@@ -28,6 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28
28
#define ASSEMBLER
29
29
30
30
#include "common.h"
31
+ #include "loongarch64_asm.S"
31
32
32
33
/* Param */
33
34
#define M $r4
@@ -57,6 +58,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
57
58
#define T2 $r28
58
59
#define T3 $r29
59
60
#define T4 $r30
61
+ #define T5 $r17
62
+ #define T6 $r16
60
63
61
64
/* LSX vectors */
62
65
#define U0 $xr31
@@ -87,75 +90,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
87
90
#define a8 $f8
88
91
#define a9 $f9
89
92
90
-
91
- PROLOGUE
92
-
93
- LDARG BUFFER, $sp , 0
94
-
95
- addi .d $sp , $sp , -88
96
-
97
- SDARG $r23, $sp , 0
98
- SDARG $r24, $sp , 8
99
- SDARG $r25, $sp , 16
100
- SDARG $r26, $sp , 32
101
- SDARG $r27, $sp , 40
102
- SDARG $r28, $sp , 48
103
- SDARG $r29, $sp , 56
104
- SDARG $r30, $sp , 64
105
- SDARG $r31, $sp , 72
106
- ST ALPHA, $sp , 80
107
-
108
- xvldrepl.w VALPHA, $sp , 80
109
-
110
- slli.d LDA, LDA, BASE_SHIFT
111
- slli.d INCX, INCX, BASE_SHIFT
112
- slli.d INCY, INCY, BASE_SHIFT
113
-
114
- bge $r0, M, .L999
115
- bge $r0, N, .L999
116
-
117
- move J, $r0
118
- move JY, $r0
119
- move JX, $r0
120
- move AO1, A
121
-
122
- beq J , N, .L999
123
-
124
- .L01:
125
- MTC a2 , $r0 //temp2
126
- fldx.s a6 , X, JX
127
- fmul .s a3, ALPHA, a6 //temp1
128
- xvreplve0.w U3, U3
129
- xvreplve0.w U2, U2
130
-
131
- mul.w T0, J, LDA
132
- slli.d T1, J, BASE_SHIFT
133
- add.w T0, T0, T1
134
- fldx.s a6 , AO1, T0
135
- fldx.s a4 , Y, JY
136
- fmadd.s a4 , a3 , a6 , a4
137
- fstx.s a4 , Y, JY
138
-
139
- move IY, JY
140
- move IX, JX
141
- addi .d II, J, 1
142
- move I, II
143
- slli.d II, II, BASE_SHIFT
144
-
145
- sub .d T0, M, J
146
- addi .d T0, T0, -1
147
- srai.d T0, T0, 3
148
- add .d T0, T0, J
149
- addi .d T0, T0, 1
150
- beq I , T0, .L03
151
- bge I , T0, .L03
152
-
153
- mul.w T1, J, LDA
154
- add .d T1, T1, II
155
-
156
- .L02: /* /8 */
157
- xvldx U1, AO1, T1
158
-
93
+ .macro LOAD_Y_8
94
+ beqz T5, .L01_Y_0
159
95
add .d T2, IY, INCY
160
96
fldx.s $f4 , Y, T2
161
97
add .d T2, T2, INCY
@@ -180,11 +116,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
180
116
vextrins.w $vr8, $vr9, 0x10
181
117
vextrins.w $vr8, $vr10, 0x20
182
118
vextrins.w $vr8, $vr11, 0x30
183
- xvpermi.q U4, U8, 0x02
184
-
185
- xvfmadd.s U4, U3, U1, U4
186
-
187
- xvpermi.d U8, U4, 0xee
119
+ xvpermi.q U4, U8, 0x02
120
+ b .L01_Y_1
121
+ .L01_Y_0:
122
+ add .d T3, IY, INCY
123
+ xvldx U4, Y, T3
124
+ .L01_Y_1:
125
+ .endm
126
+
127
+ .macro STORE_Y_8
128
+ beqz T5, .L01_Y_2
129
+ xvpermi.d U8, U4, 0xee
188
130
vextrins.w $vr5, $vr4, 0x01
189
131
vextrins.w $vr6, $vr4, 0x02
190
132
vextrins.w $vr7, $vr4, 0x03
@@ -209,10 +151,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
209
151
fstx.s $f10 , Y, T2
210
152
add .d T2, T2, INCY
211
153
fstx.s $f11 , Y, T2
212
-
213
- slli.d T2, INCY, 3
214
- add .d IY, IY, T2
215
-
154
+ b .L01_Y_3
155
+ .L01_Y_2:
156
+ xvstx U4, Y, T3
157
+ .L01_Y_3:
158
+ .endm
159
+
160
+ .macro LOAD_X_8
161
+ beqz T6, .L01_X_0
216
162
add .d T2, IX, INCX
217
163
fldx.s $f4 , X, T2
218
164
add .d T2, T2, INCX
@@ -238,39 +184,103 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
238
184
vextrins.w $vr8, $vr10, 0x20
239
185
vextrins.w $vr8, $vr11, 0x30
240
186
xvpermi.q U4, U8, 0x02
187
+ b .L01_X_1
188
+ .L01_X_0:
189
+ add .d T3, IX, INCX
190
+ xvldx U4, X, T3
191
+ .L01_X_1:
192
+ .endm
193
+
194
+ PROLOGUE
241
195
242
- xvand.v $xr12 , $xr2, $xr2
196
+ addi .d $ sp , $sp , -88
243
197
244
- xvfmadd.s U2, U1, U4, U2
245
- xvfsub.s U2, U2, $xr12
198
+ SDARG $r23, $sp , 0
199
+ SDARG $r24, $sp , 8
200
+ SDARG $r25, $sp , 16
201
+ SDARG $r26, $sp , 32
202
+ SDARG $r27, $sp , 40
203
+ SDARG $r28, $sp , 48
204
+ SDARG $r29, $sp , 56
205
+ SDARG $r30, $sp , 64
206
+ SDARG $r31, $sp , 72
207
+ ST ALPHA, $sp , 80
246
208
247
- xvpickve.w U4, U2, 0x01
248
- xvpickve.w U5, U2, 0x02
249
- xvpickve.w U6, U2, 0x03
250
- xvpickve.w U7, U2, 0x04
251
- xvpickve.w U8, U2, 0x05
252
- xvpickve.w U9, U2, 0x06
253
- xvpickve.w U10, U2, 0x07
209
+ xvldrepl.w VALPHA, $sp , 80
254
210
255
- fadd .s $f2, $f2, $f4
256
- fadd .s $f2, $f2, $f5
257
- fadd .s $f2, $f2, $f6
258
- fadd .s $f2, $f2, $f7
259
- fadd .s $f2, $f2, $f8
260
- fadd .s $f2, $f2, $f9
261
- fadd .s $f2, $f2, $f10
262
- fadd .s $f2, $f2, $f12
211
+ addi .d T5, INCY, -1
212
+ addi .d T6, INCX, -1
213
+ slli.d LDA, LDA, BASE_SHIFT
214
+ slli.d INCX, INCX, BASE_SHIFT
215
+ slli.d INCY, INCY, BASE_SHIFT
263
216
264
- xvreplve0.d U2, U2
217
+ bge $r0, M, .L999
218
+ bge $r0, N, .L999
219
+
220
+ move J, $r0
221
+ move JY, $r0
222
+ move JX, $r0
223
+ move AO1, A
265
224
266
- slli.d T2, INCX, 3
267
- add .d IX, IX, T2
225
+ beq J , N, .L999
226
+
227
+ .L01:
228
+ xvxor.v U2, U2, U2
229
+ fldx.s a6 , X, JX
230
+ fmul .s a3, ALPHA, a6 //temp1
231
+ xvreplve0.w U3, U3
232
+
233
+ mul.w T0, J, LDA
234
+ slli.d T1, J, BASE_SHIFT
235
+ add.w T0, T0, T1
236
+ fldx.s a6 , AO1, T0
237
+ fldx.s a4 , Y, JY
238
+ fmadd.s a4 , a3 , a6 , a4
239
+ fstx.s a4 , Y, JY
240
+
241
+ move IY, JY
242
+ move IX, JX
243
+ addi .d II, J, 1
244
+ move I, II
245
+ slli.d II, II, BASE_SHIFT
246
+
247
+ sub .d T0, M, J
248
+ addi .d T0, T0, -1
249
+ srai.d T0, T0, 3
250
+ add .d T0, T0, J
251
+ addi .d T0, T0, 1
252
+ beq I , T0, .L03
253
+ bge I , T0, .L03
254
+
255
+ mul.w T1, J, LDA
256
+ add .d T1, T1, II
257
+
258
+ .L02: /* /8 */
259
+ xvldx U1, AO1, T1
260
+
261
+ LOAD_Y_8
262
+
263
+ xvfmadd.s U4, U3, U1, U4
264
+
265
+ STORE_Y_8
266
+
267
+ alsl.d IY, INCY, IY, 3
268
+
269
+ LOAD_X_8
270
+
271
+ xvfmadd.s U2, U1, U4, U2
272
+
273
+ alsl.d IX, INCX, IX, 3
268
274
269
275
addi .d II, II, 32
270
276
addi .d T1, T1, 32
271
277
addi .d I, I, 1
272
278
blt I , T0, .L02
273
279
280
+ //Acc U2
281
+ GACC xvf, s, U4, U2
282
+ fmov.d $f2 , $f4
283
+
274
284
.L03: /* &4 */
275
285
sub .d T0, M, J
276
286
addi .d T0, T0, -1
@@ -433,4 +443,4 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
433
443
addi .d $sp , $sp , 88
434
444
jirl $r0, $r1, 0x0
435
445
436
- EPILOGUE
446
+ EPILOGUE
0 commit comments