Debian Codegraph @

modules/lsp-dsp-lib/include/private/dsp/arch/x86/avx/fastconv/prepare.h

raw
1 /*
2 * Copyright (C) 2020 Linux Studio Plugins Project <https://lsp-plug.in/>
3 * (C) 2020 Vladimir Sadovnikov <[email protected]>
4 *
5 * This file is part of lsp-dsp-lib
6 * Created on: 31 мар. 2020 г.
7 *
8 * lsp-dsp-lib is free software: you can redistribute it and/or modify
9 * it under the terms of the GNU Lesser General Public License as published by
10 * the Free Software Foundation, either version 3 of the License, or
11 * any later version.
12 *
13 * lsp-dsp-lib 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
16 * GNU Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public License
19 * along with lsp-dsp-lib. If not, see <https://www.gnu.org/licenses/>.
20 */
21
22 #ifndef PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_PREPARE_H_
23 #define PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_PREPARE_H_
24
25 #ifndef PRIVATE_DSP_ARCH_X86_AVX_IMPL
26 #error "This header should not be included directly"
27 #endif /* PRIVATE_DSP_ARCH_X86_AVX_IMPL */
28
29 namespace lsp
30 {
31 namespace avx
32 {
33 #define FASTCONV_DIRECT_PREPARE_BODY(FMA_SEL) \
34 size_t off; \
35 \
36 ARCH_X86_ASM( \
37 __ASM_EMIT("lea (,%[np], 8), %[off]") \
38 __ASM_EMIT("vmovups 0x00(%[ak]), %%ymm6") /* ymm6 = x_re */ \
39 __ASM_EMIT("vmovups 0x20(%[ak]), %%ymm7") /* ymm7 = x_im */ \
40 __ASM_EMIT("vmovups 0x00(%[wk]), %%ymm4") /* ymm4 = w_re */ \
41 __ASM_EMIT("vmovups 0x20(%[wk]), %%ymm5") /* ymm5 = w_im */ \
42 __ASM_EMIT("vxorps %%ymm1, %%ymm1, %%ymm1") /* ymm1 = a_im = 0 */ \
43 /* x8 blocks */ \
44 __ASM_EMIT32("subl $8, %[np]") \
45 __ASM_EMIT64("sub $8, %[np]") \
46 __ASM_EMIT64("jb 2f") \
47 __ASM_EMIT("1:") \
48 __ASM_EMIT("vmovups 0x00(%[src]), %%ymm0") /* ymm0 = a_re = re */ \
49 __ASM_EMIT("vmulps %%ymm0, %%ymm7, %%ymm3") /* ymm3 = x_im * re */ \
50 __ASM_EMIT("vmulps %%ymm0, %%ymm6, %%ymm2") /* ymm2 = b_re = x_re * re */ \
51 __ASM_EMIT("vsubps %%ymm3, %%ymm1, %%ymm3") /* ymm3 = b_im = -x_im * re */ \
52 __ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") \
53 __ASM_EMIT("vmovups %%ymm1, 0x20(%[dst])") \
54 __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst], %[off])") \
55 __ASM_EMIT("vmovups %%ymm3, 0x20(%[dst], %[off])") \
56 __ASM_EMIT("add $0x20, %[src]") \
57 __ASM_EMIT("add $0x40, %[dst]") \
58 __ASM_EMIT32("subl $8, %[np]") \
59 __ASM_EMIT64("sub $8, %[np]") \
60 __ASM_EMIT("jb 2f") \
61 /* Rotate angle */ \
62 __ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \
63 __ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \
64 __ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \
65 __ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \
66 __ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \
67 __ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \
68 __ASM_EMIT("jmp 1b") \
69 __ASM_EMIT("2:") \
70 : [dst] "+r" (dst), [src] "+r" (src), \
71 [off] "=&r" (off), [np] __ASM_ARG_RW(np) \
72 : [ak] "r" (ak), [wk] "r" (wk) \
73 : "%xmm0", "%xmm1", "%xmm2", "%xmm3", \
74 "%xmm4", "%xmm5", "%xmm6", "%xmm7" \
75 )
76
77 #define FASTCONV_REVERSE_PREPARE_BODY(FMA_SEL) \
78 ARCH_X86_ASM( \
79 /* 2x blocks of 4x-butterfly loop */ \
80 __ASM_EMIT("sub $2, %[nb]") \
81 __ASM_EMIT("jb 2f") \
82 __ASM_EMIT("1:") \
83 __ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ \
84 __ASM_EMIT("vmovups 0x10(%[dst]), %%xmm4") /* xmm4 = r4 r5 r6 r7 */ \
85 __ASM_EMIT("vmovups 0x20(%[dst]), %%xmm2") /* xmm2 = i0 i1 i2 i3 */ \
86 __ASM_EMIT("vmovups 0x30(%[dst]), %%xmm6") /* xmm6 = i4 i5 i6 i7 */ \
87 __ASM_EMIT("vinsertf128 $1, 0x40(%[dst]), %%ymm0, %%ymm0") /* ymm0 = r0 r1 r2 r3 r8 r9 r10 r11 */ \
88 __ASM_EMIT("vinsertf128 $1, 0x50(%[dst]), %%ymm4, %%ymm4") /* ymm4 = r4 r5 r6 r7 r12 r13 r14 r15 */ \
89 __ASM_EMIT("vinsertf128 $1, 0x60(%[dst]), %%ymm2, %%ymm2") /* ymm2 = i0 i1 i2 i3 i8 i9 i10 i11 */ \
90 __ASM_EMIT("vinsertf128 $1, 0x70(%[dst]), %%ymm6, %%ymm6") /* ymm6 = i4 i5 i6 i7 i12 i13 i14 i15 */ \
91 /* 1st-order 4x butterfly */ \
92 __ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ \
93 __ASM_EMIT("vhsubps %%ymm6, %%ymm2, %%ymm3") /* ymm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ \
94 __ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ \
95 __ASM_EMIT("vhaddps %%ymm6, %%ymm2, %%ymm2") /* ymm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ \
96 /* 2nd-order 4x butterfly */ \
97 __ASM_EMIT("vblendps $0xaa, %%ymm3, %%ymm1, %%ymm4") /* ymm4 = r1' i3' r5' i7' */ \
98 __ASM_EMIT("vblendps $0xaa, %%ymm1, %%ymm3, %%ymm5") /* ymm5 = i1' r3' i5' r7' */ \
99 __ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ \
100 __ASM_EMIT("vhsubps %%ymm5, %%ymm2, %%ymm3") /* ymm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ \
101 __ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ \
102 __ASM_EMIT("vhaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ \
103 __ASM_EMIT("vblendps $0xcc, %%ymm1, %%ymm0, %%ymm4") /* ymm4 = r0" i4" r1" r5" */ \
104 __ASM_EMIT("vblendps $0xcc, %%ymm0, %%ymm1, %%ymm5") /* ymm5 = r2" r6" r3" r7" */ \
105 __ASM_EMIT("vshufps $0x88, %%ymm3, %%ymm2, %%ymm6") /* ymm6 = i0" i1" i2" i3" */ \
106 __ASM_EMIT("vshufps $0xdd, %%ymm3, %%ymm2, %%ymm7") /* ymm7 = i4" i5" i6" i7" */ \
107 __ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm2") /* ymm2 = r0" r1" r2" r3" */ \
108 __ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm3") /* ymm3 = r4" r5" r6" r7" */ \
109 /* 3rd-order 8x butterfly */ \
110 __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm3, %%ymm4") /* ymm4 = x_im * b_re */ \
111 __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm7, %%ymm5") /* ymm5 = x_im * b_im */ \
112 __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_re */ \
113 __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%ymm7, %%ymm7", "")) /* ymm7 = x_re * b_im */ \
114 __ASM_EMIT(FMA_SEL("vsubps %%ymm5, %%ymm3, %%ymm5", "vfmsub231ps 0x00 + %[FFT_A], %%ymm3, %%ymm5")) /* ymm5 = c_re = x_re * b_re - x_im * b_im */ \
115 __ASM_EMIT(FMA_SEL("vaddps %%ymm4, %%ymm7, %%ymm4", "vfmadd231ps 0x00 + %[FFT_A], %%ymm7, %%ymm4")) /* ymm4 = c_im = x_re * b_im + x_im * b_re */ \
116 __ASM_EMIT("vsubps %%ymm5, %%ymm2, %%ymm0") /* ymm0 = a_re - c_re */ \
117 __ASM_EMIT("vsubps %%ymm4, %%ymm6, %%ymm1") /* ymm1 = a_im - c_im */ \
118 __ASM_EMIT("vaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = a_re + c_re */ \
119 __ASM_EMIT("vaddps %%ymm4, %%ymm6, %%ymm3") /* ymm3 = a_im + c_im */ \
120 /* Store */ \
121 __ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") \
122 __ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") \
123 __ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") \
124 __ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") \
125 __ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") \
126 __ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") \
127 __ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") \
128 __ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") \
129 __ASM_EMIT("add $0x80, %[dst]") \
130 __ASM_EMIT("sub $2, %[nb]") \
131 __ASM_EMIT("jae 1b") \
132 /* 1x block of 4-butterfly */ \
133 __ASM_EMIT("2:") \
134 __ASM_EMIT("add $1, %[nb]") \
135 __ASM_EMIT("jl 4f") \
136 __ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ \
137 __ASM_EMIT("vmovups 0x10(%[dst]), %%xmm4") /* xmm4 = r4 r5 r6 r7 */ \
138 __ASM_EMIT("vmovups 0x20(%[dst]), %%xmm2") /* xmm2 = i0 i1 i2 i3 */ \
139 __ASM_EMIT("vmovups 0x30(%[dst]), %%xmm6") /* xmm6 = i4 i5 i6 i7 */ \
140 /* 1st-order 4x butterfly */ \
141 __ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ \
142 __ASM_EMIT("vhsubps %%xmm6, %%xmm2, %%xmm3") /* xmm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ \
143 __ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ \
144 __ASM_EMIT("vhaddps %%xmm6, %%xmm2, %%xmm2") /* xmm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ \
145 /* 2nd-order 4x butterfly */ \
146 __ASM_EMIT("vblendps $0xaa, %%xmm3, %%xmm1, %%xmm4") /* xmm4 = r1' i3' r5' i7' */ \
147 __ASM_EMIT("vblendps $0xaa, %%xmm1, %%xmm3, %%xmm5") /* xmm5 = i1' r3' i5' r7' */ \
148 __ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ \
149 __ASM_EMIT("vhsubps %%xmm5, %%xmm2, %%xmm3") /* xmm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ \
150 __ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ \
151 __ASM_EMIT("vhaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ \
152 __ASM_EMIT("vblendps $0xcc, %%xmm1, %%xmm0, %%xmm4") /* xmm4 = r0" i4" r1" r5" */ \
153 __ASM_EMIT("vblendps $0xcc, %%xmm0, %%xmm1, %%xmm5") /* xmm5 = r2" r6" r3" r7" */ \
154 __ASM_EMIT("vshufps $0x88, %%xmm3, %%xmm2, %%xmm6") /* xmm6 = i0" i1" i2" i3" */ \
155 __ASM_EMIT("vshufps $0xdd, %%xmm3, %%xmm2, %%xmm7") /* xmm7 = i4" i5" i6" i7" */ \
156 __ASM_EMIT("vshufps $0x88, %%xmm5, %%xmm4, %%xmm2") /* xmm2 = r0" r1" r2" r3" */ \
157 __ASM_EMIT("vshufps $0xdd, %%xmm5, %%xmm4, %%xmm3") /* xmm3 = r4" r5" r6" r7" */ \
158 /* 3rd-order 8x butterfly */ \
159 __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm3, %%xmm4") /* xmm4 = x_im * b_re */ \
160 __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm7, %%xmm5") /* xmm5 = x_im * b_im */ \
161 __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%xmm3, %%xmm3", "")) /* xmm3 = x_re * b_re */ \
162 __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%xmm7, %%xmm7", "")) /* xmm7 = x_re * b_im */ \
163 __ASM_EMIT(FMA_SEL("vsubps %%xmm5, %%xmm3, %%xmm5", "vfmsub231ps 0x00 + %[FFT_A], %%xmm3, %%xmm5")) /* xmm5 = c_re = x_re * b_re - x_im * b_im */ \
164 __ASM_EMIT(FMA_SEL("vaddps %%xmm4, %%xmm7, %%xmm4", "vfmadd231ps 0x00 + %[FFT_A], %%xmm7, %%xmm4")) /* xmm4 = c_im = x_re * b_im + x_im * b_re */ \
165 __ASM_EMIT("vsubps %%xmm5, %%xmm2, %%xmm0") /* xmm0 = a_re - c_re */ \
166 __ASM_EMIT("vsubps %%xmm4, %%xmm6, %%xmm1") /* xmm1 = a_im - c_im */ \
167 __ASM_EMIT("vaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = a_re + c_re */ \
168 __ASM_EMIT("vaddps %%xmm4, %%xmm6, %%xmm3") /* xmm3 = a_im + c_im */ \
169 /* Store */ \
170 __ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") \
171 __ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") \
172 __ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") \
173 __ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") \
174 __ASM_EMIT("4:") \
175 : [dst] "+r" (dst), [nb] "+r" (nb) \
176 : [FFT_A] "o" (FFT_A) \
177 : "cc", "memory", \
178 "%xmm0", "%xmm1", "%xmm2", "%xmm3", \
179 "%xmm4", "%xmm5", "%xmm6", "%xmm7" \
180 )
181
182 #define FMA_OFF(a, b) a
183 #define FMA_ON(a, b) b
184
185 static inline void fastconv_direct_prepare(float *dst, const float *src, const float *ak, const float *wk, size_t np)
186 {
187 FASTCONV_DIRECT_PREPARE_BODY(FMA_OFF);
188 }
189
190 static inline void fastconv_reverse_prepare(float *dst, size_t nb)
191 {
192 FASTCONV_REVERSE_PREPARE_BODY(FMA_OFF);
193 }
194
195 static inline void fastconv_direct_prepare_fma3(float *dst, const float *src, const float *ak, const float *wk, size_t np)
196 {
197 FASTCONV_DIRECT_PREPARE_BODY(FMA_OFF);
198 }
199
200 static inline void fastconv_reverse_prepare_fma3(float *dst, size_t nb)
201 {
202 FASTCONV_REVERSE_PREPARE_BODY(FMA_OFF);
203 }
204
205 static inline void fastconv_direct_unpack(float *dst, const float *src)
206 {
207 ARCH_X86_ASM(
208 __ASM_EMIT("vmovups (%[src]), %%xmm0")
209 __ASM_EMIT("vxorps %%ymm1, %%ymm1, %%ymm1")
210 __ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])")
211 __ASM_EMIT("vmovups %%ymm1, 0x20(%[dst])")
212 :
213 : [dst] "r" (dst), [src] "r" (src)
214 : "%xmm0", "%xmm1"
215 );
216 }
217
218 static inline void fastconv_reverse_unpack(float *dst, const float *src, size_t rank)
219 {
220 size_t blocks = 1 << rank;
221 float norm = 1.0f / float(blocks);
222
223 // Perform 4-element butterflies
224 ARCH_X86_ASM
225 (
226 __ASM_EMIT("vbroadcastss %[norm], %%ymm0")
227 // 16x blocks
228 __ASM_EMIT("sub $16, %[blocks]")
229 __ASM_EMIT("jb 2f")
230 __ASM_EMIT("vmovaps %%ymm0, %%ymm1")
231 __ASM_EMIT("1:")
232 __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */
233 __ASM_EMIT("vmulps 0x40(%[src]), %%ymm1, %%ymm3") /* ymm3 = r8 r9 r10 r11 r12 r13 r14 r15 */
234 __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])")
235 __ASM_EMIT("vmovups %%ymm3, 0x20(%[dst])")
236 __ASM_EMIT("add $0x80, %[dst]")
237 __ASM_EMIT("add $0x40, %[src]")
238 __ASM_EMIT("sub $16, %[blocks]")
239 __ASM_EMIT("jae 1b")
240 __ASM_EMIT("2:")
241 // 8x block
242 __ASM_EMIT("add $8, %[blocks]")
243 __ASM_EMIT("jl 4f")
244 __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */
245 __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])")
246 __ASM_EMIT("4:")
247
248 : [dst] "+r"(dst), [src] "+r" (src), [blocks] "+r" (blocks)
249 : [norm] "o" (norm)
250 : "cc", "memory",
251 "%xmm0", "%xmm1", "%xmm2", "%xmm3",
252 "%xmm4", "%xmm5", "%xmm6", "%xmm7"
253 );
254 }
255
256 static inline void fastconv_reverse_unpack_adding(float *dst, const float *src, size_t rank)
257 {
258 size_t blocks = 1 << rank;
259 float norm = 1.0f / float(blocks);
260
261 // Perform 4-element butterflies
262 ARCH_X86_ASM
263 (
264 __ASM_EMIT("vbroadcastss %[norm], %%ymm0")
265 // 16x blocks
266 __ASM_EMIT("sub $16, %[blocks]")
267 __ASM_EMIT("jb 2f")
268 __ASM_EMIT("vmovaps %%ymm0, %%ymm1")
269 __ASM_EMIT("1:")
270 __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */
271 __ASM_EMIT("vmulps 0x40(%[src]), %%ymm1, %%ymm3") /* ymm3 = r8 r9 r10 r11 r12 r13 r14 r15 */
272 __ASM_EMIT("vaddps 0x00(%[dst]), %%ymm2, %%ymm2")
273 __ASM_EMIT("vaddps 0x20(%[dst]), %%ymm3, %%ymm3")
274 __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])")
275 __ASM_EMIT("vmovups %%ymm3, 0x20(%[dst])")
276 __ASM_EMIT("add $0x80, %[dst]")
277 __ASM_EMIT("add $0x40, %[src]")
278 __ASM_EMIT("sub $16, %[blocks]")
279 __ASM_EMIT("jae 1b")
280 __ASM_EMIT("2:")
281 // 8x block
282 __ASM_EMIT("add $8, %[blocks]")
283 __ASM_EMIT("jl 4f")
284 __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */
285 __ASM_EMIT("vaddps 0x00(%[dst]), %%ymm2, %%ymm2")
286 __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])")
287 __ASM_EMIT("4:")
288
289 : [dst] "+r"(dst), [src] "+r" (src), [blocks] "+r" (blocks)
290 : [norm] "o" (norm)
291 : "cc", "memory",
292 "%xmm0", "%xmm1", "%xmm2", "%xmm3",
293 "%xmm4", "%xmm5", "%xmm6", "%xmm7"
294 );
295 }
296
297 #undef FASTCONV_DIRECT_PREPARE_BODY
298 #undef FASTCONV_REVERSE_PREPARE_BODY
299 #undef FMA_OFF
300 #undef FMA_ON
301 }
302 }
303
304 #endif /* PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_PREPARE_H_ */