I am trying to create a view of an array object to better utilise SIMD vectors on the x86_64 platform.
Here's the main idea:
type Char_Set_Index is range 0 .. 7;
type Char_Set_Element is mod 2 ** 32;
type Character_Set_Vector is array (Char_Set_Index) of Char_Set_Element
with Alignment => 32,Component_Size => 32, Object_Size => 256, Size => 256;
type Character_Set is array (Character) of Boolean
with Alignment => 32, Component_Size => 1, Object_Size => 256, Size => 256;
Essentially, some of the operations in Ada.Character.Maps can better be processed using SIMD arithmetic. For instance the "=" operation, perhaps coded as,
function "="
(Left, Right : in Character_Set)
return Boolean
is
(for all k in Character_Set'Range =>
(Left(k) = Right(k)));
.. gives us the following output
.LFB4:
.cfi_startproc
movq %rdi, %r8
movq %rsi, %rdi
xorl %esi, %esi
jmp .L6
.p2align 4,,10
.p2align 3
.L10:
addl $1, %esi
cmpl $256, %esi
je .L9
.L6:
movl %esi, %edx
movl %esi, %ecx
sarl $3, %edx
andl $7, %ecx
movslq %edx, %rdx
movzbl (%rdi,%rdx), %eax
xorb (%r8,%rdx), %al
shrb %cl, %al
testb $1, %al
je .L10
xorl %eax, %eax
ret
.L9:
movl $1, %eax
ret
.cfi_endproc
Critically, it is comparing each bit, and GCC won't vectorise it. However, if we write,
function "="
(Left, Right : in Character_Set)
return Boolean
is
u : aliased constant Character_Set_Vector
with Import, Address => Left'Address;
v : aliased constant Character_Set_Vector
with Import, Address => Right'Address;
Temp : array (Char_Set_Index) of Integer;
Sum : Integer;
begin
for j in Temp'Range loop
pragma Loop_Optimize (Vector);
Temp(j) := (if u(j) = v(j) then 0 else 1);
end loop;
Sum := 0;
for j in Temp'Range loop
Sum := Sum + Temp(j);
end loop;
return Sum = 0;
end "=";
We get the branch-free SIMD instructions that we kind of expect,
.cfi_startproc
vmovdqa (%rdi), %ymm1
vpcmpeqd (%rsi), %ymm1, %ymm1
vpandn .LC0(%rip), %ymm1, %ymm1
vextracti128 $0x1, %ymm1, %xmm0
vpaddd %xmm1, %xmm0, %xmm0
vpsrldq $8, %xmm0, %xmm1
vpaddd %xmm1, %xmm0, %xmm0
vpsrldq $4, %xmm0, %xmm1
vpaddd %xmm1, %xmm0, %xmm0
vmovd %xmm0, %eax
testl %eax, %eax
sete %al
vzeroupper
ret
.cfi_endproc
Which all works rather well. Now, the problem at hand. If you push this code through SPARK Ada, there are a number of complaints regarding alignment, aliasing, and constants, so you have to end up writing,
function "="
(Left, Right : in Character_Set)
return Boolean
is
Left_Aligned : constant Character_Set := Left
with Alignment => 32;
Right_Aligned : constant Character_Set := Right
with Alignment => 32;
u : aliased constant Character_Set_Vector
with Import, Alignment => 32, Address => Left_Aligned'Address;
v : aliased constant Character_Set_Vector
with Import, Alignment => 32, Address => Right_Aligned'Address;
Temp : array (Char_Set_Index) of Integer;
Sum : Integer;
begin
for j in Temp'Range loop
pragma Loop_Optimize (Vector);
Temp(j) := (if u(j) = v(j) then 0 else 1);
end loop;
Sum := 0;
for j in Temp'Range loop
Sum := Sum + Temp(j);
end loop;
return Sum = 0;
end "=";
which gives us an awful lot of precopying, presumably to ensure that everything is aligned OK - even though the declarations already have the correct alignment,
.cfi_startproc
pushq %rbp
.cfi_def_cfa_offset 16
.cfi_offset 6, -16
movq %rsp, %rbp
.cfi_def_cfa_register 6
andq $-32, %rsp
vmovdqa (%rdi), %xmm2
vmovdqa 16(%rdi), %xmm3
vmovdqa (%rsi), %xmm4
vmovdqa 16(%rsi), %xmm5
vmovdqa %xmm2, -64(%rsp)
vmovdqa %xmm3, -48(%rsp)
vmovdqa -64(%rsp), %ymm6
vmovdqa %xmm4, -32(%rsp)
vmovdqa %xmm5, -16(%rsp)
vpcmpeqd -32(%rsp), %ymm6, %ymm1
vpandn .LC0(%rip), %ymm1, %ymm1
vextracti128 $0x1, %ymm1, %xmm0
vpaddd %xmm1, %xmm0, %xmm0
vpsrldq $8, %xmm0, %xmm1
vpaddd %xmm1, %xmm0, %xmm0
vpsrldq $4, %xmm0, %xmm1
vpaddd %xmm1, %xmm0, %xmm0
vmovd %xmm0, %eax
testl %eax, %eax
sete %al
vzeroupper
leave
.cfi_def_cfa 7, 8
ret
.cfi_endproc
Obviously, the only reason one would even bother with this is for greater performance, however, the SPARK Ada rules seem too restrictive in this case, hurting performance. So, my question is, is there a better way of doing this that doesn't result in the excessive moving data around, where, as far as I can tell, it's not required.
Incidentally, Ada.Unchecked_Conversion similarly does a lot of moving data around at the beginning, too.
Also, I realise that I can justify the SPARK Ada checks (false-positive) so I can use the Ada version, but I am hoping that I am missing something, here, and that there is an easier way to do this.
Perhaps there is a way of vectorising arrays of Booleans?
EDIT: I am compiling it using
gnatmake -O3 -mavx2 -gnatn -gnatp -S name-of-package.adb
Login
