500 lines
7.0 KiB
ArmAsm
500 lines
7.0 KiB
ArmAsm
.text
|
|
.attribute 4, 16
|
|
.attribute 5, "rv64i2p0_m2p0_a2p0_f2p0_d2p0_c2p0"
|
|
.file "fft2.sy"
|
|
.globl multiply
|
|
.p2align 1
|
|
.type multiply,@function
|
|
multiply:
|
|
addi sp, sp, -48
|
|
sd ra, 40(sp)
|
|
sd s0, 32(sp)
|
|
sd s1, 24(sp)
|
|
sd s2, 16(sp)
|
|
sd s3, 8(sp)
|
|
mv s1, a1
|
|
beqz a1, .LBB0_3
|
|
mv s0, a0
|
|
li s2, 1
|
|
bne s1, s2, .LBB0_4
|
|
lui a0, 70493
|
|
addiw a0, a0, -2031
|
|
mul a0, s0, a0
|
|
srli a1, a0, 63
|
|
srai a0, a0, 58
|
|
addw a0, a0, a1
|
|
lui a1, 243712
|
|
addiw a1, a1, 1
|
|
mulw a0, a0, a1
|
|
subw a0, s0, a0
|
|
j .LBB0_6
|
|
.LBB0_3:
|
|
mv a0, s1
|
|
j .LBB0_6
|
|
.LBB0_4:
|
|
srliw a0, s1, 31
|
|
add s3, s1, a0
|
|
sraiw a1, s3, 1
|
|
mv a0, s0
|
|
call multiply
|
|
slliw a0, a0, 1
|
|
lui a1, 70493
|
|
addiw a1, a1, -2031
|
|
mul a2, a0, a1
|
|
srli a3, a2, 63
|
|
srai a2, a2, 58
|
|
addw a3, a3, a2
|
|
lui a2, 243712
|
|
addiw a2, a2, 1
|
|
mulw a3, a3, a2
|
|
andi a4, s3, -2
|
|
subw a4, s1, a4
|
|
subw a0, a0, a3
|
|
bne a4, s2, .LBB0_6
|
|
addw a0, a0, s0
|
|
mul a1, a0, a1
|
|
srli a3, a1, 63
|
|
srai a1, a1, 58
|
|
addw a1, a1, a3
|
|
mulw a1, a1, a2
|
|
subw a0, a0, a1
|
|
.LBB0_6:
|
|
ld ra, 40(sp)
|
|
ld s0, 32(sp)
|
|
ld s1, 24(sp)
|
|
ld s2, 16(sp)
|
|
ld s3, 8(sp)
|
|
addi sp, sp, 48
|
|
ret
|
|
.Lfunc_end0:
|
|
.size multiply, .Lfunc_end0-multiply
|
|
|
|
.globl power
|
|
.p2align 1
|
|
.type power,@function
|
|
power:
|
|
addi sp, sp, -32
|
|
sd ra, 24(sp)
|
|
sd s0, 16(sp)
|
|
sd s1, 8(sp)
|
|
sd s2, 0(sp)
|
|
beqz a1, .LBB1_3
|
|
mv s1, a1
|
|
mv s2, a0
|
|
srliw a0, a1, 31
|
|
add s0, a1, a0
|
|
sraiw a1, s0, 1
|
|
mv a0, s2
|
|
call power
|
|
mv a1, a0
|
|
call multiply
|
|
andi a1, s0, -2
|
|
subw a1, s1, a1
|
|
li a2, 1
|
|
bne a1, a2, .LBB1_4
|
|
mv a1, s2
|
|
ld ra, 24(sp)
|
|
ld s0, 16(sp)
|
|
ld s1, 8(sp)
|
|
ld s2, 0(sp)
|
|
addi sp, sp, 32
|
|
tail multiply
|
|
.LBB1_3:
|
|
li a0, 1
|
|
.LBB1_4:
|
|
ld ra, 24(sp)
|
|
ld s0, 16(sp)
|
|
ld s1, 8(sp)
|
|
ld s2, 0(sp)
|
|
addi sp, sp, 32
|
|
ret
|
|
.Lfunc_end1:
|
|
.size power, .Lfunc_end1-power
|
|
|
|
.globl memmove
|
|
.p2align 1
|
|
.type memmove,@function
|
|
memmove:
|
|
blez a3, .LBB2_4
|
|
slli a4, a3, 32
|
|
srli a4, a4, 32
|
|
slli a1, a1, 2
|
|
add a0, a0, a1
|
|
.LBB2_2:
|
|
lw a1, 0(a2)
|
|
sw a1, 0(a0)
|
|
addi a0, a0, 4
|
|
addi a4, a4, -1
|
|
addi a2, a2, 4
|
|
bnez a4, .LBB2_2
|
|
mv a0, a3
|
|
ret
|
|
.LBB2_4:
|
|
li a0, 0
|
|
ret
|
|
.Lfunc_end2:
|
|
.size memmove, .Lfunc_end2-memmove
|
|
|
|
.globl fft
|
|
.p2align 1
|
|
.type fft,@function
|
|
fft:
|
|
addi sp, sp, -80
|
|
sd ra, 72(sp)
|
|
sd s0, 64(sp)
|
|
sd s1, 56(sp)
|
|
sd s2, 48(sp)
|
|
sd s3, 40(sp)
|
|
sd s4, 32(sp)
|
|
sd s5, 24(sp)
|
|
sd s6, 16(sp)
|
|
sd s7, 8(sp)
|
|
li a4, 1
|
|
beq a2, a4, .LBB3_14
|
|
mv s7, a3
|
|
mv s4, a2
|
|
mv s2, a1
|
|
mv s3, a0
|
|
slli s6, a1, 2
|
|
blez a2, .LBB3_10
|
|
li a1, 0
|
|
srliw a2, s4, 1
|
|
slli a0, s4, 32
|
|
srli a0, a0, 32
|
|
add a3, s3, s6
|
|
lui a4, %hi(temp)
|
|
addi a4, a4, %lo(temp)
|
|
j .LBB3_5
|
|
.LBB3_3:
|
|
srliw a5, a1, 1
|
|
add a5, a5, a2
|
|
.LBB3_4:
|
|
lw s1, 0(a3)
|
|
slli a5, a5, 2
|
|
add a5, a5, a4
|
|
sw s1, 0(a5)
|
|
addi a1, a1, 1
|
|
addi a3, a3, 4
|
|
beq a0, a1, .LBB3_7
|
|
.LBB3_5:
|
|
andi a5, a1, 1
|
|
bnez a5, .LBB3_3
|
|
srliw a5, a1, 1
|
|
j .LBB3_4
|
|
.LBB3_7:
|
|
blez s4, .LBB3_10
|
|
add a1, s3, s6
|
|
lui a2, %hi(temp)
|
|
addi a2, a2, %lo(temp)
|
|
.LBB3_9:
|
|
lw a3, 0(a2)
|
|
sw a3, 0(a1)
|
|
addi a2, a2, 4
|
|
addi a0, a0, -1
|
|
addi a1, a1, 4
|
|
bnez a0, .LBB3_9
|
|
.LBB3_10:
|
|
srliw a0, s4, 31
|
|
addw a0, s4, a0
|
|
sraiw s1, a0, 1
|
|
mv a0, s7
|
|
mv a1, s7
|
|
call multiply
|
|
mv s5, a0
|
|
mv a0, s3
|
|
mv a1, s2
|
|
mv a2, s1
|
|
mv a3, s5
|
|
call fft
|
|
addw a1, s1, s2
|
|
mv a0, s3
|
|
mv a2, s1
|
|
mv a3, s5
|
|
call fft
|
|
li a0, 2
|
|
blt s4, a0, .LBB3_13
|
|
slli a0, s1, 32
|
|
srli s4, a0, 32
|
|
add a0, s1, s2
|
|
slli a0, a0, 2
|
|
add s1, s3, a0
|
|
add s0, s3, s6
|
|
li a0, 1
|
|
lui a1, 70493
|
|
addiw s3, a1, -2031
|
|
lui a1, 243712
|
|
addiw s5, a1, 1
|
|
.LBB3_12:
|
|
lw s6, 0(s0)
|
|
lw a1, 0(s1)
|
|
sext.w s2, a0
|
|
mv a0, s2
|
|
call multiply
|
|
addw a1, a0, s6
|
|
mul a2, a1, s3
|
|
srli a3, a2, 63
|
|
srai a2, a2, 58
|
|
addw a2, a2, a3
|
|
mulw a2, a2, s5
|
|
subw a1, a1, a2
|
|
sw a1, 0(s0)
|
|
subw a0, s6, a0
|
|
addw a0, a0, s5
|
|
mul a1, a0, s3
|
|
srli a2, a1, 63
|
|
srai a1, a1, 58
|
|
addw a1, a1, a2
|
|
mulw a1, a1, s5
|
|
subw a0, a0, a1
|
|
sw a0, 0(s1)
|
|
mv a0, s2
|
|
mv a1, s7
|
|
call multiply
|
|
addi s1, s1, 4
|
|
addi s4, s4, -1
|
|
addi s0, s0, 4
|
|
bnez s4, .LBB3_12
|
|
.LBB3_13:
|
|
li a4, 0
|
|
.LBB3_14:
|
|
mv a0, a4
|
|
ld ra, 72(sp)
|
|
ld s0, 64(sp)
|
|
ld s1, 56(sp)
|
|
ld s2, 48(sp)
|
|
ld s3, 40(sp)
|
|
ld s4, 32(sp)
|
|
ld s5, 24(sp)
|
|
ld s6, 16(sp)
|
|
ld s7, 8(sp)
|
|
addi sp, sp, 80
|
|
ret
|
|
.Lfunc_end3:
|
|
.size fft, .Lfunc_end3-fft
|
|
|
|
.globl main
|
|
.p2align 1
|
|
.type main,@function
|
|
main:
|
|
addi sp, sp, -80
|
|
sd ra, 72(sp)
|
|
sd s0, 64(sp)
|
|
sd s1, 56(sp)
|
|
sd s2, 48(sp)
|
|
sd s3, 40(sp)
|
|
sd s4, 32(sp)
|
|
sd s5, 24(sp)
|
|
sd s6, 16(sp)
|
|
sd s7, 8(sp)
|
|
lui a0, %hi(a)
|
|
addi a0, a0, %lo(a)
|
|
call getarray
|
|
mv s0, a0
|
|
lui a0, %hi(b)
|
|
addi a0, a0, %lo(b)
|
|
call getarray
|
|
mv s1, a0
|
|
li a0, 61
|
|
call _sysy_starttime
|
|
addw a0, s0, s1
|
|
addiw s6, a0, -1
|
|
li a0, 1
|
|
.LBB4_1:
|
|
mv s1, a0
|
|
slliw a0, a0, 1
|
|
blt s1, s6, .LBB4_1
|
|
lui s4, %hi(d)
|
|
sw s1, %lo(d)(s4)
|
|
lui s5, 243712
|
|
divuw a1, s5, s1
|
|
li a0, 3
|
|
call power
|
|
mv a3, a0
|
|
lui a0, %hi(a)
|
|
addi s2, a0, %lo(a)
|
|
mv a0, s2
|
|
li a1, 0
|
|
mv a2, s1
|
|
call fft
|
|
lw s3, %lo(d)(s4)
|
|
divw a1, s5, s3
|
|
li a0, 3
|
|
call power
|
|
mv a3, a0
|
|
lui a0, %hi(b)
|
|
addi s7, a0, %lo(b)
|
|
mv a0, s7
|
|
li a1, 0
|
|
mv a2, s3
|
|
call fft
|
|
lwu s0, %lo(d)(s4)
|
|
sext.w s3, s0
|
|
blez s3, .LBB4_5
|
|
mv s1, s2
|
|
.LBB4_4:
|
|
lw a0, 0(s1)
|
|
lw a1, 0(s7)
|
|
call multiply
|
|
sw a0, 0(s1)
|
|
addi s7, s7, 4
|
|
addi s0, s0, -1
|
|
addi s1, s1, 4
|
|
bnez s0, .LBB4_4
|
|
.LBB4_5:
|
|
divw a0, s5, s3
|
|
subw a1, s5, a0
|
|
li a0, 3
|
|
call power
|
|
mv a3, a0
|
|
mv a0, s2
|
|
li a1, 0
|
|
mv a2, s3
|
|
call fft
|
|
lwu s0, %lo(d)(s4)
|
|
sext.w a0, s0
|
|
blez a0, .LBB4_8
|
|
lui a1, 243712
|
|
addiw a1, a1, -1
|
|
call power
|
|
mv s2, a0
|
|
lui a0, %hi(a)
|
|
addi s1, a0, %lo(a)
|
|
.LBB4_7:
|
|
lw a0, 0(s1)
|
|
mv a1, s2
|
|
call multiply
|
|
sw a0, 0(s1)
|
|
addi s0, s0, -1
|
|
addi s1, s1, 4
|
|
bnez s0, .LBB4_7
|
|
.LBB4_8:
|
|
li a0, 80
|
|
call _sysy_stoptime
|
|
lui a0, %hi(a)
|
|
addi a1, a0, %lo(a)
|
|
mv a0, s6
|
|
call putarray
|
|
li a0, 0
|
|
ld ra, 72(sp)
|
|
ld s0, 64(sp)
|
|
ld s1, 56(sp)
|
|
ld s2, 48(sp)
|
|
ld s3, 40(sp)
|
|
ld s4, 32(sp)
|
|
ld s5, 24(sp)
|
|
ld s6, 16(sp)
|
|
ld s7, 8(sp)
|
|
addi sp, sp, 80
|
|
ret
|
|
.Lfunc_end4:
|
|
.size main, .Lfunc_end4-main
|
|
|
|
.type temp,@object
|
|
.bss
|
|
.globl temp
|
|
.p2align 2
|
|
temp:
|
|
.zero 8388608
|
|
.size temp, 8388608
|
|
|
|
.type a,@object
|
|
.globl a
|
|
.p2align 2
|
|
a:
|
|
.zero 8388608
|
|
.size a, 8388608
|
|
|
|
.type b,@object
|
|
.globl b
|
|
.p2align 2
|
|
b:
|
|
.zero 8388608
|
|
.size b, 8388608
|
|
|
|
.type d,@object
|
|
.section .sbss,"aw",@nobits
|
|
.globl d
|
|
.p2align 2
|
|
d:
|
|
.word 0
|
|
.size d, 4
|
|
|
|
.type _sysy_start,@object
|
|
.bss
|
|
.globl _sysy_start
|
|
.p2align 3
|
|
_sysy_start:
|
|
.zero 16
|
|
.size _sysy_start, 16
|
|
|
|
.type _sysy_end,@object
|
|
.globl _sysy_end
|
|
.p2align 3
|
|
_sysy_end:
|
|
.zero 16
|
|
.size _sysy_end, 16
|
|
|
|
.type _sysy_l1,@object
|
|
.globl _sysy_l1
|
|
.p2align 2
|
|
_sysy_l1:
|
|
.zero 4096
|
|
.size _sysy_l1, 4096
|
|
|
|
.type _sysy_l2,@object
|
|
.globl _sysy_l2
|
|
.p2align 2
|
|
_sysy_l2:
|
|
.zero 4096
|
|
.size _sysy_l2, 4096
|
|
|
|
.type _sysy_h,@object
|
|
.globl _sysy_h
|
|
.p2align 2
|
|
_sysy_h:
|
|
.zero 4096
|
|
.size _sysy_h, 4096
|
|
|
|
.type _sysy_m,@object
|
|
.globl _sysy_m
|
|
.p2align 2
|
|
_sysy_m:
|
|
.zero 4096
|
|
.size _sysy_m, 4096
|
|
|
|
.type _sysy_s,@object
|
|
.globl _sysy_s
|
|
.p2align 2
|
|
_sysy_s:
|
|
.zero 4096
|
|
.size _sysy_s, 4096
|
|
|
|
.type _sysy_us,@object
|
|
.globl _sysy_us
|
|
.p2align 2
|
|
_sysy_us:
|
|
.zero 4096
|
|
.size _sysy_us, 4096
|
|
|
|
.type _sysy_idx,@object
|
|
.section .sbss,"aw",@nobits
|
|
.globl _sysy_idx
|
|
.p2align 2
|
|
_sysy_idx:
|
|
.word 0
|
|
.size _sysy_idx, 4
|
|
|
|
.type c,@object
|
|
.bss
|
|
.globl c
|
|
.p2align 2
|
|
c:
|
|
.zero 8388608
|
|
.size c, 8388608
|
|
|
|
.ident "Debian clang version 14.0.6"
|
|
.section ".note.GNU-stack","",@progbits
|
|
.addrsig
|
|
.addrsig_sym a
|
|
.addrsig_sym b
|