avcodec/mips: loongson optimize hpeldsp with mmi v1

1.the codes are compatible with O32 ABI
2.use uld and mtc1 to workaround cpu 3A2000 gslwlc1 bug (gslwlc1 instruction extension bug in O32 ABI)

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

libavcodec/mips/Makefile

@@ -77,3 +77,4 @@ MMI-OBJS-$(CONFIG_MPEG4_DECODER)          += mips/xvid_idct_mmi.o
 MMI-OBJS-$(CONFIG_BLOCKDSP)               += mips/blockdsp_mmi.o
 MMI-OBJS-$(CONFIG_PIXBLOCKDSP)            += mips/pixblockdsp_mmi.o
 MMI-OBJS-$(CONFIG_H264QPEL)               += mips/h264qpel_mmi.o
+MMI-OBJS-$(CONFIG_HPELDSP)                += mips/hpeldsp_mmi.o

libavcodec/mips/hpeldsp_init_mips.c

@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2015 Parag Salasakar (Parag.Salasakar@imgtec.com)
+ * Copyright (c) 2016 Zhou Xiaoyong <zhouxiaoyong@loongson.cn>
  *
  * This file is part of FFmpeg.
  *
@@ -65,9 +66,57 @@ static void ff_hpeldsp_init_msa(HpelDSPContext *c, int flags)
 }
 #endif  // #if HAVE_MSA
 
+#if HAVE_MMI
+static void ff_hpeldsp_init_mmi(HpelDSPContext *c, int flags)
+{
+    c->put_pixels_tab[0][0] = ff_put_pixels16_8_mmi;
+    c->put_pixels_tab[0][1] = ff_put_pixels16_x2_8_mmi;
+    c->put_pixels_tab[0][2] = ff_put_pixels16_y2_8_mmi;
+    c->put_pixels_tab[0][3] = ff_put_pixels16_xy2_8_mmi;
+
+    c->put_pixels_tab[1][0] = ff_put_pixels8_8_mmi;
+    c->put_pixels_tab[1][1] = ff_put_pixels8_x2_8_mmi;
+    c->put_pixels_tab[1][2] = ff_put_pixels8_y2_8_mmi;
+    c->put_pixels_tab[1][3] = ff_put_pixels8_xy2_8_mmi;
+
+    c->put_pixels_tab[2][0] = ff_put_pixels4_8_mmi;
+    c->put_pixels_tab[2][1] = ff_put_pixels4_x2_8_mmi;
+    c->put_pixels_tab[2][2] = ff_put_pixels4_y2_8_mmi;
+    c->put_pixels_tab[2][3] = ff_put_pixels4_xy2_8_mmi;
+
+    c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_8_mmi;
+    c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_8_mmi;
+    c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_8_mmi;
+    c->put_no_rnd_pixels_tab[0][3] = ff_put_no_rnd_pixels16_xy2_8_mmi;
+
+    c->put_no_rnd_pixels_tab[1][0] = ff_put_pixels8_8_mmi;
+    c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_8_mmi;
+    c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_8_mmi;
+    c->put_no_rnd_pixels_tab[1][3] = ff_put_no_rnd_pixels8_xy2_8_mmi;
+
+    c->avg_pixels_tab[0][0] = ff_avg_pixels16_8_mmi;
+    c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_8_mmi;
+    c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_8_mmi;
+    c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_8_mmi;
+
+    c->avg_pixels_tab[1][0] = ff_avg_pixels8_8_mmi;
+    c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_8_mmi;
+    c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_8_mmi;
+    c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_8_mmi;
+
+    c->avg_pixels_tab[2][0] = ff_avg_pixels4_8_mmi;
+    c->avg_pixels_tab[2][1] = ff_avg_pixels4_x2_8_mmi;
+    c->avg_pixels_tab[2][2] = ff_avg_pixels4_y2_8_mmi;
+    c->avg_pixels_tab[2][3] = ff_avg_pixels4_xy2_8_mmi;
+}
+#endif  // #if HAVE_MMI
+
 void ff_hpeldsp_init_mips(HpelDSPContext *c, int flags)
 {
 #if HAVE_MSA
     ff_hpeldsp_init_msa(c, flags);
 #endif  // #if HAVE_MSA
+#if HAVE_MMI
+    ff_hpeldsp_init_mmi(c, flags);
+#endif  // #if HAVE_MMI
 }

libavcodec/mips/hpeldsp_mips.h

@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2015 Parag Salasakar (Parag.Salasakar@imgtec.com)
+ * Copyright (c) 2016 Zhou Xiaoyong <zhouxiaoyong@loongson.cn>
  *
  * This file is part of FFmpeg.
  *
@@ -84,4 +85,90 @@ void ff_avg_pixels4_y2_msa(uint8_t *block, const uint8_t *pixels,
 void ff_avg_pixels4_xy2_msa(uint8_t *block, const uint8_t *pixels,
                             ptrdiff_t line_size, int32_t h);
 
+void ff_put_pixels16_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_put_pixels8_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_put_pixels4_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_avg_pixels16_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_avg_pixels8_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_avg_pixels4_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_put_no_rnd_pixels16_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+void ff_put_no_rnd_pixels8_l2_8_mmi(uint8_t *dst, const uint8_t *src1,
+    const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2,
+    int h);
+
+void ff_put_pixels16_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels16_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels16_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels16_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels8_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels8_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels8_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels8_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels4_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels4_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels4_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_pixels4_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_no_rnd_pixels16_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_no_rnd_pixels16_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_no_rnd_pixels16_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_no_rnd_pixels8_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_no_rnd_pixels8_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_put_no_rnd_pixels8_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels16_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels16_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels16_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels16_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels8_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels8_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels8_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels8_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels4_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels4_x2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels4_y2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+void ff_avg_pixels4_xy2_8_mmi(uint8_t *block, const uint8_t *pixels,
+    ptrdiff_t line_size, int32_t h);
+
 #endif  // #ifndef AVCODEC_MIPS_HPELDSP_MIPS_H