[FFmpeg-devel] [PATCH] vp9: sse2/ssse3/avx 16bpp loopfilter x86 simd.

Ronald S. Bultje rsbultje at gmail.com
Wed Sep 30 21:22:11 CEST 2015


---
 libavcodec/x86/Makefile                     |   1 +
 libavcodec/x86/constants.c                  |   2 +
 libavcodec/x86/constants.h                  |   1 +
 libavcodec/x86/vp9dsp_init_16bpp_template.c |  90 +++
 libavcodec/x86/vp9lpf_16bpp.asm             | 823 ++++++++++++++++++++++++++++
 libavcodec/x86/vp9mc_16bpp.asm              |   2 +-
 6 files changed, 918 insertions(+), 1 deletion(-)
 create mode 100644 libavcodec/x86/vp9lpf_16bpp.asm

diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index febaccd..01e5f18 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -160,6 +160,7 @@ YASM-OBJS-$(CONFIG_VP6_DECODER)        += x86/vp6dsp.o
 YASM-OBJS-$(CONFIG_VP9_DECODER)        += x86/vp9intrapred.o            \
                                           x86/vp9itxfm.o                \
                                           x86/vp9lpf.o                  \
+                                          x86/vp9lpf_16bpp.o            \
                                           x86/vp9mc.o                   \
                                           x86/vp9mc_16bpp.o
 YASM-OBJS-$(CONFIG_WEBP_DECODER)       += x86/vp8dsp.o
diff --git a/libavcodec/x86/constants.c b/libavcodec/x86/constants.c
index 553dd49..9f3c8b4 100644
--- a/libavcodec/x86/constants.c
+++ b/libavcodec/x86/constants.c
@@ -55,6 +55,8 @@ DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_1024) = { 0x0400040004000400ULL, 0x040
                                                     0x0400040004000400ULL, 0x0400040004000400ULL};
 DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_2048) = { 0x0800080008000800ULL, 0x0800080008000800ULL,
                                                     0x0800080008000800ULL, 0x0800080008000800ULL };
+DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_4095) = { 0x0fff0fff0fff0fffULL, 0x0fff0fff0fff0fffULL,
+                                                    0x0fff0fff0fff0fffULL, 0x0fff0fff0fff0fffULL };
 DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_4096) = { 0x1000100010001000ULL, 0x1000100010001000ULL,
                                                     0x1000100010001000ULL, 0x1000100010001000ULL };
 DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_8192) = { 0x2000200020002000ULL, 0x2000200020002000ULL,
diff --git a/libavcodec/x86/constants.h b/libavcodec/x86/constants.h
index 33dbb65..37a1869 100644
--- a/libavcodec/x86/constants.h
+++ b/libavcodec/x86/constants.h
@@ -47,6 +47,7 @@ extern const ymm_reg  ff_pw_512;
 extern const ymm_reg  ff_pw_1023;
 extern const ymm_reg  ff_pw_1024;
 extern const ymm_reg  ff_pw_2048;
+extern const ymm_reg  ff_pw_4095;
 extern const ymm_reg  ff_pw_4096;
 extern const ymm_reg  ff_pw_8192;
 extern const ymm_reg  ff_pw_m1;
diff --git a/libavcodec/x86/vp9dsp_init_16bpp_template.c b/libavcodec/x86/vp9dsp_init_16bpp_template.c
index f48225c..56cd79e 100644
--- a/libavcodec/x86/vp9dsp_init_16bpp_template.c
+++ b/libavcodec/x86/vp9dsp_init_16bpp_template.c
@@ -65,6 +65,62 @@ filters_8tap_1d_fn2(put, 16, BPC, avx2, 16bpp)
 filters_8tap_1d_fn2(avg, 16, BPC, avx2, 16bpp)
 #endif
 
+#define decl_lpf_func(dir, wd, bpp, opt) \
+void ff_vp9_loop_filter_##dir##_##wd##_##bpp##_##opt(uint8_t *dst, ptrdiff_t stride, \
+                                                     int E, int I, int H)
+
+#define decl_lpf_funcs(dir, wd, bpp) \
+decl_lpf_func(dir, wd, bpp, sse2); \
+decl_lpf_func(dir, wd, bpp, ssse3); \
+decl_lpf_func(dir, wd, bpp, avx)
+
+#define decl_lpf_funcs_wd(dir) \
+decl_lpf_funcs(dir,  4, BPC); \
+decl_lpf_funcs(dir,  8, BPC); \
+decl_lpf_funcs(dir, 16, BPC)
+
+decl_lpf_funcs_wd(h);
+decl_lpf_funcs_wd(v);
+
+#define lpf_16_wrapper(dir, off, bpp, opt) \
+static void loop_filter_##dir##_16_##bpp##_##opt(uint8_t *dst, ptrdiff_t stride, \
+                                                 int E, int I, int H) \
+{ \
+    ff_vp9_loop_filter_##dir##_16_##bpp##_##opt(dst,       stride, E, I, H); \
+    ff_vp9_loop_filter_##dir##_16_##bpp##_##opt(dst + off, stride, E, I, H); \
+}
+
+#define lpf_16_wrappers(bpp, opt) \
+lpf_16_wrapper(h, 8 * stride, bpp, opt); \
+lpf_16_wrapper(v, 16,         bpp, opt)
+
+lpf_16_wrappers(BPC, sse2);
+lpf_16_wrappers(BPC, ssse3);
+lpf_16_wrappers(BPC, avx);
+
+#define lpf_mix2_wrapper(dir, off, wd1, wd2, bpp, opt) \
+static void loop_filter_##dir##_##wd1##wd2##_##bpp##_##opt(uint8_t *dst, ptrdiff_t stride, \
+                                                           int E, int I, int H) \
+{ \
+    ff_vp9_loop_filter_##dir##_##wd1##_##bpp##_##opt(dst,       stride, \
+                                                     E & 0xff, I & 0xff, H & 0xff); \
+    ff_vp9_loop_filter_##dir##_##wd2##_##bpp##_##opt(dst + off, stride, \
+                                                     E >> 8,   I >> 8,   H >> 8); \
+}
+
+#define lpf_mix2_wrappers(wd1, wd2, bpp, opt) \
+lpf_mix2_wrapper(h, 8 * stride, wd1, wd2, bpp, opt); \
+lpf_mix2_wrapper(v, 16,         wd1, wd2, bpp, opt)
+
+#define lpf_mix2_wrappers_set(bpp, opt) \
+lpf_mix2_wrappers(4, 4, bpp, opt); \
+lpf_mix2_wrappers(4, 8, bpp, opt); \
+lpf_mix2_wrappers(8, 4, bpp, opt); \
+lpf_mix2_wrappers(8, 8, bpp, opt); \
+
+lpf_mix2_wrappers_set(BPC, sse2);
+lpf_mix2_wrappers_set(BPC, ssse3);
+lpf_mix2_wrappers_set(BPC, avx);
 #endif /* HAVE_YASM */
 
 av_cold void INIT_FUNC(VP9DSPContext *dsp)
@@ -72,9 +128,43 @@ av_cold void INIT_FUNC(VP9DSPContext *dsp)
 #if HAVE_YASM
     int cpu_flags = av_get_cpu_flags();
 
+#define init_lpf_8_func(idx1, idx2, dir, wd, bpp, opt) \
+    dsp->loop_filter_8[idx1][idx2] = ff_vp9_loop_filter_##dir##_##wd##_##bpp##_##opt
+#define init_lpf_16_func(idx, dir, bpp, opt) \
+    dsp->loop_filter_16[idx] = loop_filter_##dir##_16_##bpp##_##opt
+#define init_lpf_mix2_func(idx1, idx2, idx3, dir, wd1, wd2, bpp, opt) \
+    dsp->loop_filter_mix2[idx1][idx2][idx3] = loop_filter_##dir##_##wd1##wd2##_##bpp##_##opt
+
+#define init_lpf_funcs(bpp, opt) \
+    init_lpf_8_func(0, 0, h,  4, bpp, opt); \
+    init_lpf_8_func(0, 1, v,  4, bpp, opt); \
+    init_lpf_8_func(1, 0, h,  8, bpp, opt); \
+    init_lpf_8_func(1, 1, v,  8, bpp, opt); \
+    init_lpf_8_func(2, 0, h, 16, bpp, opt); \
+    init_lpf_8_func(2, 1, v, 16, bpp, opt); \
+    init_lpf_16_func(0, h, bpp, opt); \
+    init_lpf_16_func(1, v, bpp, opt); \
+    init_lpf_mix2_func(0, 0, 0, h, 4, 4, bpp, opt); \
+    init_lpf_mix2_func(0, 1, 0, h, 4, 8, bpp, opt); \
+    init_lpf_mix2_func(1, 0, 0, h, 8, 4, bpp, opt); \
+    init_lpf_mix2_func(1, 1, 0, h, 8, 8, bpp, opt); \
+    init_lpf_mix2_func(0, 0, 1, v, 4, 4, bpp, opt); \
+    init_lpf_mix2_func(0, 1, 1, v, 4, 8, bpp, opt); \
+    init_lpf_mix2_func(1, 0, 1, v, 8, 4, bpp, opt); \
+    init_lpf_mix2_func(1, 1, 1, v, 8, 8, bpp, opt)
+
     if (EXTERNAL_SSE2(cpu_flags)) {
         init_subpel3(0, put, BPC, sse2);
         init_subpel3(1, avg, BPC, sse2);
+        init_lpf_funcs(BPC, sse2);
+    }
+
+    if (EXTERNAL_SSSE3(cpu_flags)) {
+        init_lpf_funcs(BPC, ssse3);
+    }
+
+    if (EXTERNAL_AVX(cpu_flags)) {
+        init_lpf_funcs(BPC, avx);
     }
 
     if (EXTERNAL_AVX2(cpu_flags)) {
diff --git a/libavcodec/x86/vp9lpf_16bpp.asm b/libavcodec/x86/vp9lpf_16bpp.asm
new file mode 100644
index 0000000..c15437b
--- /dev/null
+++ b/libavcodec/x86/vp9lpf_16bpp.asm
@@ -0,0 +1,823 @@
+;******************************************************************************
+;* VP9 loop filter SIMD optimizations
+;*
+;* Copyright (C) 2015 Ronald S. Bultje <rsbultje at gmail.com>
+;*
+;* 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 "libavutil/x86/x86util.asm"
+
+SECTION_RODATA
+
+pw_511: times 16 dw 511
+pw_2047: times 16 dw 2047
+pw_16384: times 16 dw 16384
+pw_m512: times 16 dw -512
+pw_m2048: times 16 dw -2048
+
+cextern pw_1
+cextern pw_3
+cextern pw_4
+cextern pw_8
+cextern pw_16
+cextern pw_256
+cextern pw_1023
+cextern pw_4095
+cextern pw_m1
+
+SECTION .text
+
+%macro SCRATCH 3-4
+%if ARCH_X86_64
+    SWAP                %1, %2
+%if %0 == 4
+%define reg_%4 m%2
+%endif
+%else
+    mova              [%3], m%1
+%if %0 == 4
+%define reg_%4 [%3]
+%endif
+%endif
+%endmacro
+
+%macro UNSCRATCH 3-4
+%if ARCH_X86_64
+    SWAP                %1, %2
+%else
+    mova               m%1, [%3]
+%endif
+%if %0 == 4
+%undef reg_%4
+%endif
+%endmacro
+
+%macro PRELOAD 2-3
+%if ARCH_X86_64
+    mova               m%1, [%2]
+%if %0 == 3
+%define reg_%3 m%1
+%endif
+%elif %0 == 3
+%define reg_%3 [%2]
+%endif
+%endmacro
+
+; calulate p or q portion of flat8out
+%macro FLAT8OUT_HALF 0
+    psubw               m4, m0                      ; q4-q0
+    psubw               m5, m0                      ; q5-q0
+    psubw               m6, m0                      ; q6-q0
+    psubw               m7, m0                      ; q7-q0
+    ABS2                m4, m5, m2, m3              ; abs(q4-q0) | abs(q5-q0)
+    ABS2                m6, m7, m2, m3              ; abs(q6-q0) | abs(q7-q0)
+    pcmpgtw             m4, reg_F                   ; abs(q4-q0) > F
+    pcmpgtw             m5, reg_F                   ; abs(q5-q0) > F
+    pcmpgtw             m6, reg_F                   ; abs(q6-q0) > F
+    pcmpgtw             m7, reg_F                   ; abs(q7-q0) > F
+    por                 m5, m4
+    por                 m7, m6
+    por                 m7, m5                      ; !flat8out, q portion
+%endmacro
+
+; calculate p or q portion of flat8in/hev/fm (excluding mb_edge condition)
+%macro FLAT8IN_HALF 1
+%if %1 > 4
+    psubw               m4, m3, m0                  ; q3-q0
+    psubw               m5, m2, m0                  ; q2-q0
+    ABS2                m4, m5, m6, m7              ; abs(q3-q0) | abs(q2-q0)
+    pcmpgtw             m4, reg_F                   ; abs(q3-q0) > F
+    pcmpgtw             m5, reg_F                   ; abs(q2-q0) > F
+%endif
+    psubw               m3, m2                      ; q3-q2
+    psubw               m2, m1                      ; q2-q1
+    ABS2                m3, m2, m6, m7              ; abs(q3-q2) | abs(q2-q1)
+    pcmpgtw             m3, reg_I                   ; abs(q3-q2) > I
+    pcmpgtw             m2, reg_I                   ; abs(q2-q1) > I
+%if %1 > 4
+    por                 m4, m5
+%endif
+    por                 m2, m3
+    psubw               m3, m1, m0                  ; q1-q0
+    ABS1                m3, m5                      ; abs(q1-q0)
+%if %1 > 4
+    pcmpgtw             m6, m3, reg_F               ; abs(q1-q0) > F
+%endif
+    pcmpgtw             m7, m3, reg_H               ; abs(q1-q0) > H
+    pcmpgtw             m3, reg_I                   ; abs(q1-q0) > I
+%if %1 > 4
+    por                 m4, m6
+%endif
+    por                 m2, m3
+%endmacro
+
+; one step in filter_14/filter_6
+;
+; take sum $reg, downshift, apply mask and write into dst
+;
+; if sub2/add1-2 are present, add/sub as appropriate to prepare for the next
+; step's sum $reg. This is omitted for the last row in each filter.
+;
+; if dont_store is set, don't write the result into memory, instead keep the
+; values in register so we can write it out later
+%macro FILTER_STEP 6-10 "", "", "", 0 ; tmp, reg, mask, shift, dst, \
+                                      ; src/sub1, sub2, add1, add2, dont_store
+    psrlw               %1, %2, %4
+    psubw               %1, %6                      ; abs->delta
+%ifnidn %7, ""
+    psubw               %2, %6
+    psubw               %2, %7
+    paddw               %2, %8
+    paddw               %2, %9
+%endif
+    pand                %1, reg_%3                  ; apply mask
+%if %10 == 1
+    paddw               %6, %1                      ; delta->abs
+%else
+    paddw               %1, %6                      ; delta->abs
+    mova              [%5], %1
+%endif
+%endmacro
+
+; FIXME avx2 versions for 16_16 and mix2_{4,8}{4,8}
+
+%macro LOOP_FILTER 3 ; dir[h/v], wd[4/8/16], bpp[10/12]
+
+%if ARCH_X86_64
+%if %2 == 16
+%assign %%num_xmm_regs 16
+%elif %2 == 8
+%assign %%num_xmm_regs 15
+%else ; %2 == 4
+%assign %%num_xmm_regs 14
+%endif ; %2
+%assign %%bak_mem 0
+%else ; ARCH_X86_32
+%assign %%num_xmm_regs 8
+%if %2 == 16
+%assign %%bak_mem 7
+%elif %2 == 8
+%assign %%bak_mem 6
+%else ; %2 == 4
+%assign %%bak_mem 5
+%endif ; %2
+%endif ; ARCH_X86_64/32
+
+%if %2 == 16
+%ifidn %1, v
+%assign %%num_gpr_regs 6
+%else ; %1 == h
+%assign %%num_gpr_regs 5
+%endif ; %1
+%assign %%wd_mem 6
+%else ; %2 == 8/4
+%assign %%num_gpr_regs 5
+%if ARCH_X86_32 && %2 == 8
+%assign %%wd_mem 2
+%else ; ARCH_X86_64 || %2 == 4
+%assign %%wd_mem 0
+%endif ; ARCH_X86_64/32 etc.
+%endif ; %2
+
+%ifidn %1, v
+%assign %%tsp_mem 0
+%elif %2 == 16 ; && %1 == h
+%assign %%tsp_mem 16
+%else ; %1 == h && %1 == 8/4
+%assign %%tsp_mem 8
+%endif ; %1/%2
+
+%assign %%off %%wd_mem
+%assign %%tspoff %%bak_mem+%%wd_mem
+%assign %%stack_mem ((%%bak_mem+%%wd_mem+%%tsp_mem)*mmsize)
+
+%if %3 == 10
+%define %%maxsgn 511
+%define %%minsgn m512
+%define %%maxusgn 1023
+%define %%maxf 4
+%else ; %3 == 12
+%define %%maxsgn 2047
+%define %%minsgn m2048
+%define %%maxusgn 4095
+%define %%maxf 16
+%endif ; %3
+
+cglobal vp9_loop_filter_%1_%2_%3, 5, %%num_gpr_regs, %%num_xmm_regs, %%stack_mem, dst, stride, E, I, H
+    ; prepare E, I and H masks
+    shl                 Ed, %3-8
+    shl                 Id, %3-8
+    shl                 Hd, %3-8
+%if cpuflag(ssse3)
+    mova                m0, [pw_256]
+%endif
+    movd                m1, Ed
+    movd                m2, Id
+    movd                m3, Hd
+%if cpuflag(ssse3)
+    pshufb              m1, m0                      ; E << (bit_depth - 8)
+    pshufb              m2, m0                      ; I << (bit_depth - 8)
+    pshufb              m3, m0                      ; H << (bit_depth - 8)
+%else
+    punpcklwd           m1, m1
+    punpcklwd           m2, m2
+    punpcklwd           m3, m3
+    pshufd              m1, m1, q0000
+    pshufd              m2, m2, q0000
+    pshufd              m3, m3, q0000
+%endif
+    SCRATCH              1,  8, rsp+(%%off+0)*mmsize,  E
+    SCRATCH              2,  9, rsp+(%%off+1)*mmsize,  I
+    SCRATCH              3, 10, rsp+(%%off+2)*mmsize,  H
+%if %2 > 4
+    PRELOAD                 11, pw_ %+ %%maxf, F
+%endif
+
+    ; set up variables to load data
+%ifidn %1, v
+    DEFINE_ARGS dst8, stride, stride3, dst0, dst4, dst12
+    lea           stride3q, [strideq*3]
+    neg            strideq
+%if %2 == 16
+    lea              dst0q, [dst8q+strideq*8]
+%else
+    lea              dst4q, [dst8q+strideq*4]
+%endif
+    neg            strideq
+%if %2 == 16
+    lea             dst12q, [dst8q+strideq*4]
+    lea              dst4q, [dst0q+strideq*4]
+%endif
+
+%if %2 == 16
+%define %%p7 dst0q
+%define %%p6 dst0q+strideq
+%define %%p5 dst0q+strideq*2
+%define %%p4 dst0q+stride3q
+%endif
+%define %%p3 dst4q
+%define %%p2 dst4q+strideq
+%define %%p1 dst4q+strideq*2
+%define %%p0 dst4q+stride3q
+%define %%q0 dst8q
+%define %%q1 dst8q+strideq
+%define %%q2 dst8q+strideq*2
+%define %%q3 dst8q+stride3q
+%if %2 == 16
+%define %%q4 dst12q
+%define %%q5 dst12q+strideq
+%define %%q6 dst12q+strideq*2
+%define %%q7 dst12q+stride3q
+%endif
+%else ; %1 == h
+    DEFINE_ARGS dst0, stride, stride3, dst4
+    lea           stride3q, [strideq*3]
+    lea              dst4q, [dst0q+strideq*4]
+
+%define %%p3 rsp+(%%tspoff+0)*mmsize
+%define %%p2 rsp+(%%tspoff+1)*mmsize
+%define %%p1 rsp+(%%tspoff+2)*mmsize
+%define %%p0 rsp+(%%tspoff+3)*mmsize
+%define %%q0 rsp+(%%tspoff+4)*mmsize
+%define %%q1 rsp+(%%tspoff+5)*mmsize
+%define %%q2 rsp+(%%tspoff+6)*mmsize
+%define %%q3 rsp+(%%tspoff+7)*mmsize
+
+%if %2 < 16
+    movu                m0, [dst0q+strideq*0-8]
+    movu                m1, [dst0q+strideq*1-8]
+    movu                m2, [dst0q+strideq*2-8]
+    movu                m3, [dst0q+stride3q -8]
+    movu                m4, [dst4q+strideq*0-8]
+    movu                m5, [dst4q+strideq*1-8]
+    movu                m6, [dst4q+strideq*2-8]
+    movu                m7, [dst4q+stride3q -8]
+
+%if ARCH_X86_64
+    TRANSPOSE8x8W        0, 1, 2, 3, 4, 5, 6, 7, 12
+%else
+    TRANSPOSE8x8W        0, 1, 2, 3, 4, 5, 6, 7, [%%p0], [%%q0]
+%endif
+
+    mova            [%%p3], m0
+    mova            [%%p2], m1
+    mova            [%%p1], m2
+    mova            [%%p0], m3
+%if ARCH_X86_64
+    mova            [%%q0], m4
+%endif
+    mova            [%%q1], m5
+    mova            [%%q2], m6
+    mova            [%%q3], m7
+
+    ; FIXME investigate if we can _not_ load q0-3 below if h, and adjust register
+    ; order here accordingly
+%else ; %2 == 16
+
+%define %%p7 rsp+(%%tspoff+ 8)*mmsize
+%define %%p6 rsp+(%%tspoff+ 9)*mmsize
+%define %%p5 rsp+(%%tspoff+10)*mmsize
+%define %%p4 rsp+(%%tspoff+11)*mmsize
+%define %%q4 rsp+(%%tspoff+12)*mmsize
+%define %%q5 rsp+(%%tspoff+13)*mmsize
+%define %%q6 rsp+(%%tspoff+14)*mmsize
+%define %%q7 rsp+(%%tspoff+15)*mmsize
+
+    mova                m0, [dst0q+strideq*0-16]
+    mova                m1, [dst0q+strideq*1-16]
+    mova                m2, [dst0q+strideq*2-16]
+    mova                m3, [dst0q+stride3q -16]
+    mova                m4, [dst4q+strideq*0-16]
+    mova                m5, [dst4q+strideq*1-16]
+%if ARCH_X86_64
+    mova                m6, [dst4q+strideq*2-16]
+%endif
+    mova                m7, [dst4q+stride3q -16]
+
+%if ARCH_X86_64
+    TRANSPOSE8x8W        0, 1, 2, 3, 4, 5, 6, 7, 12
+%else
+    TRANSPOSE8x8W        0, 1, 2, 3, 4, 5, 6, 7, [dst4q+strideq*2-16], [%%p3], 1
+%endif
+
+    mova            [%%p7], m0
+    mova            [%%p6], m1
+    mova            [%%p5], m2
+    mova            [%%p4], m3
+%if ARCH_X86_64
+    mova            [%%p3], m4
+%endif
+    mova            [%%p2], m5
+    mova            [%%p1], m6
+    mova            [%%p0], m7
+
+    mova                m0, [dst0q+strideq*0]
+    mova                m1, [dst0q+strideq*1]
+    mova                m2, [dst0q+strideq*2]
+    mova                m3, [dst0q+stride3q ]
+    mova                m4, [dst4q+strideq*0]
+    mova                m5, [dst4q+strideq*1]
+%if ARCH_X86_64
+    mova                m6, [dst4q+strideq*2]
+%endif
+    mova                m7, [dst4q+stride3q ]
+
+%if ARCH_X86_64
+    TRANSPOSE8x8W        0, 1, 2, 3, 4, 5, 6, 7, 12
+%else
+    TRANSPOSE8x8W        0, 1, 2, 3, 4, 5, 6, 7, [dst4q+strideq*2], [%%q4], 1
+%endif
+
+    mova            [%%q0], m0
+    mova            [%%q1], m1
+    mova            [%%q2], m2
+    mova            [%%q3], m3
+%if ARCH_X86_64
+    mova            [%%q4], m4
+%endif
+    mova            [%%q5], m5
+    mova            [%%q6], m6
+    mova            [%%q7], m7
+
+    ; FIXME investigate if we can _not_ load q0|q4-7 below if h, and adjust register
+    ; order here accordingly
+%endif ; %2
+%endif ; %1
+
+    ; load q0|q4-7 data
+    mova                m0, [%%q0]
+%if %2 == 16
+    mova                m4, [%%q4]
+    mova                m5, [%%q5]
+    mova                m6, [%%q6]
+    mova                m7, [%%q7]
+
+    ; flat8out q portion
+    FLAT8OUT_HALF
+    SCRATCH              7, 15, rsp+(%%off+6)*mmsize, F8O
+%endif
+
+    ; load q1-3 data
+    mova                m1, [%%q1]
+    mova                m2, [%%q2]
+    mova                m3, [%%q3]
+
+    ; r6-8|pw_4[m8-11]=reg_E/I/H/F
+    ; r9[m15]=!flatout[q]
+    ; m12-14=free
+    ; m0-3=q0-q3
+    ; m4-7=free
+
+    ; flat8in|fm|hev q portion
+    FLAT8IN_HALF        %2
+    SCRATCH              7, 13, rsp+(%%off+4)*mmsize, HEV
+%if %2 > 4
+    SCRATCH              4, 14, rsp+(%%off+5)*mmsize, F8I
+%endif
+
+    ; r6-8|pw_4[m8-11]=reg_E/I/H/F
+    ; r9[m15]=!flat8out[q]
+    ; r10[m13]=hev[q]
+    ; r11[m14]=!flat8in[q]
+    ; m2=!fm[q]
+    ; m0,1=q0-q1
+    ; m2-7=free
+    ; m12=free
+
+    ; load p0-1
+    mova                m3, [%%p0]
+    mova                m4, [%%p1]
+
+    ; fm mb_edge portion
+    psubw               m5, m3, m0                  ; q0-p0
+    psubw               m6, m4, m1                  ; q1-p1
+%if ARCH_X86_64
+    ABS2                m5, m6, m7, m12             ; abs(q0-p0) | abs(q1-p1)
+%else
+    ABS1                m5, m7                      ; abs(q0-p0)
+    ABS1                m6, m7                      ; abs(q1-p1)
+%endif
+    paddw               m5, m5
+    psraw               m6, 1
+    paddw               m6, m5                      ; abs(q0-p0)*2+(abs(q1-p1)>>1)
+    pcmpgtw             m6, reg_E
+    por                 m2, m6
+    SCRATCH              2, 12, rsp+(%%off+3)*mmsize, FM
+
+    ; r6-8|pw_4[m8-11]=reg_E/I/H/F
+    ; r9[m15]=!flat8out[q]
+    ; r10[m13]=hev[q]
+    ; r11[m14]=!flat8in[q]
+    ; r12[m12]=!fm[q]
+    ; m3-4=q0-1
+    ; m0-2/5-7=free
+
+    ; load p4-7 data
+    SWAP                 3, 0                       ; p0
+    SWAP                 4, 1                       ; p1
+%if %2 == 16
+    mova                m7, [%%p7]
+    mova                m6, [%%p6]
+    mova                m5, [%%p5]
+    mova                m4, [%%p4]
+
+    ; flat8out p portion
+    FLAT8OUT_HALF
+    por                 m7, reg_F8O
+    SCRATCH              7, 15, rsp+(%%off+6)*mmsize, F8O
+%endif
+
+    ; r6-8|pw_4[m8-11]=reg_E/I/H/F
+    ; r9[m15]=!flat8out
+    ; r10[m13]=hev[q]
+    ; r11[m14]=!flat8in[q]
+    ; r12[m12]=!fm[q]
+    ; m0=p0
+    ; m1-7=free
+
+    ; load p2-3 data
+    mova                m2, [%%p2]
+    mova                m3, [%%p3]
+
+    ; flat8in|fm|hev p portion
+    FLAT8IN_HALF        %2
+    por                 m7, reg_HEV
+%if %2 > 4
+    por                 m4, reg_F8I
+%endif
+    por                 m2, reg_FM
+%if %2 > 4
+    por                 m4, m2                      ; !flat8|!fm
+%if %2 == 16
+    por                 m5, m4, reg_F8O             ; !flat16|!fm
+    pandn               m2, m4                      ; filter4_mask
+    pandn               m4, m5                      ; filter8_mask
+    pxor                m5, [pw_m1]                 ; filter16_mask
+    SCRATCH              5, 15, rsp+(%%off+6)*mmsize, F16M
+%else
+    pandn               m2, m4                      ; filter4_mask
+    pxor                m4, [pw_m1]                 ; filter8_mask
+%endif
+    SCRATCH              4, 14, rsp+(%%off+5)*mmsize, F8M
+%else
+    pxor                m2, [pw_m1]                 ; filter4_mask
+%endif
+    SCRATCH              7, 13, rsp+(%%off+4)*mmsize, HEV
+    SCRATCH              2, 12, rsp+(%%off+3)*mmsize, F4M
+
+    ; r9[m15]=filter16_mask
+    ; r10[m13]=hev
+    ; r11[m14]=filter8_mask
+    ; r12[m12]=filter4_mask
+    ; m0,1=p0-p1
+    ; m2-7=free
+    ; m8-11=free
+
+%if %2 > 4
+%if %2 == 16
+    ; filter_14
+    mova                m2, [%%p7]
+    mova                m3, [%%p6]
+    mova                m6, [%%p5]
+    mova                m7, [%%p4]
+    PRELOAD              8, %%p3, P3
+    PRELOAD              9, %%p2, P2
+%endif
+    PRELOAD             10, %%q0, Q0
+    PRELOAD             11, %%q1, Q1
+%if %2 == 16
+    psllw               m4, m2, 3
+    paddw               m5, m3, m3
+    paddw               m4, m6
+    paddw               m5, m7
+    paddw               m4, reg_P3
+    paddw               m5, reg_P2
+    paddw               m4, m1
+    paddw               m5, m0
+    paddw               m4, reg_Q0                  ; q0+p1+p3+p5+p7*8
+    psubw               m5, m2                      ; p0+p2+p4+p6*2-p7
+    paddw               m4, [pw_8]
+    paddw               m5, m4                      ; q0+p0+p1+p2+p3+p4+p5+p6*2+p7*7+8
+
+    ; below, we use r0-5 for storing pre-filter pixels for subsequent subtraction
+    ; at the end of the filter
+
+    mova    [rsp+0*mmsize], m3
+    FILTER_STEP         m4, m5, F16M, 4, %%p6, m3,     m2,             m6,     reg_Q1
+%endif
+    mova                m3, [%%q2]
+%if %2 == 16
+    mova    [rsp+1*mmsize], m6
+    FILTER_STEP         m4, m5, F16M, 4, %%p5, m6,     m2,             m7,     m3
+%endif
+    mova                m6, [%%q3]
+%if %2 == 16
+    mova    [rsp+2*mmsize], m7
+    FILTER_STEP         m4, m5, F16M, 4, %%p4, m7,     m2,             reg_P3, m6
+    mova                m7, [%%q4]
+%if ARCH_X86_64
+    mova    [rsp+3*mmsize], reg_P3
+%else
+    mova                m4, reg_P3
+    mova    [rsp+3*mmsize], m4
+%endif
+    FILTER_STEP         m4, m5, F16M, 4, %%p3, reg_P3, m2,             reg_P2, m7
+    PRELOAD              8, %%q5, Q5
+%if ARCH_X86_64
+    mova    [rsp+4*mmsize], reg_P2
+%else
+    mova                m4, reg_P2
+    mova    [rsp+4*mmsize], m4
+%endif
+    FILTER_STEP         m4, m5, F16M, 4, %%p2, reg_P2, m2,             m1,     reg_Q5
+    PRELOAD              9, %%q6, Q6
+    mova    [rsp+5*mmsize], m1
+    FILTER_STEP         m4, m5, F16M, 4, %%p1, m1,     m2,             m0,     reg_Q6
+    mova                m1, [%%q7]
+    FILTER_STEP         m4, m5, F16M, 4, %%p0, m0,     m2,             reg_Q0, m1,     1
+    FILTER_STEP         m4, m5, F16M, 4, %%q0, reg_Q0, [rsp+0*mmsize], reg_Q1, m1,     ARCH_X86_64
+    FILTER_STEP         m4, m5, F16M, 4, %%q1, reg_Q1, [rsp+1*mmsize], m3,     m1,     ARCH_X86_64
+    FILTER_STEP         m4, m5, F16M, 4, %%q2, m3,     [rsp+2*mmsize], m6,     m1,     1
+    FILTER_STEP         m4, m5, F16M, 4, %%q3, m6,     [rsp+3*mmsize], m7,     m1
+    FILTER_STEP         m4, m5, F16M, 4, %%q4, m7,     [rsp+4*mmsize], reg_Q5, m1
+    FILTER_STEP         m4, m5, F16M, 4, %%q5, reg_Q5, [rsp+5*mmsize], reg_Q6, m1
+    FILTER_STEP         m4, m5, F16M, 4, %%q6, reg_Q6
+
+    mova                m7, [%%p1]
+%else
+    SWAP                 1, 7
+%endif
+
+    mova                m2, [%%p3]
+    mova                m1, [%%p2]
+
+    ; reg_Q0-1 (m10-m11)
+    ; m0=p0
+    ; m1=p2
+    ; m2=p3
+    ; m3=q2
+    ; m4-5=free
+    ; m6=q3
+    ; m7=p1
+    ; m8-9 unused
+
+    ; filter_6
+    psllw               m4, m2, 2
+    paddw               m5, m1, m1
+    paddw               m4, m7
+    psubw               m5, m2
+    paddw               m4, m0
+    paddw               m5, reg_Q0
+    paddw               m4, [pw_4]
+    paddw               m5, m4
+
+%if ARCH_X86_64
+    mova                m8, m1
+    mova                m9, m7
+%else
+    mova    [rsp+0*mmsize], m1
+    mova    [rsp+1*mmsize], m7
+%endif
+%ifidn %1, v
+    FILTER_STEP         m4, m5, F8M, 3, %%p2, m1,     m2,             m7,     reg_Q1
+%else
+    FILTER_STEP         m4, m5, F8M, 3, %%p2, m1,     m2,             m7,     reg_Q1, 1
+%endif
+    FILTER_STEP         m4, m5, F8M, 3, %%p1, m7,     m2,             m0,     m3, 1
+    FILTER_STEP         m4, m5, F8M, 3, %%p0, m0,     m2,             reg_Q0, m6, 1
+%if ARCH_X86_64
+    FILTER_STEP         m4, m5, F8M, 3, %%q0, reg_Q0, m8,             reg_Q1, m6, ARCH_X86_64
+    FILTER_STEP         m4, m5, F8M, 3, %%q1, reg_Q1, m9,             m3,     m6, ARCH_X86_64
+%else
+    FILTER_STEP         m4, m5, F8M, 3, %%q0, reg_Q0, [rsp+0*mmsize], reg_Q1, m6, ARCH_X86_64
+    FILTER_STEP         m4, m5, F8M, 3, %%q1, reg_Q1, [rsp+1*mmsize], m3,     m6, ARCH_X86_64
+%endif
+    FILTER_STEP         m4, m5, F8M, 3, %%q2, m3
+
+    UNSCRATCH            2, 10, %%q0
+    UNSCRATCH            6, 11, %%q1
+%else
+    SWAP                 1, 7
+    mova                m2, [%%q0]
+    mova                m6, [%%q1]
+%endif
+    UNSCRATCH            3, 13, rsp+(%%off+4)*mmsize, HEV
+
+    ; m0=p0
+    ; m1=p2
+    ; m2=q0
+    ; m3=hev_mask
+    ; m4-5=free
+    ; m6=q1
+    ; m7=p1
+
+    ; filter_4
+    psubw               m4, m7, m6              ; p1-q1
+    psubw               m5, m2, m0              ; q0-p0
+    pand                m4, m3
+    pminsw              m4, [pw_ %+ %%maxsgn]
+    pmaxsw              m4, [pw_ %+ %%minsgn]   ; clip_intp2(p1-q1, 9) -> f
+    paddw               m4, m5
+    paddw               m5, m5
+    paddw               m4, m5                  ; 3*(q0-p0)+f
+    pminsw              m4, [pw_ %+ %%maxsgn]
+    pmaxsw              m4, [pw_ %+ %%minsgn]   ; clip_intp2(3*(q0-p0)+f, 9) -> f
+    pand                m4, reg_F4M
+    paddw               m5, m4, [pw_4]
+    paddw               m4, [pw_3]
+    pminsw              m5, [pw_ %+ %%maxsgn]
+    pminsw              m4, [pw_ %+ %%maxsgn]
+    psraw               m5, 3                   ; min_intp2(f+4, 9)>>3 -> f1
+    psraw               m4, 3                   ; min_intp2(f+3, 9)>>3 -> f2
+    psubw               m2, m5                  ; q0-f1
+    paddw               m0, m4                  ; p0+f2
+    pandn               m3, m5                  ; f1 & !hev (for p1/q1 adj)
+    pxor                m4, m4
+    mova                m5, [pw_ %+ %%maxusgn]
+    pmaxsw              m2, m4
+    pmaxsw              m0, m4
+    pminsw              m2, m5
+    pminsw              m0, m5
+%if cpuflag(ssse3)
+    pmulhrsw            m3, [pw_16384]          ; (f1+1)>>1
+%else
+    paddw               m3, [pw_1]
+    psraw               m3, 1
+%endif
+    paddw               m7, m3                  ; p1+f
+    psubw               m6, m3                  ; q1-f
+    pmaxsw              m7, m4
+    pmaxsw              m6, m4
+    pminsw              m7, m5
+    pminsw              m6, m5
+
+    ; store
+%ifidn %1, v
+    mova            [%%p1], m7
+    mova            [%%p0], m0
+    mova            [%%q0], m2
+    mova            [%%q1], m6
+%else ; %1 == h
+%if %2 == 4
+    TRANSPOSE4x4W        7, 0, 2, 6, 1
+    movh   [dst0q+strideq*0-4], m7
+    movhps [dst0q+strideq*1-4], m7
+    movh   [dst0q+strideq*2-4], m0
+    movhps [dst0q+stride3q -4], m0
+    movh   [dst4q+strideq*0-4], m2
+    movhps [dst4q+strideq*1-4], m2
+    movh   [dst4q+strideq*2-4], m6
+    movhps [dst4q+stride3q -4], m6
+%elif %2 == 8
+    mova                m3, [%%p3]
+    mova                m4, [%%q2]
+    mova                m5, [%%q3]
+
+%if ARCH_X86_64
+    TRANSPOSE8x8W        3, 1, 7, 0, 2, 6, 4, 5, 8
+%else
+    TRANSPOSE8x8W        3, 1, 7, 0, 2, 6, 4, 5, [%%q2], [%%q0], 1
+    mova                m2, [%%q0]
+%endif
+
+    movu [dst0q+strideq*0-8], m3
+    movu [dst0q+strideq*1-8], m1
+    movu [dst0q+strideq*2-8], m7
+    movu [dst0q+stride3q -8], m0
+    movu [dst4q+strideq*0-8], m2
+    movu [dst4q+strideq*1-8], m6
+    movu [dst4q+strideq*2-8], m4
+    movu [dst4q+stride3q -8], m5
+%else ; %2 == 16
+    SCRATCH              2, 8, %%q0
+    SCRATCH              6, 9, %%q1
+    mova                m2, [%%p7]
+    mova                m3, [%%p6]
+    mova                m4, [%%p5]
+    mova                m5, [%%p4]
+    mova                m6, [%%p3]
+
+%if ARCH_X86_64
+    TRANSPOSE8x8W        2, 3, 4, 5, 6, 1, 7, 0, 10
+%else
+    mova            [%%p1], m7
+    TRANSPOSE8x8W        2, 3, 4, 5, 6, 1, 7, 0, [%%p1], [dst4q+strideq*0-16], 1
+%endif
+
+    mova [dst0q+strideq*0-16], m2
+    mova [dst0q+strideq*1-16], m3
+    mova [dst0q+strideq*2-16], m4
+    mova [dst0q+stride3q -16], m5
+%if ARCH_X86_64
+    mova [dst4q+strideq*0-16], m6
+%endif
+    mova [dst4q+strideq*1-16], m1
+    mova [dst4q+strideq*2-16], m7
+    mova [dst4q+stride3q -16], m0
+
+    UNSCRATCH            2, 8, %%q0
+    UNSCRATCH            6, 9, %%q1
+    mova                m0, [%%q2]
+    mova                m1, [%%q3]
+    mova                m3, [%%q4]
+    mova                m4, [%%q5]
+%if ARCH_X86_64
+    mova                m5, [%%q6]
+%endif
+    mova                m7, [%%q7]
+
+%if ARCH_X86_64
+    TRANSPOSE8x8W        2, 6, 0, 1, 3, 4, 5, 7, 8
+%else
+    TRANSPOSE8x8W        2, 6, 0, 1, 3, 4, 5, 7, [%%q6], [dst4q+strideq*0], 1
+%endif
+
+    mova [dst0q+strideq*0], m2
+    mova [dst0q+strideq*1], m6
+    mova [dst0q+strideq*2], m0
+    mova [dst0q+stride3q ], m1
+%if ARCH_X86_64
+    mova [dst4q+strideq*0], m3
+%endif
+    mova [dst4q+strideq*1], m4
+    mova [dst4q+strideq*2], m5
+    mova [dst4q+stride3q ], m7
+%endif ; %2
+%endif ; %1
+    RET
+%endmacro
+
+%macro LOOP_FILTER_CPUSETS 3
+INIT_XMM sse2
+LOOP_FILTER %1, %2, %3
+INIT_XMM ssse3
+LOOP_FILTER %1, %2, %3
+INIT_XMM avx
+LOOP_FILTER %1, %2, %3
+%endmacro
+
+%macro LOOP_FILTER_WDSETS 2
+LOOP_FILTER_CPUSETS %1,  4, %2
+LOOP_FILTER_CPUSETS %1,  8, %2
+LOOP_FILTER_CPUSETS %1, 16, %2
+%endmacro
+
+LOOP_FILTER_WDSETS h, 10
+LOOP_FILTER_WDSETS v, 10
+LOOP_FILTER_WDSETS h, 12
+LOOP_FILTER_WDSETS v, 12
diff --git a/libavcodec/x86/vp9mc_16bpp.asm b/libavcodec/x86/vp9mc_16bpp.asm
index f60dab7..9a462ea 100644
--- a/libavcodec/x86/vp9mc_16bpp.asm
+++ b/libavcodec/x86/vp9mc_16bpp.asm
@@ -24,10 +24,10 @@
 
 SECTION_RODATA 32
 
-pw_4095: times 16 dw 0xfff
 pd_64: times 8 dd 64
 
 cextern pw_1023
+cextern pw_4095
 
 SECTION .text
 
-- 
2.1.2



More information about the ffmpeg-devel mailing list