avcodec/mips: [loongson] optimize put_hevc_pel_bi_pixels_8 with mmi.
authorShiyou Yin <yinshiyou-hf@loongson.cn>
Fri, 18 Jan 2019 06:42:20 +0000 (14:42 +0800)
committerMichael Niedermayer <michael@niedermayer.cc>
Sat, 19 Jan 2019 23:44:02 +0000 (00:44 +0100)
Optimize put_hevc_pel_bi_pixels_8 with mmi in the case width=8/16/24/32/48/64.
This optimization improved HEVC decoding performance 2%(1.77x to 1.81x, tested on loongson 3A3000).

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
libavcodec/mips/Makefile
libavcodec/mips/hevcdsp_init_mips.c
libavcodec/mips/hevcdsp_mips.h
libavcodec/mips/hevcdsp_mmi.c [new file with mode: 0644]

index 3571207..3029872 100644 (file)
@@ -86,3 +86,4 @@ MMI-OBJS-$(CONFIG_VP8_DECODER)            += mips/vp8dsp_mmi.o
 MMI-OBJS-$(CONFIG_HPELDSP)                += mips/hpeldsp_mmi.o
 MMI-OBJS-$(CONFIG_VC1_DECODER)            += mips/vc1dsp_mmi.o
 MMI-OBJS-$(CONFIG_WMV2DSP)                += mips/wmv2dsp_mmi.o
+MMI-OBJS-$(CONFIG_HEVC_DECODER)           += mips/hevcdsp_mmi.o
index 776d13e..41c9001 100644 (file)
 
 #include "libavcodec/mips/hevcdsp_mips.h"
 
+#if HAVE_MMI
+static av_cold void hevc_dsp_init_mmi(HEVCDSPContext *c,
+                                      const int bit_depth)
+{
+    if (8 == bit_depth) {
+        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;
+        c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_pel_bi_pixels32_8_mmi;
+        c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_pel_bi_pixels48_8_mmi;
+        c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_pel_bi_pixels64_8_mmi;
+
+        c->put_hevc_epel_bi[3][0][0] = ff_hevc_put_hevc_pel_bi_pixels8_8_mmi;
+        c->put_hevc_epel_bi[5][0][0] = ff_hevc_put_hevc_pel_bi_pixels16_8_mmi;
+        c->put_hevc_epel_bi[6][0][0] = ff_hevc_put_hevc_pel_bi_pixels24_8_mmi;
+        c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_pel_bi_pixels32_8_mmi;
+    }
+}
+#endif // #if HAVE_MMI
+
 #if HAVE_MSA
 static av_cold void hevc_dsp_init_msa(HEVCDSPContext *c,
                                       const int bit_depth)
@@ -448,6 +468,9 @@ static av_cold void hevc_dsp_init_msa(HEVCDSPContext *c,
 
 void ff_hevc_dsp_init_mips(HEVCDSPContext *c, const int bit_depth)
 {
+#if HAVE_MMI
+    hevc_dsp_init_mmi(c, bit_depth);
+#endif  // #if HAVE_MMI
 #if HAVE_MSA
     hevc_dsp_init_msa(c, bit_depth);
 #endif  // #if HAVE_MSA
index 1573d1c..ff9401c 100644 (file)
@@ -479,4 +479,23 @@ void ff_hevc_addblk_32x32_msa(uint8_t *dst, int16_t *pi16Coeffs,
                               ptrdiff_t stride);
 void ff_hevc_idct_luma_4x4_msa(int16_t *pi16Coeffs);
 
+/* Loongson optimization */
+#define L_BI_MC(PEL, DIR, WIDTH, TYPE)                                         \
+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,    \
+                                                        int height,            \
+                                                        intptr_t mx,           \
+                                                        intptr_t my,           \
+                                                        int width)
+
+L_BI_MC(pel, pixels, 8, mmi);
+L_BI_MC(pel, pixels, 16, mmi);
+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);
+#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
new file mode 100644 (file)
index 0000000..60b9c18
--- /dev/null
@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) 2019 Shiyou Yin (yinshiyou-hf@loongson.cn)
+ *
+ * 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 "libavcodec/bit_depth_template.c"
+#include "libavcodec/mips/hevcdsp_mips.h"
+#include "libavutil/mips/mmiutils.h"
+
+#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,      \
+                                               uint8_t *_src,             \
+                                               ptrdiff_t _srcstride,      \
+                                               int16_t *src2, int height, \
+                                               intptr_t mx, intptr_t my,  \
+                                               int width)                 \
+{                                                                         \
+    int x, y;                                                             \
+    pixel *src          = (pixel *)_src;                                  \
+    ptrdiff_t srcstride = _srcstride / sizeof(pixel);                     \
+    pixel *dst          = (pixel *)_dst;                                  \
+    ptrdiff_t dststride = _dststride / sizeof(pixel);                     \
+    uint64_t ftmp[12];                                                    \
+    uint64_t rtmp[1];                                                     \
+    int shift = 7;                                                        \
+                                                                          \
+    y = height;                                                           \
+    x = width >> 3;                                                       \
+    __asm__ volatile(                                                     \
+        "xor          %[ftmp0],      %[ftmp0],      %[ftmp0]    \n\t"     \
+        "li           %[rtmp0],      0x06                       \n\t"     \
+        "dmtc1        %[rtmp0],      %[ftmp1]                   \n\t"     \
+        "li           %[rtmp0],      0x10                       \n\t"     \
+        "dmtc1        %[rtmp0],      %[ftmp10]                  \n\t"     \
+        "li           %[rtmp0],      0x40                       \n\t"     \
+        "dmtc1        %[rtmp0],      %[offset]                  \n\t"     \
+        "punpcklhw    %[offset],     %[offset],     %[offset]   \n\t"     \
+        "punpcklwd    %[offset],     %[offset],     %[offset]   \n\t"     \
+                                                                          \
+        "1:                                                     \n\t"     \
+        "2:                                                     \n\t"     \
+        "gsldlc1      %[ftmp5],      0x07(%[src])               \n\t"     \
+        "gsldrc1      %[ftmp5],      0x00(%[src])               \n\t"     \
+        "gsldlc1      %[ftmp2],      0x07(%[src2])              \n\t"     \
+        "gsldrc1      %[ftmp2],      0x00(%[src2])              \n\t"     \
+        "gsldlc1      %[ftmp3],      0x0f(%[src2])              \n\t"     \
+        "gsldrc1      %[ftmp3],      0x08(%[src2])              \n\t"     \
+        "punpcklbh    %[ftmp4],      %[ftmp5],      %[ftmp0]    \n\t"     \
+        "punpckhbh    %[ftmp5],      %[ftmp5],      %[ftmp0]    \n\t"     \
+        "psllh        %[ftmp4],      %[ftmp4],      %[ftmp1]    \n\t"     \
+        "psllh        %[ftmp5],      %[ftmp5],      %[ftmp1]    \n\t"     \
+        "paddh        %[ftmp4],      %[ftmp4],      %[offset]   \n\t"     \
+        "paddh        %[ftmp5],      %[ftmp5],      %[offset]   \n\t"     \
+        "punpcklhw    %[ftmp6],      %[ftmp4],      %[ftmp0]    \n\t"     \
+        "punpckhhw    %[ftmp7],      %[ftmp4],      %[ftmp0]    \n\t"     \
+        "punpcklhw    %[ftmp8],      %[ftmp5],      %[ftmp0]    \n\t"     \
+        "punpckhhw    %[ftmp9],      %[ftmp5],      %[ftmp0]    \n\t"     \
+        "punpcklhw    %[ftmp4],      %[ftmp0],      %[ftmp3]    \n\t"     \
+        "punpckhhw    %[ftmp5],      %[ftmp0],      %[ftmp3]    \n\t"     \
+        "punpckhhw    %[ftmp3],      %[ftmp0],      %[ftmp2]    \n\t"     \
+        "punpcklhw    %[ftmp2],      %[ftmp0],      %[ftmp2]    \n\t"     \
+        "psraw        %[ftmp2],      %[ftmp2],      %[ftmp10]   \n\t"     \
+        "psraw        %[ftmp3],      %[ftmp3],      %[ftmp10]   \n\t"     \
+        "psraw        %[ftmp4],      %[ftmp4],      %[ftmp10]   \n\t"     \
+        "psraw        %[ftmp5],      %[ftmp5],      %[ftmp10]   \n\t"     \
+        "paddw        %[ftmp2],      %[ftmp2],      %[ftmp6]    \n\t"     \
+        "paddw        %[ftmp3],      %[ftmp3],      %[ftmp7]    \n\t"     \
+        "paddw        %[ftmp4],      %[ftmp4],      %[ftmp8]    \n\t"     \
+        "paddw        %[ftmp5],      %[ftmp5],      %[ftmp9]    \n\t"     \
+        "psraw        %[ftmp2],      %[ftmp2],      %[shift]    \n\t"     \
+        "psraw        %[ftmp3],      %[ftmp3],      %[shift]    \n\t"     \
+        "psraw        %[ftmp4],      %[ftmp4],      %[shift]    \n\t"     \
+        "psraw        %[ftmp5],      %[ftmp5],      %[shift]    \n\t"     \
+        "packsswh     %[ftmp2],      %[ftmp2],      %[ftmp3]    \n\t"     \
+        "packsswh     %[ftmp4],      %[ftmp4],      %[ftmp5]    \n\t"     \
+        "pcmpgth      %[ftmp3],      %[ftmp2],      %[ftmp0]    \n\t"     \
+        "pcmpgth      %[ftmp5],      %[ftmp4],      %[ftmp0]    \n\t"     \
+        "and          %[ftmp2],      %[ftmp2],      %[ftmp3]    \n\t"     \
+        "and          %[ftmp4],      %[ftmp4],      %[ftmp5]    \n\t"     \
+        "packushb     %[ftmp2],      %[ftmp2],      %[ftmp4]    \n\t"     \
+        "gssdlc1      %[ftmp2],      0x07(%[dst])               \n\t"     \
+        "gssdrc1      %[ftmp2],      0x00(%[dst])               \n\t"     \
+                                                                          \
+        "daddi        %[x],          %[x],         -0x01        \n\t"     \
+        PTR_ADDIU    "%[src],        %[src],        0x08        \n\t"     \
+        PTR_ADDIU    "%[dst],        %[dst],        0x08        \n\t"     \
+        PTR_ADDIU    "%[src2],       %[src2],       0x10        \n\t"     \
+        "bnez         %[x],          2b                         \n\t"     \
+                                                                          \
+        PTR_ADDIU    "%[src],        %[src],     " #src_step "  \n\t"     \
+        PTR_ADDIU    "%[dst],        %[dst],     " #dst_step "  \n\t"     \
+        PTR_ADDIU    "%[src2],       %[src2],    " #src2_step " \n\t"     \
+        "li           %[x],        " #x_step "                  \n\t"     \
+        "daddi        %[y],          %[y],         -0x01        \n\t"     \
+        PTR_ADDU     "%[src],        %[src],       %[srcstride] \n\t"     \
+        PTR_ADDU     "%[dst],        %[dst],       %[dststride] \n\t"     \
+        PTR_ADDIU    "%[src2],       %[src2],       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]), [offset]"=&f"(ftmp[11]),               \
+          [src2]"+&r"(src2), [dst]"+&r"(dst), [src]"+&r"(src),            \
+          [x]"+&r"(x), [y]"+&r"(y), [rtmp0]"=&r"(rtmp[0])                 \
+        : [dststride]"r"(dststride), [shift]"f"(shift),                   \
+          [srcstride]"r"(srcstride)                                       \
+        : "memory"                                                        \
+    );                                                                    \
+}                                                                         \
+
+PUT_HEVC_PEL_BI_PIXELS(8, 1, -8, -8, -16);
+PUT_HEVC_PEL_BI_PIXELS(16, 2, -16, -16, -32);
+PUT_HEVC_PEL_BI_PIXELS(24, 3, -24, -24, -48);
+PUT_HEVC_PEL_BI_PIXELS(32, 4, -32, -32, -64);
+PUT_HEVC_PEL_BI_PIXELS(48, 6, -48, -48, -96);
+PUT_HEVC_PEL_BI_PIXELS(64, 8, -64, -64, -128);