floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove all trailing whitespace except lapack-netlib

Browse Source 
Signed-off-by: Timothy Gu <timothygu99@gmail.com>
This commit is contained in:
Timothy Gu
2014-06-27 12:05:18 -07:00
parent d10db52edb
commit 6c2ead30f0
1423 changed files with 21229 additions and 21229 deletions
Download Patch File Download Diff File Expand all files Collapse all files
kernel/mips64/trsm_kernel_LT_loongson3a.S
+122 -122
View File
@@ -70,7 +70,7 @@
#define ALPHA $f15
PROLOGUE
daddiu $sp, $sp, -144
SDARG $16, 0($sp)
@@ -97,7 +97,7 @@
sdc1 $f23,136($sp)
#endif
# LT compute from left to right, top to bottom
LDARG OFFSET, 144($sp)
LDARG OFFSET, 144($sp)
dsll LDC, LDC, BASE_SHIFT # ldc
dsra J, N, 2 # j = nc/4
@@ -122,7 +122,7 @@
dsra I, M, 2 # i = mc/4
move KK, OFFSET # kk is the length of the rectangular data part of panel Ai
move AO, A # reset A
move AO, A # reset A
daddu C, CO4, LDC # fixed pointer C, the write back address
blez I, .L20
nop
@@ -137,7 +137,7 @@
LD b1, 0 * SIZE(B) # get 4b
LD b2, 1 * SIZE(B)
LD b3, 2 * SIZE(B)
LD b4, 3 * SIZE(B)
LD b4, 3 * SIZE(B)
MOV t13, t11 # clear result registers
MOV t23, t11
@@ -155,7 +155,7 @@
.align 3
.L12:
LD a5, 4 * SIZE(AO)
LD a5, 4 * SIZE(AO)
LD a6, 5 * SIZE(AO)
LD a7, 6 * SIZE(AO)
LD a8, 7 * SIZE(AO)
@@ -183,7 +183,7 @@
MADD t14, t14, a1, b4
MADD t24, t24, a2, b4
MADD t34, t34, a3, b4
MADD t44, t44, a4, b4
MADD t44, t44, a4, b4
LD a1, 8 * SIZE(AO)
LD a2, 9 * SIZE(AO)
@@ -213,7 +213,7 @@
MADD t14, t14, a5, b8
MADD t24, t24, a6, b8
MADD t34, t34, a7, b8
MADD t44, t44, a8, b8
MADD t44, t44, a8, b8
LD a5, 12 * SIZE(AO)
LD a6, 13 * SIZE(AO)
@@ -243,12 +243,12 @@
MADD t14, t14, a1, b4
MADD t24, t24, a2, b4
MADD t34, t34, a3, b4
MADD t44, t44, a4, b4
MADD t44, t44, a4, b4
daddiu AO, AO, 16 * SIZE # AO += 4mr*4kr
daddiu AO, AO, 16 * SIZE # AO += 4mr*4kr
daddiu BO, BO, 16 * SIZE # BO += 4nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -276,12 +276,12 @@
MADD t14, t14, a5, b8
MADD t24, t24, a6, b8
MADD t34, t34, a7, b8
MADD t44, t44, a8, b8
MADD t44, t44, a8, b8
daddiu L, L, -1
bgtz L, .L12
nop
.align 3
.L15:
@@ -291,7 +291,7 @@
.align 3
.L16:
MADD t11, t11, a1, b1
MADD t11, t11, a1, b1
MADD t21, t21, a2, b1
MADD t31, t31, a3, b1
MADD t41, t41, a4, b1
@@ -309,12 +309,12 @@
MADD t14, t14, a1, b4
MADD t24, t24, a2, b4
MADD t34, t34, a3, b4
MADD t44, t44, a4, b4
MADD t44, t44, a4, b4
daddiu AO, AO, 4 * SIZE # AO += 4mr
daddiu AO, AO, 4 * SIZE # AO += 4mr
daddiu BO, BO, 4 * SIZE # BO += 4nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -344,7 +344,7 @@
LD b6, 5 * SIZE(BO)
LD b7, 6 * SIZE(BO)
LD b8, 7 * SIZE(BO)
SUB t21, b5, t21
SUB t22, b6, t22
SUB t23, b7, t23
@@ -354,12 +354,12 @@
LD b2, 9 * SIZE(BO)
LD b3, 10 * SIZE(BO)
LD b4, 11 * SIZE(BO)
SUB t31, b1, t31
SUB t32, b2, t32
SUB t33, b3, t33
SUB t34, b4, t34
LD b5, 12 * SIZE(BO)
LD b6, 13 * SIZE(BO)
LD b7, 14 * SIZE(BO)
@@ -371,7 +371,7 @@
SUB t44, b8, t44
LD a1, 0 * SIZE(AO) # sa stores in col major
LD a1, 0 * SIZE(AO) # sa stores in col major
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -392,7 +392,7 @@
NMSUB t43, t43, a4, t13
NMSUB t44, t44, a4, t14
LD a5, 5 * SIZE(AO)
LD a6, 6 * SIZE(AO)
LD a7, 7 * SIZE(AO)
@@ -445,7 +445,7 @@
ST t43, 14 * SIZE(BO)
ST t44, 15 * SIZE(BO)
ST t11, 0 * SIZE(CO1) # write back
ST t11, 0 * SIZE(CO1) # write back
ST t21, 1 * SIZE(CO1)
ST t31, 2 * SIZE(CO1)
ST t41, 3 * SIZE(CO1)
@@ -491,7 +491,7 @@
.align 3
.L20:
andi I, M, 2 # mr=2,nr=4
andi I, M, 2 # mr=2,nr=4
blez I, .L50
nop
@@ -510,7 +510,7 @@
LD b1, 0 * SIZE(B) # get 4b
LD b2, 1 * SIZE(B)
LD b3, 2 * SIZE(B)
LD b4, 3 * SIZE(B)
LD b4, 3 * SIZE(B)
dsra L, KK, 2
blez L, .L25
@@ -519,7 +519,7 @@
.align 3
.L22:
LD a5, 2 * SIZE(AO)
LD a5, 2 * SIZE(AO)
LD a6, 3 * SIZE(AO)
LD b5, 4 * SIZE(BO)
LD b6, 5 * SIZE(BO)
@@ -567,10 +567,10 @@
MADD t14, t14, a3, b4
MADD t24, t24, a4, b4
daddiu AO, AO, 8 * SIZE # AO += 2mr*4kr
daddiu AO, AO, 8 * SIZE # AO += 2mr*4kr
daddiu BO, BO, 16 * SIZE # BO += 4nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
@@ -596,7 +596,7 @@
andi L, KK, 3
blez L, .L28
nop
.align 3
.L26:
MADD t11, t11, a1, b1 # 3rd compute
@@ -608,10 +608,10 @@
MADD t14, t14, a1, b4
MADD t24, t24, a2, b4
daddiu AO, AO, 2 * SIZE # AO += 2mr
daddiu AO, AO, 2 * SIZE # AO += 2mr
daddiu BO, BO, 4 * SIZE # BO += 4nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
@@ -643,7 +643,7 @@
SUB t24, b8, t24
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b2, 1 * SIZE(AO)
MUL t11, b1, t11
MUL t12, b1, t12
@@ -653,7 +653,7 @@
NMSUB t22, t22, b2, t12
NMSUB t23, t23, b2, t13
NMSUB t24, t24, b2, t14
LD b3, 3 * SIZE(AO)
MUL t21, b3, t21
MUL t22, b3, t22
@@ -705,7 +705,7 @@
.align 3
.L50:
andi I, M, 1 # mr=1,nr=4
andi I, M, 1 # mr=1,nr=4
blez I, .L29
nop
@@ -723,7 +723,7 @@
LD b1, 0 * SIZE(B) # get 4b
LD b2, 1 * SIZE(B)
LD b3, 2 * SIZE(B)
LD b4, 3 * SIZE(B)
LD b4, 3 * SIZE(B)
dsra L, KK, 2
blez L, .L55
@@ -732,7 +732,7 @@
.align 3
.L52:
LD a5, 1 * SIZE(AO)
LD a5, 1 * SIZE(AO)
LD b5, 4 * SIZE(BO)
LD b6, 5 * SIZE(BO)
LD b7, 6 * SIZE(BO)
@@ -765,10 +765,10 @@
MADD t13, t13, a3, b3
MADD t14, t14, a3, b4
daddiu AO, AO, 4 * SIZE # AO += mr*4kr
daddiu AO, AO, 4 * SIZE # AO += mr*4kr
daddiu BO, BO, 16 * SIZE # BO += 4nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
LD b3, 2 * SIZE(BO)
@@ -789,7 +789,7 @@
andi L, KK, 3
blez L, .L58
nop
.align 3
.L56:
MADD t11, t11, a1, b1 # 3rd compute
@@ -797,10 +797,10 @@
MADD t13, t13, a1, b3
MADD t14, t14, a1, b4
daddiu AO, AO, 1 * SIZE # AO += 2mr
daddiu AO, AO, 1 * SIZE # AO += 2mr
daddiu BO, BO, 4 * SIZE # BO += 4nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
LD b3, 2 * SIZE(BO)
@@ -823,7 +823,7 @@
SUB t14, b4, t14
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b1, 0 * SIZE(AO) # computes the triangular_part
MUL t11, b1, t11
MUL t12, b1, t12
MUL t13, b1, t13
@@ -858,8 +858,8 @@
move B, BO # fixed panel Bj
bgtz J, .L10
nop
.align 3
.L30:
andi J, N, 2 # nr=2
@@ -874,9 +874,9 @@
MOV t31, t11
MOV t41, t11
move KK, OFFSET
move KK, OFFSET
move AO, A # reset A
daddu C, CO2, LDC # fixed
daddu C, CO2, LDC # fixed
dsra I, M, 2 # I = mc/4
blez I, .L40
@@ -902,7 +902,7 @@
.align 3
.L32:
LD a5, 4 * SIZE(AO)
LD a5, 4 * SIZE(AO)
LD a6, 5 * SIZE(AO)
LD a7, 6 * SIZE(AO)
LD a8, 7 * SIZE(AO)
@@ -950,10 +950,10 @@
MADD t32, t32, a3, b4
MADD t42, t42, a4, b4
daddiu AO, AO, 16 * SIZE # AO += 4mr*4kr
daddiu AO, AO, 16 * SIZE # AO += 4mr*4kr
daddiu BO, BO, 8 * SIZE # BO += 2nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -973,30 +973,30 @@
bgtz L, .L32
nop
.align 3
.L35:
andi L, KK, 3
blez L, .L38
nop
.align 3
.L36:
MADD t11, t11, a1, b1 # 3rd compute
MADD t21, t21, a2, b1
MADD t31, t31, a3, b1
MADD t41, t41, a4, b1
MADD t12, t12, a1, b2
MADD t22, t22, a2, b2
MADD t32, t32, a3, b2
MADD t42, t42, a4, b2
daddiu AO, AO, 4 * SIZE # AO += 4mr
daddiu AO, AO, 4 * SIZE # AO += 4mr
daddiu BO, BO, 2 * SIZE # BO += 2nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -1027,7 +1027,7 @@
SUB t41, b7, t41
SUB t42, b8, t42
LD a1, 0 * SIZE(AO) # sa stores in col major
LD a1, 0 * SIZE(AO) # sa stores in col major
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -1040,7 +1040,7 @@
NMSUB t41, t41, a4, t11
NMSUB t42, t42, a4, t12
LD a5, 5 * SIZE(AO)
LD a6, 6 * SIZE(AO)
LD a7, 7 * SIZE(AO)
@@ -1091,7 +1091,7 @@
daddu AO, AO, L # move AO to the end of Ai
daddu BO, BO, TEMP
daddiu KK, KK, 4 #
daddiu KK, KK, 4 #
MTC $0, a1
MOV t11, a1
@@ -1122,12 +1122,12 @@
dsra L, KK, 2
blez L, .L45
move BO, B # reset B
move BO, B # reset B
.align 3
.L42:
LD a5, 2 * SIZE(AO)
LD a5, 2 * SIZE(AO)
LD a6, 3 * SIZE(AO)
LD b5, 2 * SIZE(BO)
LD b6, 3 * SIZE(BO)
@@ -1157,10 +1157,10 @@
MADD t12, t12, a3, b4
MADD t22, t22, a4, b4
daddiu AO, AO, 8 * SIZE # AO += 2mr*4kr
daddiu AO, AO, 8 * SIZE # AO += 2mr*4kr
daddiu BO, BO, 8 * SIZE # BO += 2nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
@@ -1173,14 +1173,14 @@
daddiu L, L, -1
bgtz L, .L42
nop
.align 3
.L45:
andi L, KK, 3
blez L, .L48
nop
.align 3
.L46:
MADD t11, t11, a1, b1 # 3rd compute
@@ -1188,10 +1188,10 @@
MADD t12, t12, a1, b2
MADD t22, t22, a2, b2
daddiu AO, AO, 2 * SIZE # AO += 2mr
daddiu AO, AO, 2 * SIZE # AO += 2mr
daddiu BO, BO, 2 * SIZE # BO += 2nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
@@ -1211,13 +1211,13 @@
SUB t21, b3, t21
SUB t22, b4, t22
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b2, 1 * SIZE(AO)
MUL t11, b1, t11
MUL t12, b1, t12
NMSUB t21, t21, b2, t11
NMSUB t22, t22, b2, t12
LD b3, 3 * SIZE(AO)
MUL t21, b3, t21
MUL t22, b3, t22
@@ -1266,12 +1266,12 @@
dsra L, KK, 2
blez L, .L65
move BO, B # reset B
move BO, B # reset B
.align 3
.L62:
LD a5, 1 * SIZE(AO)
LD a5, 1 * SIZE(AO)
LD b5, 2 * SIZE(BO)
LD b6, 3 * SIZE(BO)
@@ -1292,10 +1292,10 @@
MADD t11, t11, a3, b3 # 3rd compute
MADD t12, t12, a3, b4
daddiu AO, AO, 4 * SIZE # AO += mr*4kr
daddiu AO, AO, 4 * SIZE # AO += mr*4kr
daddiu BO, BO, 8 * SIZE # BO += 2nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
@@ -1305,23 +1305,23 @@
daddiu L, L, -1
bgtz L, .L62
nop
.align 3
.L65:
andi L, KK, 3
blez L, .L68
nop
.align 3
.L66:
MADD t11, t11, a1, b1 # 3rd compute
MADD t12, t12, a1, b2
daddiu AO, AO, 1 * SIZE # AO += 1mr
daddiu AO, AO, 1 * SIZE # AO += 1mr
daddiu BO, BO, 2 * SIZE # BO += 2nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD b1, 0 * SIZE(BO)
LD b2, 1 * SIZE(BO)
@@ -1336,7 +1336,7 @@
SUB t11, b1, t11
SUB t12, b2, t12
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b1, 0 * SIZE(AO) # computes the triangular_part
MUL t11, b1, t11
MUL t12, b1, t12
@@ -1360,7 +1360,7 @@
.align 3
.L49:
move B, BO
.align 3
.L70:
@@ -1396,7 +1396,7 @@
.align 3
.L72:
LD a5, 4 * SIZE(AO)
LD a5, 4 * SIZE(AO)
LD a6, 5 * SIZE(AO)
LD a7, 6 * SIZE(AO)
LD a8, 7 * SIZE(AO)
@@ -1432,10 +1432,10 @@
MADD t31, t31, a3, b3
MADD t41, t41, a4, b3
daddiu AO, AO, 16 * SIZE # AO += 4mr*4kr
daddiu AO, AO, 16 * SIZE # AO += 4mr*4kr
daddiu BO, BO, 4 * SIZE # BO += 1nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -1450,7 +1450,7 @@
daddiu L, L, -1
bgtz L, .L72
nop
.align 3
.L75:
@@ -1465,10 +1465,10 @@
MADD t31, t31, a3, b1
MADD t41, t41, a4, b1
daddiu AO, AO, 4 * SIZE # AO += 4mr
daddiu AO, AO, 4 * SIZE # AO += 4mr
daddiu BO, BO, 1 * SIZE # BO += 1nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -1490,7 +1490,7 @@
SUB t31, b3, t31
SUB t41, b4, t41
LD a1, 0 * SIZE(AO) # sa stores in col major
LD a1, 0 * SIZE(AO) # sa stores in col major
LD a2, 1 * SIZE(AO)
LD a3, 2 * SIZE(AO)
LD a4, 3 * SIZE(AO)
@@ -1556,10 +1556,10 @@
dsra L, KK, 2
blez L, .L85
move BO, B
.align 3
.L82:
LD a5, 2 * SIZE(AO)
LD a5, 2 * SIZE(AO)
LD a6, 3 * SIZE(AO)
LD b5, 1 * SIZE(BO)
@@ -1583,10 +1583,10 @@
MADD t11, t11, a3, b3 # 3rd compute
MADD t21, t21, a4, b3
daddiu AO, AO, 8 * SIZE # AO += 2mr*4kr
daddiu AO, AO, 8 * SIZE # AO += 2mr*4kr
daddiu BO, BO, 4 * SIZE # BO += 1nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD b1, 0 * SIZE(BO)
@@ -1597,7 +1597,7 @@
daddiu L, L, -1
bgtz L, .L82
nop
.align 3
.L85:
@@ -1610,10 +1610,10 @@
MADD t11, t11, a1, b1 # 3rd compute
MADD t21, t21, a2, b1
daddiu AO, AO, 2 * SIZE # AO += 2mr
daddiu AO, AO, 2 * SIZE # AO += 2mr
daddiu BO, BO, 1 * SIZE # BO += 1nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD a2, 1 * SIZE(AO)
LD b1, 0 * SIZE(BO)
@@ -1630,11 +1630,11 @@
SUB t11, b1, t11
SUB t21, b2, t21
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b2, 1 * SIZE(AO)
MUL t11, b1, t11
NMSUB t21, t21, b2, t11
LD b3, 3 * SIZE(AO)
MUL t21, b3, t21
@@ -1646,7 +1646,7 @@
daddiu CO1, CO1, 2 * SIZE
dsubu TEMP, K, KK
dsll L, TEMP, 1 + BASE_SHIFT
dsll TEMP, TEMP, 0 + BASE_SHIFT
@@ -1655,7 +1655,7 @@
daddiu KK, KK, 2
.align 3
.L90:
andi I, M, 1 # mr=1
@@ -1670,10 +1670,10 @@
dsra L, KK, 2
blez L, .L95
move BO, B
.align 3
.L92:
LD a5, 1 * SIZE(AO)
LD a5, 1 * SIZE(AO)
LD b5, 1 * SIZE(BO)
MADD t11, t11, a1, b1 # 1st compute
@@ -1688,10 +1688,10 @@
MADD t11, t11, a3, b3 # 3rd compute
daddiu AO, AO, 4 * SIZE # AO += 2mr*4kr
daddiu AO, AO, 4 * SIZE # AO += 2mr*4kr
daddiu BO, BO, 4 * SIZE # BO += 1nr*4kr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD b1, 0 * SIZE(BO)
MADD t11, t11, a7, b7 # 4th compute
@@ -1699,7 +1699,7 @@
daddiu L, L, -1
bgtz L, .L92
nop
.align 3
.L95:
andi L, KK, 3
@@ -1710,10 +1710,10 @@
.L96:
MADD t11, t11, a1, b1 # 3rd compute
daddiu AO, AO, 1 * SIZE # AO += 2mr
daddiu AO, AO, 1 * SIZE # AO += 2mr
daddiu BO, BO, 1 * SIZE # BO += 1nr
LD a1, 0 * SIZE(AO) # next
LD a1, 0 * SIZE(AO) # next
LD b1, 0 * SIZE(BO)
daddiu L, L, -1
@@ -1726,16 +1726,16 @@
SUB t11, b1, t11
LD b1, 0 * SIZE(AO) # computes the triangular_part
LD b1, 0 * SIZE(AO) # computes the triangular_part
MUL t11, b1, t11
ST t11, 0 * SIZE(BO)
ST t11, 0 * SIZE(CO1)
daddiu CO1, CO1, 1 * SIZE
dsubu TEMP, K, KK
dsll L, TEMP, BASE_SHIFT
dsll TEMP, TEMP, BASE_SHIFT