_8_msa;
+        c->put_hevc_qpel_bi[3][0][1] = ff_hevc_put_hevc_bi_qpel_h8_8_msa;
+        c->put_hevc_qpel_bi[4][0][1] = ff_hevc_put_hevc_bi_qpel_h12_8_msa;
+        c->put_hevc_qpel_bi[5][0][1] = ff_hevc_put_hevc_bi_qpel_h16_8_msa;
+        c->put_hevc_qpel_bi[6][0][1] = ff_hevc_put_hevc_bi_qpel_h24_8_msa;
+        c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_hevc_bi_qpel_h32_8_msa;
+        c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_hevc_bi_qpel_h48_8_msa;
+        c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_hevc_bi_qpel_h64_8_msa;
+
+        c->put_hevc_qpel_bi[1][1][0] = ff_hevc_put_hevc_bi_qpel_v4_8_msa;
+        c->put_hevc_qpel_bi[3][1][0] = ff_hevc_put_hevc_bi_qpel_v8_8_msa;
+        c->put_hevc_qpel_bi[4][1][0] = ff_hevc_put_hevc_bi_qpel_v12_8_msa;
+        c->put_hevc_qpel_bi[5][1][0] = ff_hevc_put_hevc_bi_qpel_v16_8_msa;
+        c->put_hevc_qpel_bi[6][1][0] = ff_hevc_put_hevc_bi_qpel_v24_8_msa;
+        c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_hevc_bi_qpel_v32_8_msa;
+        c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_hevc_bi_qpel_v48_8_msa;
+        c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_hevc_bi_qpel_v64_8_msa;
+
+        c->put_hevc_qpel_bi[1][1][1] = ff_hevc_put_hevc_bi_qpel_hv4_8_msa;
+        c->put_hevc_qpel_bi[3][1][1] = ff_hevc_put_hevc_bi_qpel_hv8_8_msa;
+        c->put_hevc_qpel_bi[4][1][1] = ff_hevc_put_hevc_bi_qpel_hv12_8_msa;
+        c->put_hevc_qpel_bi[5][1][1] = ff_hevc_put_hevc_bi_qpel_hv16_8_msa;
+        c->put_hevc_qpel_bi[6][1][1] = ff_hevc_put_hevc_bi_qpel_hv24_8_msa;
+        c->put_hevc_qpel_bi[7][1][1] = ff_hevc_put_hevc_bi_qpel_hv32_8_msa;
+        c->put_hevc_qpel_bi[8][1][1] = ff_hevc_put_hevc_bi_qpel_hv48_8_msa;
+        c->put_hevc_qpel_bi[9][1][1] = ff_hevc_put_hevc_bi_qpel_hv64_8_msa;
+
+        c->put_hevc_epel_bi[1][0][0] = ff_hevc_put_hevc_bi_pel_pixels4_8_msa;
+        c->put_hevc_epel_bi[2][0][0] = ff_hevc_put_hevc_bi_pel_pixels6_8_msa;
+        c->put_hevc_epel_bi[3][0][0] = ff_hevc_put_hevc_bi_pel_pixels8_8_msa;
+        c->put_hevc_epel_bi[4][0][0] = ff_hevc_put_hevc_bi_pel_pixels12_8_msa;
+        c->put_hevc_epel_bi[5][0][0] = ff_hevc_put_hevc_bi_pel_pixels16_8_msa;
+        c->put_hevc_epel_bi[6][0][0] = ff_hevc_put_hevc_bi_pel_pixels24_8_msa;
+        c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_msa;
+
+        c->put_hevc_epel_bi[1][0][1] = ff_hevc_put_hevc_bi_epel_h4_8_msa;
+        c->put_hevc_epel_bi[2][0][1] = ff_hevc_put_hevc_bi_epel_h6_8_msa;
+        c->put_hevc_epel_bi[3][0][1] = ff_hevc_put_hevc_bi_epel_h8_8_msa;
+        c->put_hevc_epel_bi[4][0][1] = ff_hevc_put_hevc_bi_epel_h12_8_msa;
+        c->put_hevc_epel_bi[5][0][1] = ff_hevc_put_hevc_bi_epel_h16_8_msa;
+        c->put_hevc_epel_bi[6][0][1] = ff_hevc_put_hevc_bi_epel_h24_8_msa;
+        c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_hevc_bi_epel_h32_8_msa;
+
+        c->put_hevc_epel_bi[1][1][0] = ff_hevc_put_hevc_bi_epel_v4_8_msa;
+        c->put_hevc_epel_bi[2][1][0] = ff_hevc_put_hevc_bi_epel_v6_8_msa;
+        c->put_hevc_epel_bi[3][1][0] = ff_hevc_put_hevc_bi_epel_v8_8_msa;
+        c->put_hevc_epel_bi[4][1][0] = ff_hevc_put_hevc_bi_epel_v12_8_msa;
+        c->put_hevc_epel_bi[5][1][0] = ff_hevc_put_hevc_bi_epel_v16_8_msa;
+        c->put_hevc_epel_bi[6][1][0] = ff_hevc_put_hevc_bi_epel_v24_8_msa;
+        c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_hevc_bi_epel_v32_8_msa;
+
+        c->put_hevc_epel_bi[1][1][1] = ff_hevc_put_hevc_bi_epel_hv4_8_msa;
+        c->put_hevc_epel_bi[2][1][1] = ff_hevc_put_hevc_bi_epel_hv6_8_msa;
+        c->put_hevc_epel_bi[3][1][1] = ff_hevc_put_hevc_bi_epel_hv8_8_msa;
+        c->put_hevc_epel_bi[4][1][1] = ff_hevc_put_hevc_bi_epel_hv12_8_msa;
+        c->put_hevc_epel_bi[5][1][1] = ff_hevc_put_hevc_bi_epel_hv16_8_msa;
+        c->put_hevc_epel_bi[6][1][1] = ff_hevc_put_hevc_bi_epel_hv24_8_msa;
+        c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_hevc_bi_epel_hv32_8_msa;
     }
 }
 #endif  // #if HAVE_MSA

libavcodec/mips/hevcdsp_mips.h

@@ -116,3 +116,83 @@ UNI_MC(qpel, hv, 48);
 UNI_MC(qpel, hv, 64);
 
 #undef UNI_MC
+
+#define BI_MC(PEL, DIR, WIDTH)                                                 \
+void ff_hevc_put_hevc_bi_##PEL##_##DIR####WIDTH##_8_msa(uint8_t *dst,          \
+                                                        ptrdiff_t dst_stride,  \
+                                                        uint8_t *src,          \
+                                                        ptrdiff_t src_stride,  \
+                                                        int16_t *src_16bit,    \
+                                                        int height,            \
+                                                        intptr_t mx,           \
+                                                        intptr_t my,           \
+                                                        int width)
+
+BI_MC(pel, pixels, 4);
+BI_MC(pel, pixels, 6);
+BI_MC(pel, pixels, 8);
+BI_MC(pel, pixels, 12);
+BI_MC(pel, pixels, 16);
+BI_MC(pel, pixels, 24);
+BI_MC(pel, pixels, 32);
+BI_MC(pel, pixels, 48);
+BI_MC(pel, pixels, 64);
+
+BI_MC(qpel, h, 4);
+BI_MC(qpel, h, 8);
+BI_MC(qpel, h, 12);
+BI_MC(qpel, h, 16);
+BI_MC(qpel, h, 24);
+BI_MC(qpel, h, 32);
+BI_MC(qpel, h, 48);
+BI_MC(qpel, h, 64);
+
+BI_MC(qpel, v, 4);
+BI_MC(qpel, v, 8);
+BI_MC(qpel, v, 12);
+BI_MC(qpel, v, 16);
+BI_MC(qpel, v, 24);
+BI_MC(qpel, v, 32);
+BI_MC(qpel, v, 48);
+BI_MC(qpel, v, 64);
+
+BI_MC(qpel, hv, 4);
+BI_MC(qpel, hv, 8);
+BI_MC(qpel, hv, 12);
+BI_MC(qpel, hv, 16);
+BI_MC(qpel, hv, 24);
+BI_MC(qpel, hv, 32);
+BI_MC(qpel, hv, 48);
+BI_MC(qpel, hv, 64);
+
+BI_MC(epel, h, 4);
+BI_MC(epel, h, 6);
+BI_MC(epel, h, 8);
+BI_MC(epel, h, 12);
+BI_MC(epel, h, 16);
+BI_MC(epel, h, 24);
+BI_MC(epel, h, 32);
+BI_MC(epel, h, 48);
+BI_MC(epel, h, 64);
+
+BI_MC(epel, v, 4);
+BI_MC(epel, v, 6);
+BI_MC(epel, v, 8);
+BI_MC(epel, v, 12);
+BI_MC(epel, v, 16);
+BI_MC(epel, v, 24);
+BI_MC(epel, v, 32);
+BI_MC(epel, v, 48);
+BI_MC(epel, v, 64);
+
+BI_MC(epel, hv, 4);
+BI_MC(epel, hv, 6);
+BI_MC(epel, hv, 8);
+BI_MC(epel, hv, 12);
+BI_MC(epel, hv, 16);
+BI_MC(epel, hv, 24);
+BI_MC(epel, hv, 32);
+BI_MC(epel, hv, 48);
+BI_MC(epel, hv, 64);
+
+#undef BI_MC

libavutil/mips/generic_macros_msa.h

@@ -29,18 +29,23 @@
 #define LD_SB(...) LD_B(v16i8, __VA_ARGS__)
 
 #define LD_H(RTYPE, psrc) *((RTYPE *)(psrc))
+#define LD_UH(...) LD_H(v8u16, __VA_ARGS__)
 #define LD_SH(...) LD_H(v8i16, __VA_ARGS__)
 
 #define LD_W(RTYPE, psrc) *((RTYPE *)(psrc))
+#define LD_UW(...) LD_W(v4u32, __VA_ARGS__)
 #define LD_SW(...) LD_W(v4i32, __VA_ARGS__)
 
 #define ST_B(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
 #define ST_UB(...) ST_B(v16u8, __VA_ARGS__)
+#define ST_SB(...) ST_B(v16i8, __VA_ARGS__)
 
 #define ST_H(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_UH(...) ST_H(v8u16, __VA_ARGS__)
 #define ST_SH(...) ST_H(v8i16, __VA_ARGS__)
 
 #define ST_W(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_UW(...) ST_W(v4u32, __VA_ARGS__)
 #define ST_SW(...) ST_W(v4i32, __VA_ARGS__)
 
 #if (__mips_isa_rev >= 6)
@@ -328,6 +333,46 @@
 #define LD_UB8(...) LD_B8(v16u8, __VA_ARGS__)
 #define LD_SB8(...) LD_B8(v16i8, __VA_ARGS__)
 
+/* Description : Load vectors with 8 halfword elements with stride
+   Arguments   : Inputs  - psrc    (source pointer to load from)
+                         - stride
+                 Outputs - out0, out1
+   Details     : Loads 8 halfword elements in 'out0' from (psrc)
+                 Loads 8 halfword elements in 'out1' from (psrc + stride)
+*/
+#define LD_H2(RTYPE, psrc, stride, out0, out1)  \
+{                                               \
+    out0 = LD_H(RTYPE, (psrc));                 \
+    out1 = LD_H(RTYPE, (psrc) + (stride));      \
+}
+#define LD_UH2(...) LD_H2(v8u16, __VA_ARGS__)
+#define LD_SH2(...) LD_H2(v8i16, __VA_ARGS__)
+
+#define LD_H4(RTYPE, psrc, stride, out0, out1, out2, out3)  \
+{                                                           \
+    LD_H2(RTYPE, (psrc), stride, out0, out1);               \
+    LD_H2(RTYPE, (psrc) + 2 * stride, stride, out2, out3);  \
+}
+#define LD_UH4(...) LD_H4(v8u16, __VA_ARGS__)
+#define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__)
+
+#define LD_H6(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5)  \
+{                                                                       \
+    LD_H4(RTYPE, (psrc), stride, out0, out1, out2, out3);               \
+    LD_H2(RTYPE, (psrc) + 4 * stride, stride, out4, out5);              \
+}
+#define LD_UH6(...) LD_H6(v8u16, __VA_ARGS__)
+#define LD_SH6(...) LD_H6(v8i16, __VA_ARGS__)
+
+#define LD_H8(RTYPE, psrc, stride,                                      \
+              out0, out1, out2, out3, out4, out5, out6, out7)           \
+{                                                                       \
+    LD_H4(RTYPE, (psrc), stride, out0, out1, out2, out3);               \
+    LD_H4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7);  \
+}
+#define LD_UH8(...) LD_H8(v8u16, __VA_ARGS__)
+#define LD_SH8(...) LD_H8(v8i16, __VA_ARGS__)
+
 /* Description : Store vectors of 16 byte elements with stride
    Arguments   : Inputs  - in0, in1, stride
                  Outputs - pdst    (destination pointer to store to)
@@ -478,6 +523,55 @@
                                                                   \
     SW4(out0_m, out1_m, out2_m, out3_m, pblk_4x4_m, stride);      \
 }
+#define ST4x8_UB(in0, in1, pdst, stride)                            \
+{                                                                   \
+    uint8_t *pblk_4x8 = (uint8_t *) (pdst);                         \
+                                                                    \
+    ST4x4_UB(in0, in0, 0, 1, 2, 3, pblk_4x8, stride);               \
+    ST4x4_UB(in1, in1, 0, 1, 2, 3, pblk_4x8 + 4 * stride, stride);  \
+}
+
+/* Description : Store as 6x4 byte block to destination memory from input
+                 vectors
+   Arguments   : Inputs  - in0, in1, pdst, stride
+                 Return Type - unsigned byte
+   Details     : Index 0 word element from input vector 'in0' is copied and
+                 stored on first line followed by index 2 halfword element
+                 Index 2 word element from input vector 'in0' is copied and
+                 stored on second line followed by index 2 halfword element
+                 Index 0 word element from input vector 'in1' is copied and
+                 stored on third line followed by index 2 halfword element
+                 Index 2 word element from input vector 'in1' is copied and
+                 stored on fourth line followed by index 2 halfword element
+*/
+#define ST6x4_UB(in0, in1, pdst, stride)       \
+{                                              \
+    uint32_t out0_m, out1_m, out2_m, out3_m;   \
+    uint16_t out4_m, out5_m, out6_m, out7_m;   \
+    uint8_t *pblk_6x4_m = (uint8_t *) (pdst);  \
+                                               \
+    out0_m = __msa_copy_u_w((v4i32) in0, 0);   \
+    out1_m = __msa_copy_u_w((v4i32) in0, 2);   \
+    out2_m = __msa_copy_u_w((v4i32) in1, 0);   \
+    out3_m = __msa_copy_u_w((v4i32) in1, 2);   \
+                                               \
+    out4_m = __msa_copy_u_h((v8i16) in0, 2);   \
+    out5_m = __msa_copy_u_h((v8i16) in0, 6);   \
+    out6_m = __msa_copy_u_h((v8i16) in1, 2);   \
+    out7_m = __msa_copy_u_h((v8i16) in1, 6);   \
+                                               \
+    SW(out0_m, pblk_6x4_m);                    \
+    SH(out4_m, (pblk_6x4_m + 4));              \
+    pblk_6x4_m += stride;                      \
+    SW(out1_m, pblk_6x4_m);                    \
+    SH(out5_m, (pblk_6x4_m + 4));              \
+    pblk_6x4_m += stride;                      \
+    SW(out2_m, pblk_6x4_m);                    \
+    SH(out6_m, (pblk_6x4_m + 4));              \
+    pblk_6x4_m += stride;                      \
+    SW(out3_m, pblk_6x4_m);                    \
+    SH(out7_m, (pblk_6x4_m + 4));              \
+}
 
 /* Description : Store as 8x2 byte block to destination memory from input vector
    Arguments   : Inputs  - in, pdst, stride
@@ -529,6 +623,15 @@
     ST8x4_UB(in0, in1, pblk_8x8_m, stride);               \
     ST8x4_UB(in2, in3, pblk_8x8_m + 4 * stride, stride);  \
 }
+#define ST12x4_UB(in0, in1, in2, pdst, stride)                \
+{                                                             \
+    uint8_t *pblk_12x4_m = (uint8_t *) (pdst);                \
+                                                              \
+    /* left 8x4 */                                            \
+    ST8x4_UB(in0, in1, pblk_12x4_m, stride);                  \
+    /* right 4x4 */                                           \
+    ST4x4_UB(in2, in2, 0, 1, 2, 3, pblk_12x4_m + 8, stride);  \
+}
 
 /* Description : Store as 12x8 byte block to destination memory from
                  input vectors
@@ -1246,6 +1349,8 @@
 }
 #define PCKEV_B4_SB(...) PCKEV_B4(v16i8, __VA_ARGS__)
 #define PCKEV_B4_UB(...) PCKEV_B4(v16u8, __VA_ARGS__)
+#define PCKEV_B4_SH(...) PCKEV_B4(v8i16, __VA_ARGS__)
+#define PCKEV_B4_SW(...) PCKEV_B4(v4i32, __VA_ARGS__)
 
 /* Description : Pack even halfword elements of vector pairs
    Arguments   : Inputs  - in0, in1, in2, in3
@@ -1317,6 +1422,13 @@
 }
 #define XORI_B5_128_SB(...) XORI_B5_128(v16i8, __VA_ARGS__)
 
+#define XORI_B6_128(RTYPE, in0, in1, in2, in3, in4, in5)  \
+{                                                         \
+    XORI_B4_128(RTYPE, in0, in1, in2, in3);               \
+    XORI_B2_128(RTYPE, in4, in5);                         \
+}
+#define XORI_B6_128_SB(...) XORI_B6_128(v16i8, __VA_ARGS__)
+
 #define XORI_B7_128(RTYPE, in0, in1, in2, in3, in4, in5, in6)  \
 {                                                              \
     XORI_B4_128(RTYPE, in0, in1, in2, in3);                    \
@@ -1442,6 +1554,25 @@
 }
 #define SRAR_H4_UH(...) SRAR_H4(v8u16, __VA_ARGS__)
 #define SRAR_H4_SH(...) SRAR_H4(v8i16, __VA_ARGS__)
+
+/* Description : Shift right arithmetic rounded (immediate)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in0, in1, in2, in3 (in place)
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetic by
+                 value in 'shift'.
+                 The last discarded bit is added to shifted value for rounding
+                 and the result is in place written to 'in0'
+                 Similar for other pairs
+*/
+#define SRARI_H2(RTYPE, in0, in1, shift)              \
+{                                                     \
+    in0 = (RTYPE) __msa_srari_h((v8i16) in0, shift);  \
+    in1 = (RTYPE) __msa_srari_h((v8i16) in1, shift);  \
+}
+#define SRARI_H2_UH(...) SRARI_H2(v8u16, __VA_ARGS__)
+#define SRARI_H2_SH(...) SRARI_H2(v8i16, __VA_ARGS__)
+
 /* Description : Shift right arithmetic rounded (immediate)
    Arguments   : Inputs  - in0, in1, shift
                  Outputs - in0, in1     (in place)
@@ -1499,6 +1630,25 @@
     ILVRL_B2_SH(zero_m, in, out0, out1);              \
 }
 
+/* Description : Sign extend halfword elements from input vector and return
+                 result in pair of vectors
+   Arguments   : Inputs  - in           (1 input halfword vector)
+                 Outputs - out0, out1   (sign extended 2 word vectors)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved right with same vector 'in0' to
+                 generate 4 signed word elements in 'out0'
+                 Then interleaved left with same vector 'in0' to
+                 generate 4 signed word elements in 'out1'
+*/
+#define UNPCK_SH_SW(in, out0, out1)                  \
+{                                                    \
+    v8i16 tmp_m;                                     \
+                                                     \
+    tmp_m = __msa_clti_s_h((v8i16) in, 0);           \
+    ILVRL_H2_SW(tmp_m, in, out0, out1);              \
+}
+
 /* Description : Transposes input 4x4 byte block
    Arguments   : Inputs  - in0, in1, in2, in3      (input 4x4 byte block)
                  Outputs - out0, out1, out2, out3  (output 4x4 byte block)