[FFmpeg-devel] [PATCH] avcodec/mips: msa optimizations for vc1dsp

gxw guxiwei-hf at loongson.cn
Sat Oct 12 05:32:07 EEST 2019


Performance of WMV3 decoding has speed up from 3.66x to 5.23x tested on 3A4000.
---
 libavcodec/mips/Makefile            |   1 +
 libavcodec/mips/vc1dsp_init_mips.c  |  30 ++-
 libavcodec/mips/vc1dsp_mips.h       |  23 ++
 libavcodec/mips/vc1dsp_msa.c        | 483 ++++++++++++++++++++++++++++++++++++
 libavutil/mips/generic_macros_msa.h |   3 +
 5 files changed, 536 insertions(+), 4 deletions(-)
 create mode 100644 libavcodec/mips/vc1dsp_msa.c

diff --git a/libavcodec/mips/Makefile b/libavcodec/mips/Makefile
index c5b54d5..b4993f6 100644
--- a/libavcodec/mips/Makefile
+++ b/libavcodec/mips/Makefile
@@ -89,3 +89,4 @@ MMI-OBJS-$(CONFIG_WMV2DSP)                += mips/wmv2dsp_mmi.o
 MMI-OBJS-$(CONFIG_HEVC_DECODER)           += mips/hevcdsp_mmi.o
 MMI-OBJS-$(CONFIG_VP3DSP)                 += mips/vp3dsp_idct_mmi.o
 MMI-OBJS-$(CONFIG_VP9_DECODER)            += mips/vp9_mc_mmi.o
+MSA-OBJS-$(CONFIG_VC1_DECODER)            += mips/vc1dsp_msa.o
diff --git a/libavcodec/mips/vc1dsp_init_mips.c b/libavcodec/mips/vc1dsp_init_mips.c
index 4adc9e1..c0007ff 100644
--- a/libavcodec/mips/vc1dsp_init_mips.c
+++ b/libavcodec/mips/vc1dsp_init_mips.c
@@ -23,6 +23,10 @@
 #include "vc1dsp_mips.h"
 #include "config.h"
 
+#define FN_ASSIGN(OP, X, Y, INSN) \
+    dsp->OP##vc1_mspel_pixels_tab[1][X+4*Y] = ff_##OP##vc1_mspel_mc##X##Y##INSN; \
+    dsp->OP##vc1_mspel_pixels_tab[0][X+4*Y] = ff_##OP##vc1_mspel_mc##X##Y##_16##INSN
+
 #if HAVE_MMI
 static av_cold void vc1dsp_init_mmi(VC1DSPContext *dsp)
 {
@@ -49,10 +53,6 @@ static av_cold void vc1dsp_init_mmi(VC1DSPContext *dsp)
     dsp->vc1_v_loop_filter16 = ff_vc1_v_loop_filter16_mmi;
     dsp->vc1_h_loop_filter16 = ff_vc1_h_loop_filter16_mmi;
 
-#define FN_ASSIGN(OP, X, Y, INSN) \
-    dsp->OP##vc1_mspel_pixels_tab[1][X+4*Y] = ff_##OP##vc1_mspel_mc##X##Y##INSN; \
-    dsp->OP##vc1_mspel_pixels_tab[0][X+4*Y] = ff_##OP##vc1_mspel_mc##X##Y##_16##INSN
-
     FN_ASSIGN(put_, 0, 0, _mmi);
     FN_ASSIGN(put_, 0, 1, _mmi);
     FN_ASSIGN(put_, 0, 2, _mmi);
@@ -100,9 +100,31 @@ static av_cold void vc1dsp_init_mmi(VC1DSPContext *dsp)
 }
 #endif /* HAVE_MMI */
 
+#if HAVE_MSA
+static av_cold void vc1dsp_init_msa(VC1DSPContext *dsp)
+{
+    dsp->vc1_inv_trans_8x8 = ff_vc1_inv_trans_8x8_msa;
+    dsp->vc1_inv_trans_4x8 = ff_vc1_inv_trans_4x8_msa;
+    dsp->vc1_inv_trans_8x4 = ff_vc1_inv_trans_8x4_msa;
+
+    FN_ASSIGN(put_, 1, 1, _msa);
+    FN_ASSIGN(put_, 1, 2, _msa);
+    FN_ASSIGN(put_, 1, 3, _msa);
+    FN_ASSIGN(put_, 2, 1, _msa);
+    FN_ASSIGN(put_, 2, 2, _msa);
+    FN_ASSIGN(put_, 2, 3, _msa);
+    FN_ASSIGN(put_, 3, 1, _msa);
+    FN_ASSIGN(put_, 3, 2, _msa);
+    FN_ASSIGN(put_, 3, 3, _msa);
+}
+#endif /* HAVE_MSA */
+
 av_cold void ff_vc1dsp_init_mips(VC1DSPContext *dsp)
 {
 #if HAVE_MMI
     vc1dsp_init_mmi(dsp);
 #endif /* HAVE_MMI */
+#if HAVE_MSA
+    vc1dsp_init_msa(dsp);
+#endif /* HAVE_MSA */
 }
diff --git a/libavcodec/mips/vc1dsp_mips.h b/libavcodec/mips/vc1dsp_mips.h
index 0db85fa..5f72e60 100644
--- a/libavcodec/mips/vc1dsp_mips.h
+++ b/libavcodec/mips/vc1dsp_mips.h
@@ -191,4 +191,27 @@ void ff_avg_no_rnd_vc1_chroma_mc4_mmi(uint8_t *dst /* align 8 */,
                                       uint8_t *src /* align 1 */,
                                       int stride, int h, int x, int y);
 
+void ff_vc1_inv_trans_8x8_msa(int16_t block[64]);
+void ff_vc1_inv_trans_8x4_msa(uint8_t *dest, ptrdiff_t linesize, int16_t *block);
+void ff_vc1_inv_trans_4x8_msa(uint8_t *dest, ptrdiff_t linesize, int16_t *block);
+
+#define FF_PUT_VC1_MSPEL_MC_MSA(hmode, vmode)                                 \
+void ff_put_vc1_mspel_mc ## hmode ## vmode ## _msa(uint8_t *dst,              \
+                                                  const uint8_t *src,         \
+                                                  ptrdiff_t stride, int rnd); \
+void ff_put_vc1_mspel_mc ## hmode ## vmode ## _16_msa(uint8_t *dst,           \
+                                                  const uint8_t *src,         \
+                                                  ptrdiff_t stride, int rnd);
+
+FF_PUT_VC1_MSPEL_MC_MSA(1, 1);
+FF_PUT_VC1_MSPEL_MC_MSA(1, 2);
+FF_PUT_VC1_MSPEL_MC_MSA(1, 3);
+
+FF_PUT_VC1_MSPEL_MC_MSA(2, 1);
+FF_PUT_VC1_MSPEL_MC_MSA(2, 2);
+FF_PUT_VC1_MSPEL_MC_MSA(2, 3);
+
+FF_PUT_VC1_MSPEL_MC_MSA(3, 1);
+FF_PUT_VC1_MSPEL_MC_MSA(3, 2);
+FF_PUT_VC1_MSPEL_MC_MSA(3, 3);
 #endif /* AVCODEC_MIPS_VC1DSP_MIPS_H */
diff --git a/libavcodec/mips/vc1dsp_msa.c b/libavcodec/mips/vc1dsp_msa.c
new file mode 100644
index 0000000..1619ea4
--- /dev/null
+++ b/libavcodec/mips/vc1dsp_msa.c
@@ -0,0 +1,483 @@
+/*
+ * Loongson SIMD optimized vc1dsp
+ *
+ * Copyright (c) 2019 Loongson Technology Corporation Limited
+ *                    gxw <guxiwei-hf at 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 "vc1dsp_mips.h"
+#include "constants.h"
+#include "libavutil/mips/generic_macros_msa.h"
+
+void ff_vc1_inv_trans_8x8_msa(int16_t block[64])
+{
+    v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+    v4i32 in_r0, in_r1, in_r2, in_r3, in_r4, in_r5, in_r6, in_r7;
+    v4i32 in_l0, in_l1, in_l2, in_l3, in_l4, in_l5, in_l6, in_l7;
+    v4i32 t_r1, t_r2, t_r3, t_r4, t_r5, t_r6, t_r7, t_r8;
+    v4i32 t_l1, t_l2, t_l3, t_l4, t_l5, t_l6, t_l7, t_l8;
+    v4i32 cnst_12 = {12, 12, 12, 12};
+    v4i32 cnst_4 = {4, 4, 4, 4};
+    v4i32 cnst_16 = {16, 16, 16, 16};
+    v4i32 cnst_6 = {6, 6, 6, 6};
+    v4i32 cnst_15 = {15, 15, 15, 15};
+    v4i32 cnst_9 = {9, 9, 9, 9};
+    v4i32 cnst_1 = {1, 1, 1, 1};
+    v4i32 cnst_64 = {64, 64, 64, 64};
+
+    LD_SH8(block, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+    UNPCK_SH_SW(in0, in_r0, in_l0);
+    UNPCK_SH_SW(in1, in_r1, in_l1);
+    UNPCK_SH_SW(in2, in_r2, in_l2);
+    UNPCK_SH_SW(in3, in_r3, in_l3);
+    UNPCK_SH_SW(in4, in_r4, in_l4);
+    UNPCK_SH_SW(in5, in_r5, in_l5);
+    UNPCK_SH_SW(in6, in_r6, in_l6);
+    UNPCK_SH_SW(in7, in_r7, in_l7);
+    // First loop
+    t_r1 = cnst_12 * (in_r0 + in_r4) + cnst_4;
+    t_l1 = cnst_12 * (in_l0 + in_l4) + cnst_4;
+    t_r2 = cnst_12 * (in_r0 - in_r4) + cnst_4;
+    t_l2 = cnst_12 * (in_l0 - in_l4) + cnst_4;
+    t_r3 = cnst_16 * in_r2 + cnst_6 * in_r6;
+    t_l3 = cnst_16 * in_l2 + cnst_6 * in_l6;
+    t_r4 = cnst_6 * in_r2 - cnst_16 * in_r6;
+    t_l4 = cnst_6 * in_l2 - cnst_16 * in_l6;
+
+    ADD4(t_r1, t_r3, t_l1, t_l3, t_r2, t_r4, t_l2, t_l4, t_r5, t_l5, t_r6, t_l6);
+    SUB4(t_r2, t_r4, t_l2, t_l4, t_r1, t_r3, t_l1, t_l3, t_r7, t_l7, t_r8, t_l8);
+    t_r1 = cnst_16 * in_r1 + cnst_15 * in_r3 + cnst_9 * in_r5 + cnst_4 * in_r7;
+    t_l1 = cnst_16 * in_l1 + cnst_15 * in_l3 + cnst_9 * in_l5 + cnst_4 * in_l7;
+    t_r2 = cnst_15 * in_r1 - cnst_4 * in_r3 - cnst_16 * in_r5 - cnst_9 * in_r7;
+    t_l2 = cnst_15 * in_l1 - cnst_4 * in_l3 - cnst_16 * in_l5 - cnst_9 * in_l7;
+    t_r3 = cnst_9 * in_r1 - cnst_16 * in_r3 + cnst_4 * in_r5 + cnst_15 * in_r7;
+    t_l3 = cnst_9 * in_l1 - cnst_16 * in_l3 + cnst_4 * in_l5 + cnst_15 * in_l7;
+    t_r4 = cnst_4 * in_r1 - cnst_9 * in_r3 + cnst_15 * in_r5 - cnst_16 * in_r7;
+    t_l4 = cnst_4 * in_l1 - cnst_9 * in_l3 + cnst_15 * in_l5 - cnst_16 * in_l7;
+
+    in_r0 = (t_r5 + t_r1) >> 3;
+    in_l0 = (t_l5 + t_l1) >> 3;
+    in_r1 = (t_r6 + t_r2) >> 3;
+    in_l1 = (t_l6 + t_l2) >> 3;
+    in_r2 = (t_r7 + t_r3) >> 3;
+    in_l2 = (t_l7 + t_l3) >> 3;
+    in_r3 = (t_r8 + t_r4) >> 3;
+    in_l3 = (t_l8 + t_l4) >> 3;
+
+    in_r4 = (t_r8 - t_r4) >> 3;
+    in_l4 = (t_l8 - t_l4) >> 3;
+    in_r5 = (t_r7 - t_r3) >> 3;
+    in_l5 = (t_l7 - t_l3) >> 3;
+    in_r6 = (t_r6 - t_r2) >> 3;
+    in_l6 = (t_l6 - t_l2) >> 3;
+    in_r7 = (t_r5 - t_r1) >> 3;
+    in_l7 = (t_l5 - t_l1) >> 3;
+    TRANSPOSE4x4_SW_SW(in_r0, in_r1, in_r2, in_r3, in_r0, in_r1, in_r2, in_r3);
+    TRANSPOSE4x4_SW_SW(in_l0, in_l1, in_l2, in_l3, t_l1, t_l2, t_l3, t_l4);
+    TRANSPOSE4x4_SW_SW(in_r4, in_r5, in_r6, in_r7, in_l0, in_l1, in_l2, in_l3);
+    TRANSPOSE4x4_SW_SW(in_l4, in_l5, in_l6, in_l7, in_l4, in_l5, in_l6, in_l7);
+    in_r4 = t_l1, in_r5 = t_l2, in_r6 = t_l3, in_r7 = t_l4;
+    // Second loop
+    t_r1 = cnst_12 * (in_r0 + in_r4) + cnst_64;
+    t_l1 = cnst_12 * (in_l0 + in_l4) + cnst_64;
+    t_r2 = cnst_12 * (in_r0 - in_r4) + cnst_64;
+    t_l2 = cnst_12 * (in_l0 - in_l4) + cnst_64;
+    t_r3 = cnst_16 * in_r2 + cnst_6 * in_r6;
+    t_l3 = cnst_16 * in_l2 + cnst_6 * in_l6;
+    t_r4 = cnst_6 * in_r2 - cnst_16 * in_r6;
+    t_l4 = cnst_6 * in_l2 - cnst_16 * in_l6;
+
+    ADD4(t_r1, t_r3, t_l1, t_l3, t_r2, t_r4, t_l2, t_l4, t_r5, t_l5, t_r6, t_l6);
+    SUB4(t_r2, t_r4, t_l2, t_l4, t_r1, t_r3, t_l1, t_l3, t_r7, t_l7, t_r8, t_l8);
+    t_r1 = cnst_16 * in_r1 + cnst_15 * in_r3 + cnst_9 * in_r5 + cnst_4 * in_r7;
+    t_l1 = cnst_16 * in_l1 + cnst_15 * in_l3 + cnst_9 * in_l5 + cnst_4 * in_l7;
+    t_r2 = cnst_15 * in_r1 - cnst_4 * in_r3 - cnst_16 * in_r5 - cnst_9 * in_r7;
+    t_l2 = cnst_15 * in_l1 - cnst_4 * in_l3 - cnst_16 * in_l5 - cnst_9 * in_l7;
+    t_r3 = cnst_9 * in_r1 - cnst_16 * in_r3 + cnst_4 * in_r5 + cnst_15 * in_r7;
+    t_l3 = cnst_9 * in_l1 - cnst_16 * in_l3 + cnst_4 * in_l5 + cnst_15 * in_l7;
+    t_r4 = cnst_4 * in_r1 - cnst_9 * in_r3 + cnst_15 * in_r5 - cnst_16 * in_r7;
+    t_l4 = cnst_4 * in_l1 - cnst_9 * in_l3 + cnst_15 * in_l5 - cnst_16 * in_l7;
+
+    in_r0 = (t_r5 + t_r1) >> 7;
+    in_l0 = (t_l5 + t_l1) >> 7;
+    in_r1 = (t_r6 + t_r2) >> 7;
+    in_l1 = (t_l6 + t_l2) >> 7;
+    in_r2 = (t_r7 + t_r3) >> 7;
+    in_l2 = (t_l7 + t_l3) >> 7;
+    in_r3 = (t_r8 + t_r4) >> 7;
+    in_l3 = (t_l8 + t_l4) >> 7;
+
+    in_r4 = (t_r8 - t_r4 + cnst_1) >> 7;
+    in_l4 = (t_l8 - t_l4 + cnst_1) >> 7;
+    in_r5 = (t_r7 - t_r3 + cnst_1) >> 7;
+    in_l5 = (t_l7 - t_l3 + cnst_1) >> 7;
+    in_r6 = (t_r6 - t_r2 + cnst_1) >> 7;
+    in_l6 = (t_l6 - t_l2 + cnst_1) >> 7;
+    in_r7 = (t_r5 - t_r1 + cnst_1) >> 7;
+    in_l7 = (t_l5 - t_l1 + cnst_1) >> 7;
+    PCKEV_H4_SH(in_l0, in_r0, in_l1, in_r1, in_l2, in_r2, in_l3, in_r3,
+                in0, in1, in2, in3);
+    PCKEV_H4_SH(in_l4, in_r4, in_l5, in_r5, in_l6, in_r6, in_l7, in_r7,
+                in4, in5, in6, in7);
+    ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, block, 8);
+}
+
+void ff_vc1_inv_trans_4x8_msa(uint8_t *dest, ptrdiff_t linesize, int16_t *block)
+{
+    v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+    v4i32 in_r0, in_r1, in_r2, in_r3, in_r4, in_r5, in_r6, in_r7;
+    v4i32 t1, t2, t3, t4, t5, t6, t7, t8;
+    v4i32 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+    v16i8 zero_m = { 0 };
+    v4i32 cnst_17 = {17, 17, 17, 17};
+    v4i32 cnst_22 = {22, 22, 22, 22};
+    v4i32 cnst_10 = {10, 10, 10, 10};
+    v4i32 cnst_12 = {12, 12, 12, 12};
+    v4i32 cnst_64 = {64, 64, 64, 64};
+    v4i32 cnst_16 = {16, 16, 16, 16};
+    v4i32 cnst_15 = {15, 15, 15, 15};
+    v4i32 cnst_4 = {4, 4, 4, 4};
+    v4i32 cnst_6 = {6, 6, 6, 6};
+    v4i32 cnst_9 = {9, 9, 9, 9};
+    v4i32 cnst_1 = {1, 1, 1, 1};
+
+    LD_SH8(block, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+    UNPCK_R_SH_SW(in0, in_r0);
+    UNPCK_R_SH_SW(in1, in_r1);
+    UNPCK_R_SH_SW(in2, in_r2);
+    UNPCK_R_SH_SW(in3, in_r3);
+    UNPCK_R_SH_SW(in4, in_r4);
+    UNPCK_R_SH_SW(in5, in_r5);
+    UNPCK_R_SH_SW(in6, in_r6);
+    UNPCK_R_SH_SW(in7, in_r7);
+    // First loop
+    TRANSPOSE4x4_SW_SW(in_r0, in_r1, in_r2, in_r3, in_r0, in_r1, in_r2, in_r3);
+    TRANSPOSE4x4_SW_SW(in_r4, in_r5, in_r6, in_r7, in_r4, in_r5, in_r6, in_r7);
+    t1 = cnst_17 * (in_r0 + in_r2) + cnst_4;
+    t5 = cnst_17 * (in_r4 + in_r6) + cnst_4;
+    t2 = cnst_17 * (in_r0 - in_r2) + cnst_4;
+    t6 = cnst_17 * (in_r4 - in_r6) + cnst_4;
+    t3 = cnst_22 * in_r1 + cnst_10 * in_r3;
+    t7 = cnst_22 * in_r5 + cnst_10 * in_r7;
+    t4 = cnst_22 * in_r3 - cnst_10 * in_r1;
+    t8 = cnst_22 * in_r7 - cnst_10 * in_r5;
+
+    in_r0 = (t1 + t3) >> 3;
+    in_r4 = (t5 + t7) >> 3;
+    in_r1 = (t2 - t4) >> 3;
+    in_r5 = (t6 - t8) >> 3;
+    in_r2 = (t2 + t4) >> 3;
+    in_r6 = (t6 + t8) >> 3;
+    in_r3 = (t1 - t3) >> 3;
+    in_r7 = (t5 - t7) >> 3;
+    TRANSPOSE4x4_SW_SW(in_r0, in_r1, in_r2, in_r3, in_r0, in_r1, in_r2, in_r3);
+    TRANSPOSE4x4_SW_SW(in_r4, in_r5, in_r6, in_r7, in_r4, in_r5, in_r6, in_r7);
+    PCKEV_H4_SH(in_r1, in_r0, in_r3, in_r2, in_r5, in_r4, in_r7, in_r6,
+                in0, in1, in2, in3);
+    ST_D8(in0, in1, in2, in3, 0, 1, 0, 1, 0, 1, 0, 1, block, 8);
+    // Second loop
+    t1 = cnst_12 * (in_r0 + in_r4) + cnst_64;
+    t2 = cnst_12 * (in_r0 - in_r4) + cnst_64;
+    t3 = cnst_16 * in_r2 + cnst_6 * in_r6;
+    t4 = cnst_6 * in_r2 - cnst_16 * in_r6;
+    t5 = t1 + t3, t6 = t2 + t4;
+    t7 = t2 - t4, t8 = t1 - t3;
+    t1 = cnst_16 * in_r1 + cnst_15 * in_r3 + cnst_9 * in_r5 + cnst_4 * in_r7;
+    t2 = cnst_15 * in_r1 - cnst_4 * in_r3 - cnst_16 * in_r5 - cnst_9 * in_r7;
+    t3 = cnst_9 * in_r1 - cnst_16 * in_r3 + cnst_4 * in_r5 + cnst_15 * in_r7;
+    t4 = cnst_4 * in_r1 - cnst_9 * in_r3 + cnst_15 * in_r5 - cnst_16 * in_r7;
+    LD_SW8(dest, linesize, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+    ILVR_B8_SW(zero_m, dst0, zero_m, dst1, zero_m, dst2, zero_m, dst3,
+               zero_m, dst4, zero_m, dst5, zero_m, dst6, zero_m, dst7,
+               dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+    ILVR_H4_SW(zero_m, dst0, zero_m, dst1, zero_m, dst2, zero_m, dst3,
+               dst0, dst1, dst2, dst3);
+    ILVR_H4_SW(zero_m, dst4, zero_m, dst5, zero_m, dst6, zero_m, dst7,
+               dst4, dst5, dst6, dst7);
+    in_r0 = (t5 + t1) >> 7;
+    in_r1 = (t6 + t2) >> 7;
+    in_r2 = (t7 + t3) >> 7;
+    in_r3 = (t8 + t4) >> 7;
+    in_r4 = (t8 - t4 + cnst_1) >> 7;
+    in_r5 = (t7 - t3 + cnst_1) >> 7;
+    in_r6 = (t6 - t2 + cnst_1) >> 7;
+    in_r7 = (t5 - t1 + cnst_1) >> 7;
+    ADD4(in_r0, dst0, in_r1, dst1, in_r2, dst2, in_r3, dst3,
+         in_r0, in_r1, in_r2, in_r3);
+    ADD4(in_r4, dst4, in_r5, dst5, in_r6, dst6, in_r7, dst7,
+         in_r4, in_r5, in_r6, in_r7);
+    CLIP_SW8_0_255(in_r0, in_r1, in_r2, in_r3, in_r4, in_r5, in_r6, in_r7);
+    PCKEV_H4_SH(in_r1, in_r0, in_r3, in_r2, in_r5, in_r4, in_r7, in_r6,
+                in0, in1, in2, in3);
+    PCKEV_B2_SH(in1, in0, in3, in2, in0, in1);
+    ST_W8(in0, in1, 0, 1, 2, 3, 0, 1, 2, 3, dest, linesize);
+}
+
+void ff_vc1_inv_trans_8x4_msa(uint8_t *dest, ptrdiff_t linesize, int16_t *block)
+{
+    v4i32 in0, in1, in2, in3, in4, in5, in6, in7;
+    v4i32 t1, t2, t3, t4, t5, t6, t7, t8;
+    v4i32 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+    v16i8 zero_m = { 0 };
+    v4i32 cnst_17 = {17, 17, 17, 17};
+    v4i32 cnst_22 = {22, 22, 22, 22};
+    v4i32 cnst_10 = {10, 10, 10, 10};
+    v4i32 cnst_12 = {12, 12, 12, 12};
+    v4i32 cnst_64 = {64, 64, 64, 64};
+    v4i32 cnst_16 = {16, 16, 16, 16};
+    v4i32 cnst_15 = {15, 15, 15, 15};
+    v4i32 cnst_4 = {4, 4, 4, 4};
+    v4i32 cnst_6 = {6, 6, 6, 6};
+    v4i32 cnst_9 = {9, 9, 9, 9};
+
+    LD_SW4(block, 8, t1, t2, t3, t4);
+    UNPCK_SH_SW(t1, in0, in4);
+    UNPCK_SH_SW(t2, in1, in5);
+    UNPCK_SH_SW(t3, in2, in6);
+    UNPCK_SH_SW(t4, in3, in7);
+    TRANSPOSE4x4_SW_SW(in0, in1, in2, in3, in0, in1, in2, in3);
+    TRANSPOSE4x4_SW_SW(in4, in5, in6, in7, in4, in5, in6, in7);
+    // First loop
+    t1 = cnst_12 * (in0 + in4) + cnst_4;
+    t2 = cnst_12 * (in0 - in4) + cnst_4;
+    t3 = cnst_16 * in2 + cnst_6 * in6;
+    t4 = cnst_6 * in2 - cnst_16 * in6;
+    t5 = t1 + t3, t6 = t2 + t4;
+    t7 = t2 - t4, t8 = t1 - t3;
+    t1 = cnst_16 * in1 + cnst_15 * in3 + cnst_9 * in5 + cnst_4 * in7;
+    t2 = cnst_15 * in1 - cnst_4 * in3 - cnst_16 * in5 - cnst_9 * in7;
+    t3 = cnst_9 * in1 - cnst_16 * in3 + cnst_4 * in5 + cnst_15 * in7;
+    t4 = cnst_4 * in1 - cnst_9 * in3 + cnst_15 * in5 - cnst_16 * in7;
+    in0 = (t5 + t1) >> 3;
+    in1 = (t6 + t2) >> 3;
+    in2 = (t7 + t3) >> 3;
+    in3 = (t8 + t4) >> 3;
+    in4 = (t8 - t4) >> 3;
+    in5 = (t7 - t3) >> 3;
+    in6 = (t6 - t2) >> 3;
+    in7 = (t5 - t1) >> 3;
+    TRANSPOSE4x4_SW_SW(in0, in1, in2, in3, in0, in1, in2, in3);
+    TRANSPOSE4x4_SW_SW(in4, in5, in6, in7, in4, in5, in6, in7);
+    PCKEV_H4_SW(in4, in0, in5, in1, in6, in2, in7, in3, t1, t2, t3, t4);
+    ST_SW4(t1, t2, t3, t4, block, 8);
+    // Second loop
+    LD_SW4(dest, linesize, dst0, dst1, dst2, dst3);
+    ILVR_B4_SW(zero_m, dst0, zero_m, dst1, zero_m, dst2, zero_m, dst3,
+               dst0, dst1, dst2, dst3);
+    ILVL_H4_SW(zero_m, dst0, zero_m, dst1, zero_m, dst2, zero_m, dst3,
+               dst4, dst5, dst6, dst7);
+    ILVR_H4_SW(zero_m, dst0, zero_m, dst1, zero_m, dst2, zero_m, dst3,
+               dst0, dst1, dst2, dst3);
+    // Right part
+    t1 = cnst_17 * (in0 + in2) + cnst_64;
+    t2 = cnst_17 * (in0 - in2) + cnst_64;
+    t3 = cnst_22 * in1 + cnst_10 * in3;
+    t4 = cnst_22 * in3 - cnst_10 * in1;
+    in0 = (t1 + t3) >> 7;
+    in1 = (t2 - t4) >> 7;
+    in2 = (t2 + t4) >> 7;
+    in3 = (t1 - t3) >> 7;
+    ADD4(in0, dst0, in1, dst1, in2, dst2, in3, dst3, in0, in1, in2, in3);
+    CLIP_SW4_0_255(in0, in1, in2, in3);
+    // Left part
+    t5 = cnst_17 * (in4 + in6) + cnst_64;
+    t6 = cnst_17 * (in4 - in6) + cnst_64;
+    t7 = cnst_22 * in5 + cnst_10 * in7;
+    t8 = cnst_22 * in7 - cnst_10 * in5;
+    in4 = (t5 + t7) >> 7;
+    in5 = (t6 - t8) >> 7;
+    in6 = (t6 + t8) >> 7;
+    in7 = (t5 - t7) >> 7;
+    ADD4(in4, dst4, in5, dst5, in6, dst6, in7, dst7, in4, in5, in6, in7);
+    CLIP_SW4_0_255(in4, in5, in6, in7);
+    PCKEV_H4_SW(in4, in0, in5, in1, in6, in2, in7, in3, in0, in1, in2, in3);
+    PCKEV_B2_SW(in1, in0, in3, in2, in0, in1);
+    ST_D4(in0, in1, 0, 1, 0, 1, dest, linesize);
+}
+
+static void put_vc1_mspel_mc_h_v_msa(uint8_t *dst, const uint8_t *src,
+                                     ptrdiff_t stride, int hmode, int vmode,
+                                     int rnd)
+{
+    v8i16 in_r0, in_r1, in_r2, in_r3, in_l0, in_l1, in_l2, in_l3;
+    v8i16 t0, t1, t2, t3, t4, t5, t6, t7;
+    v8i16 t8, t9, t10, t11, t12, t13, t14, t15;
+    v8i16 cnst_para0, cnst_para1, cnst_para2, cnst_para3, cnst_r;
+    static const int shift_value[] = {0, 5, 1, 5};
+    int shift = (shift_value[hmode] + shift_value[vmode]) >> 1;
+    int r = (1 << (shift - 1)) + rnd - 1;
+    cnst_r = __msa_fill_h(r);
+    src -= 1, src -= stride;
+    if (vmode == 1) {
+      cnst_para0 = __msa_ldi_h(4);
+      cnst_para1 = __msa_ldi_h(53);
+      cnst_para2 = __msa_ldi_h(18);
+      cnst_para3 = __msa_ldi_h(3);
+    } else if (vmode == 2) {
+      cnst_para0 = __msa_ldi_h(1);
+      cnst_para1 = __msa_ldi_h(9);
+      cnst_para2 = __msa_ldi_h(9);
+      cnst_para3 = __msa_ldi_h(1);
+    } else {
+      cnst_para0 = __msa_ldi_h(3);
+      cnst_para1 = __msa_ldi_h(18);
+      cnst_para2 = __msa_ldi_h(53);
+      cnst_para3 = __msa_ldi_h(4);
+    }
+    LD_SH4(src, stride, in_l0, in_l1, in_l2, in_l3);
+    UNPCK_UB_SH(in_l0, in_r0, in_l0);
+    UNPCK_UB_SH(in_l1, in_r1, in_l1);
+    UNPCK_UB_SH(in_l2, in_r2, in_l2);
+    UNPCK_UB_SH(in_l3, in_r3, in_l3);
+    // row 0
+    t0 = cnst_para1 * in_r1 + cnst_para2 * in_r2
+         - cnst_para0 * in_r0 - cnst_para3 * in_r3;
+    t8 = cnst_para1 * in_l1 + cnst_para2 * in_l2
+         - cnst_para0 * in_l0 - cnst_para3 * in_l3;
+    in_l0 = LD_SH(src + 4 * stride);
+    UNPCK_UB_SH(in_l0, in_r0, in_l0);
+    // row 1
+    t1 = cnst_para1 * in_r2 + cnst_para2 * in_r3
+         - cnst_para0 * in_r1 - cnst_para3 * in_r0;
+    t9 = cnst_para1 * in_l2 + cnst_para2 * in_l3
+         - cnst_para0 * in_l1 - cnst_para3 * in_l0;
+    in_l1 = LD_SH(src + 5 * stride);
+    UNPCK_UB_SH(in_l1, in_r1, in_l1);
+    // row 2
+    t2 = cnst_para1 * in_r3 + cnst_para2 * in_r0
+         - cnst_para0 * in_r2 - cnst_para3 * in_r1;
+    t10 = cnst_para1 * in_l3 + cnst_para2 * in_l0
+          - cnst_para0 * in_l2 - cnst_para3 * in_l1;
+    in_l2 = LD_SH(src + 6 * stride);
+    UNPCK_UB_SH(in_l2, in_r2, in_l2);
+    // row 3
+    t3 = cnst_para1 * in_r0 + cnst_para2 * in_r1
+         - cnst_para0 * in_r3 - cnst_para3 * in_r2;
+    t11 = cnst_para1 * in_l0 + cnst_para2 * in_l1
+          - cnst_para0 * in_l3 - cnst_para3 * in_l2;
+    in_l3 = LD_SH(src + 7 * stride);
+    UNPCK_UB_SH(in_l3, in_r3, in_l3);
+    // row 4
+    t4 = cnst_para1 * in_r1 + cnst_para2 * in_r2
+         - cnst_para0 * in_r0 - cnst_para3 * in_r3;
+    t12 = cnst_para1 * in_l1 + cnst_para2 * in_l2
+          - cnst_para0 * in_l0 - cnst_para3 * in_l3;
+    in_l0 = LD_SH(src + 8 * stride);
+    UNPCK_UB_SH(in_l0, in_r0, in_l0);
+    // row 5
+    t5 = cnst_para1 * in_r2 + cnst_para2 * in_r3
+         - cnst_para0 * in_r1 - cnst_para3 * in_r0;
+    t13 = cnst_para1 * in_l2 + cnst_para2 * in_l3
+          - cnst_para0 * in_l1 - cnst_para3 * in_l0;
+    in_l1 = LD_SH(src + 9 * stride);
+    UNPCK_UB_SH(in_l1, in_r1, in_l1);
+    // row 6
+    t6 = cnst_para1 * in_r3 + cnst_para2 * in_r0
+         - cnst_para0 * in_r2 - cnst_para3 * in_r1;
+    t14 = cnst_para1 * in_l3 + cnst_para2 * in_l0
+          - cnst_para0 * in_l2 - cnst_para3 * in_l1;
+    in_l2 = LD_SH(src + 10 * stride);
+    UNPCK_UB_SH(in_l2, in_r2, in_l2);
+    // row 7
+    t7 = cnst_para1 * in_r0 + cnst_para2 * in_r1
+         - cnst_para0 * in_r3 - cnst_para3 * in_r2;
+    t15 = cnst_para1 * in_l0 + cnst_para2 * in_l1
+          - cnst_para0 * in_l3 - cnst_para3 * in_l2;
+
+    ADD4(t0, cnst_r, t1, cnst_r, t2, cnst_r, t3, cnst_r, t0, t1, t2, t3);
+    ADD4(t4, cnst_r, t5, cnst_r, t6, cnst_r, t7, cnst_r, t4, t5, t6, t7);
+    ADD4(t8, cnst_r, t9, cnst_r, t10, cnst_r, t11, cnst_r,
+         t8, t9, t10, t11);
+    ADD4(t12, cnst_r, t13, cnst_r, t14, cnst_r, t15, cnst_r,
+         t12, t13, t14, t15);
+    t0 >>= shift, t1 >>= shift, t2 >>= shift, t3 >>= shift;
+    t4 >>= shift, t5 >>= shift, t6 >>= shift, t7 >>= shift;
+    t8 >>= shift, t9 >>= shift, t10 >>= shift, t11 >>= shift;
+    t12 >>= shift, t13 >>= shift, t14 >>= shift, t15 >>= shift;
+    TRANSPOSE8x8_SH_SH(t0, t1, t2, t3, t4, t5, t6, t7,
+                       t0, t1, t2, t3, t4, t5, t6, t7);
+    TRANSPOSE8x8_SH_SH(t8, t9, t10, t11, t12, t13, t14, t15,
+                       t8, t9, t10, t11, t12, t13, t14, t15);
+    if (hmode == 1) {
+      cnst_para0 = __msa_ldi_h(4);
+      cnst_para1 = __msa_ldi_h(53);
+      cnst_para2 = __msa_ldi_h(18);
+      cnst_para3 = __msa_ldi_h(3);
+    } else if (hmode == 2) {
+      cnst_para0 = __msa_ldi_h(1);
+      cnst_para1 = __msa_ldi_h(9);
+      cnst_para2 = __msa_ldi_h(9);
+      cnst_para3 = __msa_ldi_h(1);
+    } else {
+      cnst_para0 = __msa_ldi_h(3);
+      cnst_para1 = __msa_ldi_h(18);
+      cnst_para2 = __msa_ldi_h(53);
+      cnst_para3 = __msa_ldi_h(4);
+    }
+    r = 64 - rnd;
+    cnst_r = __msa_fill_h(r);
+    // col 0 ~ 7
+    t0 = cnst_para1 * t1 + cnst_para2 * t2 - cnst_para0 * t0 - cnst_para3 * t3;
+    t1 = cnst_para1 * t2 + cnst_para2 * t3 - cnst_para0 * t1 - cnst_para3 * t4;
+    t2 = cnst_para1 * t3 + cnst_para2 * t4 - cnst_para0 * t2 - cnst_para3 * t5;
+    t3 = cnst_para1 * t4 + cnst_para2 * t5 - cnst_para0 * t3 - cnst_para3 * t6;
+    t4 = cnst_para1 * t5 + cnst_para2 * t6 - cnst_para0 * t4 - cnst_para3 * t7;
+    t5 = cnst_para1 * t6 + cnst_para2 * t7 - cnst_para0 * t5 - cnst_para3 * t8;
+    t6 = cnst_para1 * t7 + cnst_para2 * t8 - cnst_para0 * t6 - cnst_para3 * t9;
+    t7 = cnst_para1 * t8 + cnst_para2 * t9 - cnst_para0 * t7 - cnst_para3 * t10;
+    ADD4(t0, cnst_r, t1, cnst_r, t2, cnst_r, t3, cnst_r, t0, t1, t2, t3);
+    ADD4(t4, cnst_r, t5, cnst_r, t6, cnst_r, t7, cnst_r, t4, t5, t6, t7);
+    t0 >>= 7, t1 >>= 7, t2 >>= 7, t3 >>= 7;
+    t4 >>= 7, t5 >>= 7, t6 >>= 7, t7 >>= 7;
+    TRANSPOSE8x8_SH_SH(t0, t1, t2, t3, t4, t5, t6, t7,
+                       t0, t1, t2, t3, t4, t5, t6, t7);
+    CLIP_SH8_0_255(t0, t1, t2, t3, t4, t5, t6, t7);
+    PCKEV_B4_SH(t1, t0, t3, t2, t5, t4, t7, t6, t0, t1, t2, t3);
+    ST_D8(t0, t1, t2, t3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
+}
+
+#define PUT_VC1_MSPEL_MC_MSA(hmode, vmode)                                    \
+void ff_put_vc1_mspel_mc ## hmode ## vmode ## _msa(uint8_t *dst,              \
+                                                const uint8_t *src,           \
+                                                ptrdiff_t stride, int rnd)    \
+{                                                                             \
+    put_vc1_mspel_mc_h_v_msa(dst, src, stride, hmode, vmode, rnd);            \
+}                                                                             \
+void ff_put_vc1_mspel_mc ## hmode ## vmode ## _16_msa(uint8_t *dst,           \
+                                                   const uint8_t *src,        \
+                                                   ptrdiff_t stride, int rnd) \
+{                                                                             \
+    put_vc1_mspel_mc_h_v_msa(dst, src, stride, hmode, vmode, rnd);            \
+    put_vc1_mspel_mc_h_v_msa(dst + 8, src + 8, stride, hmode, vmode, rnd);    \
+    dst += 8 * stride, src += 8 * stride;                                     \
+    put_vc1_mspel_mc_h_v_msa(dst, src, stride, hmode, vmode, rnd);            \
+    put_vc1_mspel_mc_h_v_msa(dst + 8, src + 8, stride, hmode, vmode, rnd);    \
+}
+
+PUT_VC1_MSPEL_MC_MSA(1, 1);
+PUT_VC1_MSPEL_MC_MSA(1, 2);
+PUT_VC1_MSPEL_MC_MSA(1, 3);
+
+PUT_VC1_MSPEL_MC_MSA(2, 1);
+PUT_VC1_MSPEL_MC_MSA(2, 2);
+PUT_VC1_MSPEL_MC_MSA(2, 3);
+
+PUT_VC1_MSPEL_MC_MSA(3, 1);
+PUT_VC1_MSPEL_MC_MSA(3, 2);
+PUT_VC1_MSPEL_MC_MSA(3, 3);
diff --git a/libavutil/mips/generic_macros_msa.h b/libavutil/mips/generic_macros_msa.h
index c25509e..267d4e6 100644
--- a/libavutil/mips/generic_macros_msa.h
+++ b/libavutil/mips/generic_macros_msa.h
@@ -299,6 +299,7 @@
 #define LD_SB4(...) LD_V4(v16i8, __VA_ARGS__)
 #define LD_UH4(...) LD_V4(v8u16, __VA_ARGS__)
 #define LD_SH4(...) LD_V4(v8i16, __VA_ARGS__)
+#define LD_SW4(...) LD_V4(v4i32, __VA_ARGS__)
 
 #define LD_V5(RTYPE, psrc, stride, out0, out1, out2, out3, out4)  \
 {                                                                 \
@@ -337,6 +338,7 @@
 #define LD_SB8(...) LD_V8(v16i8, __VA_ARGS__)
 #define LD_UH8(...) LD_V8(v8u16, __VA_ARGS__)
 #define LD_SH8(...) LD_V8(v8i16, __VA_ARGS__)
+#define LD_SW8(...) LD_V8(v4i32, __VA_ARGS__)
 
 #define LD_V16(RTYPE, psrc, stride,                                   \
                out0, out1, out2, out3, out4, out5, out6, out7,        \
@@ -1382,6 +1384,7 @@
             out4, out5, out6, out7);                              \
 }
 #define ILVR_B8_UH(...) ILVR_B8(v8u16, __VA_ARGS__)
+#define ILVR_B8_SW(...) ILVR_B8(v4i32, __VA_ARGS__)
 
 /* Description : Interleave right half of halfword elements from vectors
    Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
-- 
2.1.0




More information about the ffmpeg-devel mailing list