104
104
"movaps (%2,%0), %%xmm1 \n\t"
105
105
"movaps (%1,%0), %%xmm0 \n\t"
106
"movaps 16(%2,%0), %%xmm5 \n\t"
107
"movaps 16(%1,%0), %%xmm4 \n\t"
106
108
"movaps %%xmm1, %%xmm2 \n\t"
109
"movaps %%xmm5, %%xmm6 \n\t"
107
110
"shufps $0xA0, %%xmm1, %%xmm1 \n\t"
108
111
"shufps $0xF5, %%xmm2, %%xmm2 \n\t"
112
"shufps $0xA0, %%xmm5, %%xmm5 \n\t"
113
"shufps $0xF5, %%xmm6, %%xmm6 \n\t"
109
114
"mulps (%3,%0,2), %%xmm1 \n\t" // cre*re cim*re
110
115
"mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
116
"mulps 32(%3,%0,2), %%xmm5 \n\t" // cre*re cim*re
117
"mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
111
118
"addps %%xmm2, %%xmm1 \n\t"
119
"addps %%xmm6, %%xmm5 \n\t"
112
120
"movaps %%xmm0, %%xmm3 \n\t"
121
"movaps %%xmm4, %%xmm7 \n\t"
113
122
"addps %%xmm1, %%xmm0 \n\t"
114
123
"subps %%xmm1, %%xmm3 \n\t"
124
"addps %%xmm5, %%xmm4 \n\t"
125
"subps %%xmm5, %%xmm7 \n\t"
115
126
"movaps %%xmm0, (%1,%0) \n\t"
116
127
"movaps %%xmm3, (%2,%0) \n\t"
128
"movaps %%xmm4, 16(%1,%0) \n\t"
129
"movaps %%xmm7, 16(%2,%0) \n\t"
119
132
:"r"(p), "r"(p + nloops), "r"(cptr)
144
asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
158
asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1));
159
#define P1M1P1M1 "%%xmm8"
161
#define P1M1P1M1 "%4"
146
164
/* pre rotation */
148
166
in2 = input + n2 - 4;
150
/* Complex multiplication
151
Two complex products per iteration, we could have 4 with 8 xmm
152
registers, 8 with 16 xmm registers.
153
Maybe we should unroll more.
155
for (k = 0; k < n4; k += 2) {
168
/* Complex multiplication */
169
for (k = 0; k < n4; k += 4) {
157
171
"movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2
158
172
"movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1
173
"movaps -16+1*%0, %%xmm4 \n\t" // xmm4 = r0 X r1 X : in2
174
"movaps 16+1*%1, %%xmm7 \n\t" // xmm7 = X i1 X i0: in1
159
175
"movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
160
176
"movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
177
"movlps 8+1*%2, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos
178
"movlps 8+1*%3, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin
161
179
"shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
162
180
"shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
181
"shufps $95, %%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0
182
"shufps $160,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0
163
183
"unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
184
"unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0
164
185
"movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
165
"xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
186
"movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0
187
"xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
188
"xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0
166
189
"mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
190
"mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR
167
191
"shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
192
"shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0
168
193
"mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
194
"mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii
169
195
"addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
196
"addps %%xmm7, %%xmm4 \n\t" // xmm4 = result
170
197
::"m"(in2[-2*k]), "m"(in1[2*k]),
171
198
"m"(tcos[k]), "m"(tsin[k])
173
203
/* Should be in the same block, hack for gcc2.95 & gcc3 */
175
205
"movlps %%xmm0, %0 \n\t"
176
206
"movhps %%xmm0, %1 \n\t"
177
:"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]])
207
"movlps %%xmm4, %2 \n\t"
208
"movhps %%xmm4, %3 \n\t"
209
:"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]),
210
"=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]])
181
214
ff_fft_calc_sse(&s->fft, z);
183
/* Not currently needed, added for safety */
184
asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
218
#define P1M1P1M1 "%3"
186
221
/* post rotation + reordering */
187
for (k = 0; k < n4; k += 2) {
222
for (k = 0; k < n4; k += 4) {
189
224
"movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z
225
"movaps 16+1*%0, %%xmm4 \n\t" // xmm4 = i1 r1 i0 r0: z
190
226
"movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
227
"movlps 8+1*%1, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos
191
228
"movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0
229
"movaps %%xmm4, %%xmm7 \n\t" // xmm7 = i1 r1 i0 r0
192
230
"movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
231
"movlps 8+1*%2, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin
193
232
"shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
194
233
"shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
234
"shufps $160,%%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0
235
"shufps $245,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0
195
236
"unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
237
"unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0
196
238
"movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
197
"xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
239
"movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0
240
"xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
198
241
"mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
242
"xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0
243
"mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR
199
244
"shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
245
"shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0
200
246
"mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
247
"mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii
201
248
"addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
249
"addps %%xmm7, %%xmm4 \n\t" // xmm4 = result
202
250
"movaps %%xmm0, %0 \n\t"
251
"movaps %%xmm4, 16+1*%0\n\t"
204
253
:"m"(tcos[k]), "m"(tsin[k])