Skip to content

Commit e976557

Browse files
martin-frbgmartin-frbg
authored
Fix inline assembly constraints
rework indices to allow marking argument lda as input and output.
1 parent 9d8be15 commit e976557

File tree

1 file changed

+65
-65
lines changed

1 file changed

+65
-65
lines changed

‎kernel/x86_64/sgemv_n_microk_sandy-4.c

Lines changed: 65 additions & 65 deletions
Original file line numberDiff line numberDiff line change
@@ -39,14 +39,14 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
3939
__asm__ __volatile__
4040
(
4141
"vzeroupper \n\t"
42-
"vbroadcastss (%2), %%ymm12 \n\t" // x0
43-
"vbroadcastss 4(%2), %%ymm13 \n\t" // x1
44-
"vbroadcastss 8(%2), %%ymm14 \n\t" // x2
45-
"vbroadcastss 12(%2), %%ymm15 \n\t" // x3
46-
"vbroadcastss 16(%2), %%ymm0 \n\t" // x4
47-
"vbroadcastss 20(%2), %%ymm1 \n\t" // x5
48-
"vbroadcastss 24(%2), %%ymm2 \n\t" // x6
49-
"vbroadcastss 28(%2), %%ymm3 \n\t" // x7
42+
"vbroadcastss (%3), %%ymm12 \n\t" // x0
43+
"vbroadcastss 4(%3), %%ymm13 \n\t" // x1
44+
"vbroadcastss 8(%3), %%ymm14 \n\t" // x2
45+
"vbroadcastss 12(%3), %%ymm15 \n\t" // x3
46+
"vbroadcastss 16(%3), %%ymm0 \n\t" // x4
47+
"vbroadcastss 20(%3), %%ymm1 \n\t" // x5
48+
"vbroadcastss 24(%3), %%ymm2 \n\t" // x6
49+
"vbroadcastss 28(%3), %%ymm3 \n\t" // x7
5050

5151
"vbroadcastss (%9), %%ymm6 \n\t" // alpha
5252

@@ -55,21 +55,21 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
5555

5656
"vxorps %%xmm4 , %%xmm4 , %%xmm4 \n\t"
5757
"vxorps %%xmm5 , %%xmm5 , %%xmm5 \n\t"
58-
"vmovups (%3,%0,4), %%xmm7 \n\t" // 4 * y
58+
"vmovups (%4,%0,4), %%xmm7 \n\t" // 4 * y
5959

60-
"vmulps (%4,%0,4), %%xmm12, %%xmm8 \n\t"
61-
"vmulps (%5,%0,4), %%xmm13, %%xmm10 \n\t"
62-
"vmulps (%6,%0,4), %%xmm14, %%xmm9 \n\t"
63-
"vmulps (%7,%0,4), %%xmm15, %%xmm11 \n\t"
60+
"vmulps (%5,%0,4), %%xmm12, %%xmm8 \n\t"
61+
"vmulps (%6,%0,4), %%xmm13, %%xmm10 \n\t"
62+
"vmulps (%7,%0,4), %%xmm14, %%xmm9 \n\t"
63+
"vmulps (%8,%0,4), %%xmm15, %%xmm11 \n\t"
6464
"vaddps %%xmm4, %%xmm8 , %%xmm4 \n\t"
6565
"vaddps %%xmm5, %%xmm10, %%xmm5 \n\t"
6666
"vaddps %%xmm4, %%xmm9 , %%xmm4 \n\t"
6767
"vaddps %%xmm5, %%xmm11, %%xmm5 \n\t"
6868

69-
"vmulps (%4,%8,4), %%xmm0 , %%xmm8 \n\t"
70-
"vmulps (%5,%8,4), %%xmm1 , %%xmm10 \n\t"
71-
"vmulps (%6,%8,4), %%xmm2 , %%xmm9 \n\t"
72-
"vmulps (%7,%8,4), %%xmm3 , %%xmm11 \n\t"
69+
"vmulps (%5,%2,4), %%xmm0 , %%xmm8 \n\t"
70+
"vmulps (%6,%2,4), %%xmm1 , %%xmm10 \n\t"
71+
"vmulps (%7,%2,4), %%xmm2 , %%xmm9 \n\t"
72+
"vmulps (%8,%2,4), %%xmm3 , %%xmm11 \n\t"
7373
"vaddps %%xmm4, %%xmm8 , %%xmm4 \n\t"
7474
"vaddps %%xmm5, %%xmm10, %%xmm5 \n\t"
7575
"vaddps %%xmm4, %%xmm9 , %%xmm4 \n\t"
@@ -79,9 +79,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
7979
"vmulps %%xmm6, %%xmm4 , %%xmm5 \n\t"
8080
"vaddps %%xmm5, %%xmm7 , %%xmm5 \n\t"
8181

82-
"vmovups %%xmm5, (%3,%0,4) \n\t" // 4 * y
82+
"vmovups %%xmm5, (%4,%0,4) \n\t" // 4 * y
8383

84-
"addq $4, %8 \n\t"
84+
"addq $4, %2 \n\t"
8585
"addq $4, %0 \n\t"
8686
"subq $4, %1 \n\t"
8787

@@ -92,21 +92,21 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
9292

9393
"vxorps %%ymm4 , %%ymm4 , %%ymm4 \n\t"
9494
"vxorps %%ymm5 , %%ymm5 , %%ymm5 \n\t"
95-
"vmovups (%3,%0,4), %%ymm7 \n\t" // 8 * y
95+
"vmovups (%4,%0,4), %%ymm7 \n\t" // 8 * y
9696

97-
"vmulps (%4,%0,4), %%ymm12, %%ymm8 \n\t"
98-
"vmulps (%5,%0,4), %%ymm13, %%ymm10 \n\t"
99-
"vmulps (%6,%0,4), %%ymm14, %%ymm9 \n\t"
100-
"vmulps (%7,%0,4), %%ymm15, %%ymm11 \n\t"
97+
"vmulps (%5,%0,4), %%ymm12, %%ymm8 \n\t"
98+
"vmulps (%6,%0,4), %%ymm13, %%ymm10 \n\t"
99+
"vmulps (%7,%0,4), %%ymm14, %%ymm9 \n\t"
100+
"vmulps (%8,%0,4), %%ymm15, %%ymm11 \n\t"
101101
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
102102
"vaddps %%ymm5, %%ymm10, %%ymm5 \n\t"
103103
"vaddps %%ymm4, %%ymm9 , %%ymm4 \n\t"
104104
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
105105

106-
"vmulps (%4,%8,4), %%ymm0 , %%ymm8 \n\t"
107-
"vmulps (%5,%8,4), %%ymm1 , %%ymm10 \n\t"
108-
"vmulps (%6,%8,4), %%ymm2 , %%ymm9 \n\t"
109-
"vmulps (%7,%8,4), %%ymm3 , %%ymm11 \n\t"
106+
"vmulps (%5,%2,4), %%ymm0 , %%ymm8 \n\t"
107+
"vmulps (%6,%2,4), %%ymm1 , %%ymm10 \n\t"
108+
"vmulps (%7,%2,4), %%ymm2 , %%ymm9 \n\t"
109+
"vmulps (%8,%2,4), %%ymm3 , %%ymm11 \n\t"
110110
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
111111
"vaddps %%ymm5, %%ymm10, %%ymm5 \n\t"
112112
"vaddps %%ymm4, %%ymm9 , %%ymm4 \n\t"
@@ -116,9 +116,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
116116
"vmulps %%ymm6, %%ymm4 , %%ymm5 \n\t"
117117
"vaddps %%ymm5, %%ymm7 , %%ymm5 \n\t"
118118

119-
"vmovups %%ymm5, (%3,%0,4) \n\t" // 8 * y
119+
"vmovups %%ymm5, (%4,%0,4) \n\t" // 8 * y
120120

121-
"addq $8, %8 \n\t"
121+
"addq $8, %2 \n\t"
122122
"addq $8, %0 \n\t"
123123
"subq $8, %1 \n\t"
124124

@@ -134,45 +134,45 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
134134
"vxorps %%ymm4 , %%ymm4 , %%ymm4 \n\t"
135135
"vxorps %%ymm5 , %%ymm5 , %%ymm5 \n\t"
136136

137-
"prefetcht0 192(%4,%0,4) \n\t"
138-
"vmulps (%4,%0,4), %%ymm12, %%ymm8 \n\t"
139-
"vmulps 32(%4,%0,4), %%ymm12, %%ymm9 \n\t"
140137
"prefetcht0 192(%5,%0,4) \n\t"
141-
"vmulps (%5,%0,4), %%ymm13, %%ymm10 \n\t"
142-
"vmulps 32(%5,%0,4), %%ymm13, %%ymm11 \n\t"
138+
"vmulps (%5,%0,4), %%ymm12, %%ymm8 \n\t"
139+
"vmulps 32(%5,%0,4), %%ymm12, %%ymm9 \n\t"
140+
"prefetcht0 192(%6,%0,4) \n\t"
141+
"vmulps (%6,%0,4), %%ymm13, %%ymm10 \n\t"
142+
"vmulps 32(%6,%0,4), %%ymm13, %%ymm11 \n\t"
143143
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
144144
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
145145
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
146146
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
147147

148-
"prefetcht0 192(%6,%0,4) \n\t"
149-
"vmulps (%6,%0,4), %%ymm14, %%ymm8 \n\t"
150-
"vmulps 32(%6,%0,4), %%ymm14, %%ymm9 \n\t"
151148
"prefetcht0 192(%7,%0,4) \n\t"
152-
"vmulps (%7,%0,4), %%ymm15, %%ymm10 \n\t"
153-
"vmulps 32(%7,%0,4), %%ymm15, %%ymm11 \n\t"
149+
"vmulps (%7,%0,4), %%ymm14, %%ymm8 \n\t"
150+
"vmulps 32(%7,%0,4), %%ymm14, %%ymm9 \n\t"
151+
"prefetcht0 192(%8,%0,4) \n\t"
152+
"vmulps (%8,%0,4), %%ymm15, %%ymm10 \n\t"
153+
"vmulps 32(%8,%0,4), %%ymm15, %%ymm11 \n\t"
154154
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
155155
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
156156
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
157157
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
158158

159-
"prefetcht0 192(%4,%8,4) \n\t"
160-
"vmulps (%4,%8,4), %%ymm0 , %%ymm8 \n\t"
161-
"vmulps 32(%4,%8,4), %%ymm0 , %%ymm9 \n\t"
162-
"prefetcht0 192(%5,%8,4) \n\t"
163-
"vmulps (%5,%8,4), %%ymm1 , %%ymm10 \n\t"
164-
"vmulps 32(%5,%8,4), %%ymm1 , %%ymm11 \n\t"
159+
"prefetcht0 192(%5,%2,4) \n\t"
160+
"vmulps (%5,%2,4), %%ymm0 , %%ymm8 \n\t"
161+
"vmulps 32(%5,%2,4), %%ymm0 , %%ymm9 \n\t"
162+
"prefetcht0 192(%6,%2,4) \n\t"
163+
"vmulps (%6,%2,4), %%ymm1 , %%ymm10 \n\t"
164+
"vmulps 32(%6,%2,4), %%ymm1 , %%ymm11 \n\t"
165165
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
166166
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
167167
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
168168
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
169169

170-
"prefetcht0 192(%6,%8,4) \n\t"
171-
"vmulps (%6,%8,4), %%ymm2 , %%ymm8 \n\t"
172-
"vmulps 32(%6,%8,4), %%ymm2 , %%ymm9 \n\t"
173-
"prefetcht0 192(%7,%8,4) \n\t"
174-
"vmulps (%7,%8,4), %%ymm3 , %%ymm10 \n\t"
175-
"vmulps 32(%7,%8,4), %%ymm3 , %%ymm11 \n\t"
170+
"prefetcht0 192(%7,%2,4) \n\t"
171+
"vmulps (%7,%2,4), %%ymm2 , %%ymm8 \n\t"
172+
"vmulps 32(%7,%2,4), %%ymm2 , %%ymm9 \n\t"
173+
"prefetcht0 192(%8,%2,4) \n\t"
174+
"vmulps (%8,%2,4), %%ymm3 , %%ymm10 \n\t"
175+
"vmulps 32(%8,%2,4), %%ymm3 , %%ymm11 \n\t"
176176
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
177177
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
178178
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
@@ -181,13 +181,13 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
181181
"vmulps %%ymm6, %%ymm4 , %%ymm4 \n\t"
182182
"vmulps %%ymm6, %%ymm5 , %%ymm5 \n\t"
183183

184-
"vaddps (%3,%0,4), %%ymm4 , %%ymm4 \n\t" // 8 * y
185-
"vaddps 32(%3,%0,4), %%ymm5 , %%ymm5 \n\t" // 8 * y
184+
"vaddps (%4,%0,4), %%ymm4 , %%ymm4 \n\t" // 8 * y
185+
"vaddps 32(%4,%0,4), %%ymm5 , %%ymm5 \n\t" // 8 * y
186186

187-
"vmovups %%ymm4, (%3,%0,4) \n\t" // 8 * y
188-
"vmovups %%ymm5, 32(%3,%0,4) \n\t" // 8 * y
187+
"vmovups %%ymm4, (%4,%0,4) \n\t" // 8 * y
188+
"vmovups %%ymm5, 32(%4,%0,4) \n\t" // 8 * y
189189

190-
"addq $16, %8 \n\t"
190+
"addq $16, %2 \n\t"
191191
"addq $16, %0 \n\t"
192192
"subq $16, %1 \n\t"
193193
"jnz 1b \n\t"
@@ -197,15 +197,15 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
197197

198198
:
199199
"+r" (i), // 0
200-
"+r" (n) // 1
200+
"+r" (n), // 1
201+
"+r" (lda4) // 2
201202
:
202-
"r" (x), // 2
203-
"r" (y), // 3
204-
"r" (ap[0]), // 4
205-
"r" (ap[1]), // 5
206-
"r" (ap[2]), // 6
207-
"r" (ap[3]), // 7
208-
"r" (lda4), // 8
203+
"r" (x), // 3
204+
"r" (y), // 4
205+
"r" (ap[0]), // 5
206+
"r" (ap[1]), // 6
207+
"r" (ap[2]), // 7
208+
"r" (ap[3]), // 8
209209
"r" (alpha) // 9
210210
: "cc",
211211
"%xmm0", "%xmm1",

0 commit comments

Comments
 (0)