avcodec/mips: [loongson] optimize put_hevc_qpel_bi_hv_8 with mmi.

Optimize put_hevc_qpel_bi_hv_8 with mmi in the case width=4/8/12/16/24/32/48/64.
This optimization improved HEVC decoding performance 11.4%(2.01x to 2.24x, tested on loongson 3A3000).

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>

3 files changed, 249 insertions, 1 deletions

diff --git a/libavcodec/mips/hevcdsp_init_mips.c b/libavcodec/mips/hevcdsp_init_mips.c
index e5e0588e5c..18d97d9e0f 100644
--- a/libavcodec/mips/hevcdsp_init_mips.c
+++ b/libavcodec/mips/hevcdsp_init_mips.c
@@ -34,6 +34,15 @@ static av_cold void hevc_dsp_init_mmi(HEVCDSPContext *c,
         c->put_hevc_qpel[8][1][1] = ff_hevc_put_hevc_qpel_hv48_8_mmi;
         c->put_hevc_qpel[9][1][1] = ff_hevc_put_hevc_qpel_hv64_8_mmi;
 
+        c->put_hevc_qpel_bi[1][1][1] = ff_hevc_put_hevc_qpel_bi_hv4_8_mmi;
+        c->put_hevc_qpel_bi[3][1][1] = ff_hevc_put_hevc_qpel_bi_hv8_8_mmi;
+        c->put_hevc_qpel_bi[4][1][1] = ff_hevc_put_hevc_qpel_bi_hv12_8_mmi;
+        c->put_hevc_qpel_bi[5][1][1] = ff_hevc_put_hevc_qpel_bi_hv16_8_mmi;
+        c->put_hevc_qpel_bi[6][1][1] = ff_hevc_put_hevc_qpel_bi_hv24_8_mmi;
+        c->put_hevc_qpel_bi[7][1][1] = ff_hevc_put_hevc_qpel_bi_hv32_8_mmi;
+        c->put_hevc_qpel_bi[8][1][1] = ff_hevc_put_hevc_qpel_bi_hv48_8_mmi;
+        c->put_hevc_qpel_bi[9][1][1] = ff_hevc_put_hevc_qpel_bi_hv64_8_mmi;
+
         c->put_hevc_qpel_bi[3][0][0] = ff_hevc_put_hevc_pel_bi_pixels8_8_mmi;
         c->put_hevc_qpel_bi[5][0][0] = ff_hevc_put_hevc_pel_bi_pixels16_8_mmi;
         c->put_hevc_qpel_bi[6][0][0] = ff_hevc_put_hevc_pel_bi_pixels24_8_mmi;
diff --git a/libavcodec/mips/hevcdsp_mips.h b/libavcodec/mips/hevcdsp_mips.h
index 2351c9b7c3..9f1e44734d 100644
--- a/libavcodec/mips/hevcdsp_mips.h
+++ b/libavcodec/mips/hevcdsp_mips.h
@@ -502,7 +502,7 @@ void ff_hevc_put_hevc_##PEL##_bi_##DIR##WIDTH##_8_##TYPE(uint8_t *dst,
                                                          ptrdiff_t dst_stride,  \
                                                          uint8_t *src,          \
                                                          ptrdiff_t src_stride,  \
-                                                         int16_t *src_16bit,    \
+                                                         int16_t *src2,         \
                                                          int height,            \
                                                          intptr_t mx,           \
                                                          intptr_t my,           \
@@ -514,5 +514,15 @@ L_BI_MC(pel, pixels, 24, mmi);
 L_BI_MC(pel, pixels, 32, mmi);
 L_BI_MC(pel, pixels, 48, mmi);
 L_BI_MC(pel, pixels, 64, mmi);
+
+L_BI_MC(qpel, hv, 4, mmi);
+L_BI_MC(qpel, hv, 8, mmi);
+L_BI_MC(qpel, hv, 12, mmi);
+L_BI_MC(qpel, hv, 16, mmi);
+L_BI_MC(qpel, hv, 24, mmi);
+L_BI_MC(qpel, hv, 32, mmi);
+L_BI_MC(qpel, hv, 48, mmi);
+L_BI_MC(qpel, hv, 64, mmi);
+
 #undef L_BI_MC
 #endif  // #ifndef AVCODEC_MIPS_HEVCDSP_MIPS_H
diff --git a/libavcodec/mips/hevcdsp_mmi.c b/libavcodec/mips/hevcdsp_mmi.c
index e776a13d5c..727a718857 100644
--- a/libavcodec/mips/hevcdsp_mmi.c
+++ b/libavcodec/mips/hevcdsp_mmi.c
@@ -217,6 +217,235 @@ PUT_HEVC_QPEL_HV(32, 8, -32, -64);
 PUT_HEVC_QPEL_HV(48, 12, -48, -96);
 PUT_HEVC_QPEL_HV(64, 16, -64, -128);
 
+#define PUT_HEVC_QPEL_BI_HV(w, x_step, src_step, src2_step, dst_step)   \
+void ff_hevc_put_hevc_qpel_bi_hv##w##_8_mmi(uint8_t *_dst,              \
+                                            ptrdiff_t _dststride,       \
+                                            uint8_t *_src,              \
+                                            ptrdiff_t _srcstride,       \
+                                            int16_t *src2, int height,  \
+                                            intptr_t mx, intptr_t my,   \
+                                            int width)                  \
+{                                                                       \
+    int x, y;                                                           \
+    const int8_t *filter;                                               \
+    pixel *src = (pixel*)_src;                                          \
+    ptrdiff_t srcstride = _srcstride / sizeof(pixel);                   \
+    pixel *dst          = (pixel *)_dst;                                \
+    ptrdiff_t dststride = _dststride / sizeof(pixel);                   \
+    int16_t tmp_array[(MAX_PB_SIZE + QPEL_EXTRA) * MAX_PB_SIZE];        \
+    int16_t *tmp = tmp_array;                                           \
+    uint64_t ftmp[20];                                                  \
+    uint64_t rtmp[1];                                                   \
+    int shift = 7;                                                      \
+    int offset = 64;                                                    \
+                                                                        \
+    src   -= (QPEL_EXTRA_BEFORE * srcstride + 3);                       \
+    filter = ff_hevc_qpel_filters[mx - 1];                              \
+    x = width >> 2;                                                     \
+    y = height + QPEL_EXTRA;                                            \
+    __asm__ volatile(                                                   \
+        MMI_LDC1(%[ftmp1], %[filter], 0x00)                             \
+        "li           %[rtmp0],      0x08                       \n\t"   \
+        "dmtc1        %[rtmp0],      %[ftmp0]                   \n\t"   \
+        "punpckhbh    %[ftmp2],      %[ftmp0],      %[ftmp1]    \n\t"   \
+        "punpcklbh    %[ftmp1],      %[ftmp0],      %[ftmp1]    \n\t"   \
+        "psrah        %[ftmp1],      %[ftmp1],      %[ftmp0]    \n\t"   \
+        "psrah        %[ftmp2],      %[ftmp2],      %[ftmp0]    \n\t"   \
+        "xor          %[ftmp0],      %[ftmp0],      %[ftmp0]    \n\t"   \
+                                                                        \
+        "1:                                                     \n\t"   \
+        "2:                                                     \n\t"   \
+        "gsldlc1      %[ftmp3],      0x07(%[src])               \n\t"   \
+        "gsldrc1      %[ftmp3],      0x00(%[src])               \n\t"   \
+        "gsldlc1      %[ftmp4],      0x08(%[src])               \n\t"   \
+        "gsldrc1      %[ftmp4],      0x01(%[src])               \n\t"   \
+        "gsldlc1      %[ftmp5],      0x09(%[src])               \n\t"   \
+        "gsldrc1      %[ftmp5],      0x02(%[src])               \n\t"   \
+        "gsldlc1      %[ftmp6],      0x0a(%[src])               \n\t"   \
+        "gsldrc1      %[ftmp6],      0x03(%[src])               \n\t"   \
+        "punpcklbh    %[ftmp7],      %[ftmp3],      %[ftmp0]    \n\t"   \
+        "punpckhbh    %[ftmp8],      %[ftmp3],      %[ftmp0]    \n\t"   \
+        "pmullh       %[ftmp7],      %[ftmp7],      %[ftmp1]    \n\t"   \
+        "pmullh       %[ftmp8],      %[ftmp8],      %[ftmp2]    \n\t"   \
+        "paddh        %[ftmp3],      %[ftmp7],      %[ftmp8]    \n\t"   \
+        "punpcklbh    %[ftmp7],      %[ftmp4],      %[ftmp0]    \n\t"   \
+        "punpckhbh    %[ftmp8],      %[ftmp4],      %[ftmp0]    \n\t"   \
+        "pmullh       %[ftmp7],      %[ftmp7],      %[ftmp1]    \n\t"   \
+        "pmullh       %[ftmp8],      %[ftmp8],      %[ftmp2]    \n\t"   \
+        "paddh        %[ftmp4],      %[ftmp7],      %[ftmp8]    \n\t"   \
+        "punpcklbh    %[ftmp7],      %[ftmp5],      %[ftmp0]    \n\t"   \
+        "punpckhbh    %[ftmp8],      %[ftmp5],      %[ftmp0]    \n\t"   \
+        "pmullh       %[ftmp7],      %[ftmp7],      %[ftmp1]    \n\t"   \
+        "pmullh       %[ftmp8],      %[ftmp8],      %[ftmp2]    \n\t"   \
+        "paddh        %[ftmp5],      %[ftmp7],      %[ftmp8]    \n\t"   \
+        "punpcklbh    %[ftmp7],      %[ftmp6],      %[ftmp0]    \n\t"   \
+        "punpckhbh    %[ftmp8],      %[ftmp6],      %[ftmp0]    \n\t"   \
+        "pmullh       %[ftmp7],      %[ftmp7],      %[ftmp1]    \n\t"   \
+        "pmullh       %[ftmp8],      %[ftmp8],      %[ftmp2]    \n\t"   \
+        "paddh        %[ftmp6],      %[ftmp7],      %[ftmp8]    \n\t"   \
+        TRANSPOSE_4H(%[ftmp3], %[ftmp4], %[ftmp5], %[ftmp6],            \
+                     %[ftmp7], %[ftmp8], %[ftmp9], %[ftmp10])           \
+        "paddh        %[ftmp3],      %[ftmp3],      %[ftmp4]    \n\t"   \
+        "paddh        %[ftmp5],      %[ftmp5],      %[ftmp6]    \n\t"   \
+        "paddh        %[ftmp3],      %[ftmp3],      %[ftmp5]    \n\t"   \
+        "gssdlc1      %[ftmp3],      0x07(%[tmp])               \n\t"   \
+        "gssdrc1      %[ftmp3],      0x00(%[tmp])               \n\t"   \
+                                                                        \
+        "daddi        %[x],          %[x],         -0x01        \n\t"   \
+        PTR_ADDIU    "%[src],        %[src],        0x04        \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x08        \n\t"   \
+        "bnez         %[x],          2b                         \n\t"   \
+                                                                        \
+        "daddi        %[y],          %[y],         -0x01        \n\t"   \
+        "li           %[x],        " #x_step "                  \n\t"   \
+        PTR_ADDIU    "%[src],        %[src],      " #src_step " \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],     " #src2_step " \n\t"   \
+        PTR_ADDU     "%[src],        %[src],        %[stride]   \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "bnez         %[y],          1b                         \n\t"   \
+        : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),                 \
+          [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),                 \
+          [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),                 \
+          [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),                 \
+          [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),                 \
+          [ftmp10]"=&f"(ftmp[10]), [rtmp0]"=&r"(rtmp[0]),               \
+          [src]"+&r"(src), [tmp]"+&r"(tmp), [y]"+&r"(y),                \
+          [x]"+&r"(x)                                                   \
+        : [filter]"r"(filter), [stride]"r"(srcstride)                   \
+        : "memory"                                                      \
+    );                                                                  \
+                                                                        \
+    tmp    = tmp_array;                                                 \
+    filter = ff_hevc_qpel_filters[my - 1];                              \
+    x = width >> 2;                                                     \
+    y = height;                                                         \
+    __asm__ volatile(                                                   \
+        MMI_LDC1(%[ftmp1], %[filter], 0x00)                             \
+        "li           %[rtmp0],      0x08                       \n\t"   \
+        "dmtc1        %[rtmp0],      %[ftmp0]                   \n\t"   \
+        "punpckhbh    %[ftmp2],      %[ftmp0],      %[ftmp1]    \n\t"   \
+        "punpcklbh    %[ftmp1],      %[ftmp0],      %[ftmp1]    \n\t"   \
+        "psrah        %[ftmp1],      %[ftmp1],      %[ftmp0]    \n\t"   \
+        "psrah        %[ftmp2],      %[ftmp2],      %[ftmp0]    \n\t"   \
+        "li           %[rtmp0],      0x06                       \n\t"   \
+        "dmtc1        %[rtmp0],      %[ftmp0]                   \n\t"   \
+        "punpcklwd    %[offset],     %[offset],     %[offset]   \n\t"   \
+                                                                        \
+        "1:                                                     \n\t"   \
+        "li           %[x],        " #x_step "                  \n\t"   \
+        "2:                                                     \n\t"   \
+        "gsldlc1      %[ftmp3],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp3],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp4],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp4],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp5],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp5],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp6],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp6],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp7],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp7],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp8],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp8],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp9],      0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp9],      0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "gsldlc1      %[ftmp10],     0x07(%[tmp])               \n\t"   \
+        "gsldrc1      %[ftmp10],     0x00(%[tmp])               \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        -0x380      \n\t"   \
+        TRANSPOSE_4H(%[ftmp3], %[ftmp4], %[ftmp5], %[ftmp6],            \
+                     %[ftmp11], %[ftmp12], %[ftmp13], %[ftmp14])        \
+        TRANSPOSE_4H(%[ftmp7], %[ftmp8], %[ftmp9], %[ftmp10],           \
+                     %[ftmp11], %[ftmp12], %[ftmp13], %[ftmp14])        \
+        "pmaddhw      %[ftmp11],     %[ftmp3],      %[ftmp1]    \n\t"   \
+        "pmaddhw      %[ftmp12],     %[ftmp7],      %[ftmp2]    \n\t"   \
+        "pmaddhw      %[ftmp13],     %[ftmp4],      %[ftmp1]    \n\t"   \
+        "pmaddhw      %[ftmp14],     %[ftmp8],      %[ftmp2]    \n\t"   \
+        "paddw        %[ftmp11],     %[ftmp11],     %[ftmp12]   \n\t"   \
+        "paddw        %[ftmp13],     %[ftmp13],     %[ftmp14]   \n\t"   \
+        TRANSPOSE_2W(%[ftmp11], %[ftmp13], %[ftmp3], %[ftmp4])          \
+        "paddw        %[ftmp3],      %[ftmp3],      %[ftmp4]    \n\t"   \
+        "psraw        %[ftmp3],      %[ftmp3],      %[ftmp0]    \n\t"   \
+        "pmaddhw      %[ftmp11],     %[ftmp5],      %[ftmp1]    \n\t"   \
+        "pmaddhw      %[ftmp12],     %[ftmp9],      %[ftmp2]    \n\t"   \
+        "pmaddhw      %[ftmp13],     %[ftmp6],      %[ftmp1]    \n\t"   \
+        "pmaddhw      %[ftmp14],     %[ftmp10],     %[ftmp2]    \n\t"   \
+        "paddw        %[ftmp11],     %[ftmp11],     %[ftmp12]   \n\t"   \
+        "paddw        %[ftmp13],     %[ftmp13],     %[ftmp14]   \n\t"   \
+        TRANSPOSE_2W(%[ftmp11], %[ftmp13], %[ftmp5], %[ftmp6])          \
+        "paddw        %[ftmp5],      %[ftmp5],      %[ftmp6]    \n\t"   \
+        "psraw        %[ftmp5],      %[ftmp5],      %[ftmp0]    \n\t"   \
+        "packsswh     %[ftmp3],      %[ftmp3],      %[ftmp5]    \n\t"   \
+        "gsldlc1      %[ftmp4],      0x07(%[src2])              \n\t"   \
+        "gsldrc1      %[ftmp4],      0x00(%[src2])              \n\t"   \
+        "xor          %[ftmp7],      %[ftmp7],      %[ftmp7]    \n\t"   \
+        "li           %[rtmp0],      0x10                       \n\t"   \
+        "dmtc1        %[rtmp0],      %[ftmp8]                   \n\t"   \
+        "punpcklhw    %[ftmp5],      %[ftmp7],      %[ftmp3]    \n\t"   \
+        "punpckhhw    %[ftmp6],      %[ftmp7],      %[ftmp3]    \n\t"   \
+        "punpckhhw    %[ftmp3],      %[ftmp7],      %[ftmp4]    \n\t"   \
+        "punpcklhw    %[ftmp4],      %[ftmp7],      %[ftmp4]    \n\t"   \
+        "psraw        %[ftmp5],      %[ftmp5],      %[ftmp8]    \n\t"   \
+        "psraw        %[ftmp6],      %[ftmp6],      %[ftmp8]    \n\t"   \
+        "psraw        %[ftmp3],      %[ftmp3],      %[ftmp8]    \n\t"   \
+        "psraw        %[ftmp4],      %[ftmp4],      %[ftmp8]    \n\t"   \
+        "paddw        %[ftmp5],      %[ftmp5],      %[ftmp4]    \n\t"   \
+        "paddw        %[ftmp6],      %[ftmp6],      %[ftmp3]    \n\t"   \
+        "paddw        %[ftmp5],      %[ftmp5],      %[offset]   \n\t"   \
+        "paddw        %[ftmp6],      %[ftmp6],      %[offset]   \n\t"   \
+        "psraw        %[ftmp5],      %[ftmp5],      %[shift]    \n\t"   \
+        "psraw        %[ftmp6],      %[ftmp6],      %[shift]    \n\t"   \
+        "packsswh     %[ftmp5],      %[ftmp5],      %[ftmp6]    \n\t"   \
+        "pcmpgth      %[ftmp7],      %[ftmp5],      %[ftmp7]    \n\t"   \
+        "and          %[ftmp3],      %[ftmp5],      %[ftmp7]    \n\t"   \
+        "packushb     %[ftmp3],      %[ftmp3],      %[ftmp3]    \n\t"   \
+        "gsswlc1      %[ftmp3],      0x03(%[dst])               \n\t"   \
+        "gsswrc1      %[ftmp3],      0x00(%[dst])               \n\t"   \
+                                                                        \
+        "daddi        %[x],          %[x],         -0x01        \n\t"   \
+        PTR_ADDIU    "%[src2],       %[src2],       0x08        \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x08        \n\t"   \
+        PTR_ADDIU    "%[dst],        %[dst],        0x04        \n\t"   \
+        "bnez         %[x],          2b                         \n\t"   \
+                                                                        \
+        "daddi        %[y],          %[y],         -0x01        \n\t"   \
+        PTR_ADDIU    "%[src2],       %[src2],    " #src2_step " \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],     " #src2_step " \n\t"   \
+        PTR_ADDIU    "%[dst],        %[dst],     " #dst_step "  \n\t"   \
+        PTR_ADDIU    "%[src2],       %[src2],       0x80        \n\t"   \
+        PTR_ADDU     "%[dst],        %[dst],        %[stride]   \n\t"   \
+        PTR_ADDIU    "%[tmp],        %[tmp],        0x80        \n\t"   \
+        "bnez         %[y],          1b                         \n\t"   \
+        : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),                 \
+          [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),                 \
+          [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),                 \
+          [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),                 \
+          [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),                 \
+          [ftmp10]"=&f"(ftmp[10]), [ftmp11]"=&f"(ftmp[11]),             \
+          [ftmp12]"=&f"(ftmp[12]), [ftmp13]"=&f"(ftmp[13]),             \
+          [ftmp14]"=&f"(ftmp[14]), [src2]"+&r"(src2),                   \
+          [dst]"+&r"(dst), [tmp]"+&r"(tmp), [y]"+&r"(y), [x]"=&r"(x),   \
+          [offset]"+&f"(offset), [rtmp0]"=&r"(rtmp[0])                  \
+        : [filter]"r"(filter), [stride]"r"(dststride),                  \
+          [shift]"f"(shift)                                             \
+        : "memory"                                                      \
+    );                                                                  \
+}
+
+PUT_HEVC_QPEL_BI_HV(4, 1, -4, -8, -4);
+PUT_HEVC_QPEL_BI_HV(8, 2, -8, -16, -8);
+PUT_HEVC_QPEL_BI_HV(12, 3, -12, -24, -12);
+PUT_HEVC_QPEL_BI_HV(16, 4, -16, -32, -16);
+PUT_HEVC_QPEL_BI_HV(24, 6, -24, -48, -24);
+PUT_HEVC_QPEL_BI_HV(32, 8, -32, -64, -32);
+PUT_HEVC_QPEL_BI_HV(48, 12, -48, -96, -48);
+PUT_HEVC_QPEL_BI_HV(64, 16, -64, -128, -64);
+
 #define PUT_HEVC_PEL_BI_PIXELS(w, x_step, src_step, dst_step, src2_step)  \
 void ff_hevc_put_hevc_pel_bi_pixels##w##_8_mmi(uint8_t *_dst,             \
                                                ptrdiff_t _dststride,      \