1;******************************************************************************
2;* linear least squares model
3;*
4;* Copyright (c) 2013 Loren Merritt
5;*
6;* This file is part of FFmpeg.
7;*
8;* FFmpeg is free software; you can redistribute it and/or
9;* modify it under the terms of the GNU Lesser General Public
10;* License as published by the Free Software Foundation; either
11;* version 2.1 of the License, or (at your option) any later version.
12;*
13;* FFmpeg is distributed in the hope that it will be useful,
14;* but WITHOUT ANY WARRANTY; without even the implied warranty of
15;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16;* Lesser General Public License for more details.
17;*
18;* You should have received a copy of the GNU Lesser General Public
19;* License along with FFmpeg; if not, write to the Free Software
20;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21;******************************************************************************
22
23%include "x86util.asm"
24
25SECTION .text
26
27%define MAX_VARS 32
28%define MAX_VARS_ALIGN (MAX_VARS+4)
29%define COVAR_STRIDE MAX_VARS_ALIGN*8
30%define COVAR(x,y) [covarq + (x)*8 + (y)*COVAR_STRIDE]
31
32struc LLSModel
33    .covariance:  resq MAX_VARS_ALIGN*MAX_VARS_ALIGN
34    .coeff:       resq MAX_VARS*MAX_VARS
35    .variance:    resq MAX_VARS
36    .indep_count: resd 1
37endstruc
38
39%macro ADDPD_MEM 2
40%if cpuflag(avx)
41    vaddpd %2, %2, %1
42%else
43    addpd  %2, %1
44%endif
45    mova   %1, %2
46%endmacro
47
48INIT_XMM sse2
49%define movdqa movaps
50cglobal update_lls, 2,5,8, ctx, var, i, j, covar2
51    %define covarq ctxq
52    mov     id, [ctxq + LLSModel.indep_count]
53    lea   varq, [varq + iq*8]
54    neg     iq
55    mov covar2q, covarq
56.loopi:
57    ; Compute all 3 pairwise products of a 2x2 block that lies on the diagonal
58    mova    m1, [varq + iq*8]
59    mova    m3, [varq + iq*8 + 16]
60    pshufd  m4, m1, q1010
61    pshufd  m5, m1, q3232
62    pshufd  m6, m3, q1010
63    pshufd  m7, m3, q3232
64    mulpd   m0, m1, m4
65    mulpd   m1, m1, m5
66    lea covarq, [covar2q + 16]
67    ADDPD_MEM COVAR(-2,0), m0
68    ADDPD_MEM COVAR(-2,1), m1
69    lea     jq, [iq + 2]
70    cmp     jd, -2
71    jg .skip4x4
72.loop4x4:
73    ; Compute all 16 pairwise products of a 4x4 block
74    mulpd   m0, m4, m3
75    mulpd   m1, m5, m3
76    mulpd   m2, m6, m3
77    mulpd   m3, m3, m7
78    ADDPD_MEM COVAR(0,0), m0
79    ADDPD_MEM COVAR(0,1), m1
80    ADDPD_MEM COVAR(0,2), m2
81    ADDPD_MEM COVAR(0,3), m3
82    mova    m3, [varq + jq*8 + 16]
83    mulpd   m0, m4, m3
84    mulpd   m1, m5, m3
85    mulpd   m2, m6, m3
86    mulpd   m3, m3, m7
87    ADDPD_MEM COVAR(2,0), m0
88    ADDPD_MEM COVAR(2,1), m1
89    ADDPD_MEM COVAR(2,2), m2
90    ADDPD_MEM COVAR(2,3), m3
91    mova    m3, [varq + jq*8 + 32]
92    add covarq, 32
93    add     jq, 4
94    cmp     jd, -2
95    jle .loop4x4
96.skip4x4:
97    test    jd, jd
98    jg .skip2x4
99    mulpd   m4, m3
100    mulpd   m5, m3
101    mulpd   m6, m3
102    mulpd   m7, m3
103    ADDPD_MEM COVAR(0,0), m4
104    ADDPD_MEM COVAR(0,1), m5
105    ADDPD_MEM COVAR(0,2), m6
106    ADDPD_MEM COVAR(0,3), m7
107.skip2x4:
108    add     iq, 4
109    add covar2q, 4*COVAR_STRIDE+32
110    cmp     id, -2
111    jle .loopi
112    test    id, id
113    jg .ret
114    mov     jq, iq
115    %define covarq covar2q
116.loop2x1:
117    movsd   m0, [varq + iq*8]
118    movlhps m0, m0
119    mulpd   m0, [varq + jq*8]
120    ADDPD_MEM COVAR(0,0), m0
121    inc     iq
122    add covarq, COVAR_STRIDE
123    test    id, id
124    jle .loop2x1
125.ret:
126    REP_RET
127
128%if HAVE_AVX_EXTERNAL
129INIT_YMM avx
130cglobal update_lls, 3,6,8, ctx, var, count, i, j, count2
131    %define covarq ctxq
132    mov  countd, [ctxq + LLSModel.indep_count]
133    lea count2d, [countq-2]
134    xor     id, id
135.loopi:
136    ; Compute all 10 pairwise products of a 4x4 block that lies on the diagonal
137    mova    ymm1, [varq + iq*8]
138    vbroadcastsd ymm4, [varq + iq*8]
139    vbroadcastsd ymm5, [varq + iq*8 + 8]
140    vbroadcastsd ymm6, [varq + iq*8 + 16]
141    vbroadcastsd ymm7, [varq + iq*8 + 24]
142    vextractf128 xmm3, ymm1, 1
143    vmulpd  ymm0, ymm1, ymm4
144    vmulpd  ymm1, ymm1, ymm5
145    vmulpd  xmm2, xmm3, xmm6
146    vmulpd  xmm3, xmm3, xmm7
147    ADDPD_MEM COVAR(iq  ,0), ymm0
148    ADDPD_MEM COVAR(iq  ,1), ymm1
149    ADDPD_MEM COVAR(iq+2,2), xmm2
150    ADDPD_MEM COVAR(iq+2,3), xmm3
151    lea     jd, [iq + 4]
152    cmp     jd, count2d
153    jg .skip4x4
154.loop4x4:
155    ; Compute all 16 pairwise products of a 4x4 block
156    mova    ymm3, [varq + jq*8]
157    vmulpd  ymm0, ymm3, ymm4
158    vmulpd  ymm1, ymm3, ymm5
159    vmulpd  ymm2, ymm3, ymm6
160    vmulpd  ymm3, ymm3, ymm7
161    ADDPD_MEM COVAR(jq,0), ymm0
162    ADDPD_MEM COVAR(jq,1), ymm1
163    ADDPD_MEM COVAR(jq,2), ymm2
164    ADDPD_MEM COVAR(jq,3), ymm3
165    add     jd, 4
166    cmp     jd, count2d
167    jle .loop4x4
168.skip4x4:
169    cmp     jd, countd
170    jg .skip2x4
171    mova    xmm3, [varq + jq*8]
172    vmulpd  xmm0, xmm3, xmm4
173    vmulpd  xmm1, xmm3, xmm5
174    vmulpd  xmm2, xmm3, xmm6
175    vmulpd  xmm3, xmm3, xmm7
176    ADDPD_MEM COVAR(jq,0), xmm0
177    ADDPD_MEM COVAR(jq,1), xmm1
178    ADDPD_MEM COVAR(jq,2), xmm2
179    ADDPD_MEM COVAR(jq,3), xmm3
180.skip2x4:
181    add     id, 4
182    add covarq, 4*COVAR_STRIDE
183    cmp     id, count2d
184    jle .loopi
185    cmp     id, countd
186    jg .ret
187    mov     jd, id
188.loop2x1:
189    vmovddup xmm0, [varq + iq*8]
190    vmulpd   xmm0, [varq + jq*8]
191    ADDPD_MEM COVAR(jq,0), xmm0
192    inc     id
193    add covarq, COVAR_STRIDE
194    cmp     id, countd
195    jle .loop2x1
196.ret:
197    REP_RET
198%endif
199
200INIT_XMM sse2
201cglobal evaluate_lls, 3,4,2, ctx, var, order, i
202    ; This function is often called on the same buffer as update_lls, but with
203    ; an offset. They can't both be aligned.
204    ; Load halves rather than movu to avoid store-forwarding stalls, since the
205    ; input was initialized immediately prior to this function using scalar math.
206    %define coefsq ctxq
207    mov     id, orderd
208    imul    orderd, MAX_VARS
209    lea     coefsq, [ctxq + LLSModel.coeff + orderq*8]
210    movsd   m0, [varq]
211    movhpd  m0, [varq + 8]
212    mulpd   m0, [coefsq]
213    lea coefsq, [coefsq + iq*8]
214    lea   varq, [varq + iq*8]
215    neg     iq
216    add     iq, 2
217.loop:
218    movsd   m1, [varq + iq*8]
219    movhpd  m1, [varq + iq*8 + 8]
220    mulpd   m1, [coefsq + iq*8]
221    addpd   m0, m1
222    add     iq, 2
223    jl .loop
224    jg .skip1
225    movsd   m1, [varq + iq*8]
226    mulsd   m1, [coefsq + iq*8]
227    addpd   m0, m1
228.skip1:
229    movhlps m1, m0
230    addsd   m0, m1
231%if ARCH_X86_32
232    movsd  r0m, m0
233    fld   qword r0m
234%endif
235    RET
236