2 # Copyright 2002 Free Software Foundation, Inc.
4 # This file is part of GNU Radio
6 # GNU Radio is free software; you can redistribute it and/or modify
7 # it under the terms of the GNU General Public License as published by
8 # the Free Software Foundation; either version 2, or (at your option)
11 # GNU Radio is distributed in the hope that it will be useful,
12 # but WITHOUT ANY WARRANTY; without even the implied warranty of
13 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 # GNU General Public License for more details.
16 # You should have received a copy of the GNU General Public License
17 # along with GNU Radio; see the file COPYING. If not, write to
18 # the Free Software Foundation, Inc., 51 Franklin Street,
19 # Boston, MA 02110-1301, USA.
23 # input and taps are guarenteed to be 16 byte aligned.
24 # n_2_complex_blocks is != 0
27 # fcomplex_dotprod_generic (const float *input,
28 # const float *taps, unsigned n_2_complex_blocks, float *result)
37 # sum0 += input[0] * taps[0];
38 # sum1 += input[0] * taps[1];
39 # sum2 += input[1] * taps[2];
40 # sum3 += input[1] * taps[3];
45 # } while (--n_2_complex_blocks != 0);
48 # result[0] = sum0 + sum2;
49 # result[1] = sum1 + sum3;
53 # TODO: prefetch and better scheduling
58 .file "fcomplex_dotprod_sse.S"
62 .globl GLOB_SYMB(fcomplex_dotprod_sse)
63 DEF_FUNC_HEAD(fcomplex_dotprod_sse)
64 GLOB_SYMB(fcomplex_dotprod_sse):
67 movl 8(%ebp), %eax # input
68 movl 12(%ebp), %edx # taps
72 # xmm0 xmm1 xmm2 xmm3 are used to hold taps and the result of mults
73 # xmm4 xmm5 xmm6 xmm7 are used to hold the accumulated results
75 xorps %xmm4, %xmm4 # zero two accumulators
76 xorps %xmm5, %xmm5 # xmm5 holds zero for use below
78 # first handle any non-zero remainder of (n_2_complex_blocks % 4)
87 shufps $0x50, %xmm0, %xmm0 # b01010000
98 # set up for primary loop which is unrolled 4 times
101 movaps %xmm5, %xmm6 # zero remaining accumulators
104 shrl $2, %ecx # n_2_complex_blocks / 4
105 je .cleanup # if zero, take short path
107 # finish setup and loop priming
109 movlps 0(%eax), %xmm0
114 movlps 8(%eax), %xmm1
115 shufps $0x50, %xmm0, %xmm0
117 shufps $0x50, %xmm1, %xmm1
119 # we know ecx is not zero, we checked above,
120 # hence enter loop at top
125 movlps 0x10(%eax), %xmm2
131 movlps 0x18(%eax), %xmm3
132 shufps $0x50, %xmm2, %xmm2
134 mulps 0x10(%edx), %xmm1
136 shufps $0x50, %xmm3, %xmm3
139 movlps 0x20(%eax), %xmm0
143 mulps 0x20(%edx), %xmm2
145 movlps 0x28(%eax), %xmm1
146 shufps $0x50, %xmm0, %xmm0
148 mulps 0x30(%edx), %xmm3
150 shufps $0x50, %xmm1, %xmm1
157 # OK, now we've done with all the multiplies, but
158 # we still need to handle the unaccumulated
159 # products in xmm2 and xmm3
164 # now we want to add all accumulators into xmm4
171 # At this point, xmm4 contains 2x2 partial sums. We need
172 # to compute a "horizontal complex add" across xmm4.
174 .cleanup: # xmm4 = r1 i2 r3 i4
175 movl 20(%ebp), %eax # @result
176 movhlps %xmm4, %xmm0 # xmm0 = ?? ?? r1 r2
177 addps %xmm4, %xmm0 # xmm0 = ?? ?? r1+r3 i2+i4
178 movlps %xmm0, (%eax) # store low 2x32 bits (complex) to memory
183 FUNC_TAIL(fcomplex_dotprod_sse)
184 .ident "Hand coded x86 SSE assembly"