A very simple and basic question about avx2 in fpc

[beria2]

A big request for help... I've been sitting here for twenty-four hours and I can't solve the problem. How to load data into avx2 registers in the internal fpc assembled correctly. In the program below, they are initialized with garbage, and the program outputs the value -1.2556805098698515E+249, which is not true.
 :'( :'( :'( :'( :'( :'( :'( :'( :'( :'( :'( :'(

  function AsmTest: double; assembler;
  asm
           VMOVAPD YMM0, ymmword ptr [rip + @Data]
           VEXTRACTF128 XMM0, YMM0, 1
           JMP     @Exit
           ALIGN   32
           @Data:
           DB      $40, $08, $FA, $29, $E1, $70, $A4, $F1
           DB      $41, $09, $FB, $2A, $E2, $71, $A5, $F2
           DB      $42, $0A, $FC, $2B, $E3, $72, $A6, $F3
           DB      $43, $0B, $FD, $2C, $E4, $73, $A7, $F4
           @Exit:
  end;
     

[ALLIGATOR]

Hmmm, at first glance, it looks like the value loads normally

[beria2]

Hmmm, at first glance, it looks like the value loads normally

And where do you have such a beautiful debugger in Lazarus? This is how I look at the contents of the processor registers.  According to him, there is garbage there. And the test result, that is, the second number from the array, is garbage, not the one whose binary contents are typed in there.. It should be 3.26711296000000E+05. In my opinion, the code is correct, but the result is strange....

[LV]

Maybe use the data as a double? The result should be 10Pi ~ 31.4159265359.

Code: Pascal  [Select][+][-]

1. program Project1;
2.  
3. {$asmmode intel}
4. uses
5.   SysUtils, Math;
6.  
7. var
8.   RawPtr: Pointer;
9.   AlignedPtr: Pointer;
10.   ResultValue: Double;
11.  
12. const
13.   ALIGNMENT = 32;
14.  
15. type
16.   TAVXData = array[0..3] of Double;
17.  
18. procedure AllocateAlignedMemory(var Raw, Aligned: Pointer; Size, Align: PtrUInt);
19. var
20.   p: PtrUInt;
21. begin
22.   GetMem(Raw, Size + Align);              // Reserve extra space for alignment
23.   p := (PtrUInt(Raw) + Align - 1) and not (Align - 1); // Align upwards
24.   Aligned := Pointer(p);
25. end;
26.  
27. function AsmTest(p: Pointer): Double; assembler;
28. asm
29.   mov     rax, p                      // Address of the array
30.   vmovapd ymm0, ymmword ptr [rax]     // Load 4 double values into ymm0
31.  
32.   // Add lower and upper halves of ymm0
33.   vextractf128 xmm1, ymm0, 1          // Extract upper 2 double values into xmm1
34.   vaddpd  xmm0, xmm0, xmm1            // xmm0 = (a0+a2, a1+a3)
35.  
36.   // Horizontally add remaining values
37.   vhaddpd xmm0, xmm0, xmm0            // xmm0 = (a0+a2+a1+a3, ...)
38.  
39.   // Return the result
40.   movsd   xmm1, xmm0                  // Move result to xmm1
41. end;
42.  
43.  
44.  
45. var
46.   Data: TAVXData = (pi, 2 * pi, 3 * pi, 4 * pi);
47.  
48. begin
49.   // Allocate memory
50.   AllocateAlignedMemory(RawPtr, AlignedPtr, SizeOf(Data), ALIGNMENT);
51.  
52.   // Copy data to aligned memory
53.   Move(Data, AlignedPtr^, SizeOf(Data));
54.  
55.   Writeln('RawPtr     = ', PtrUInt(RawPtr));
56.   Writeln('AlignedPtr = ', PtrUInt(AlignedPtr));
57.  
58.   // Check if AlignedPtr is a multiple of 32
59.   if (PtrUInt(AlignedPtr) mod ALIGNMENT) <> 0 then
60.     raise Exception.Create('Alignment error!');
61.  
62.   // Call assembler function
63.   ResultValue := AsmTest(AlignedPtr);
64.   Writeln('Returned value = ', ResultValue :0:10);
65.  
66.   FreeMem(RawPtr);
67.  
68.   Readln;
69. end.
70.  

Code: Text  [Select][+][-]

1. RawPtr     = 22902112
2. AlignedPtr = 22902112
3. Returned value = 31.4159265359
4.  

[ALLIGATOR]

Hmmm, at first glance, it looks like the value loads normally

And where do you have such a beautiful debugger in Lazarus? This is how I look at the contents of the processor registers.  According to him, there is garbage there. And the test result, that is, the second number from the array, is garbage, not the one whose binary contents are typed in there.. It should be 3.26711296000000E+05. In my opinion, the code is correct, but the result is strange....

  It's right here:
https://forum.lazarus.freepascal.org/index.php?topic=68962.0

[beria2]

A big request for help... I've been sitting here for twenty-four hours and I can't solve the problem. How to load data into avx2 registers in the internal fpc assembled correctly. In the program below, they are initialized with garbage, and the program outputs the value -1.2556805098698515E+249, which is not true.
 :'( :'( :'( :'( :'( :'( :'( :'( :'( :'( :'( :'(

  function AsmTest: double; assembler;
  asm
           VMOVAPD YMM0, ymmword ptr [rip + @Data]
           VEXTRACTF128 XMM0, YMM0, 1
           JMP     @Exit
           ALIGN   32
           @Data:
           DB      $40, $08, $FA, $29, $E1, $70, $A4, $F1
           DB      $41, $09, $FB, $2A, $E2, $71, $A5, $F2
           DB      $42, $0A, $FC, $2B, $E3, $72, $A6, $F3
           DB      $43, $0B, $FD, $2C, $E4, $73, $A7, $F4
           @Exit:
  end;
     

I did it!!!


  function AsmTest: double; assembler;
  asm
           VMOVAPD YMM0, ymmword ptr [rip + @Data]
           VPERMILPD XMM1, XMM0, 1
           MOVAPS  XMM0, XMM1
           JMP     @Exit
           ALIGN   32
           @Data:
           DB      $00, $00, $00, $00, $00, $00, $F0, $3F  // 1.0
           DB      $00, $00, $00, $00, $00, $00, $00, $40  // 2.0
           DB      $00, $00, $00, $00, $00, $00, $08, $40  // 3.0
           DB      $00, $00, $00, $00, $00, $00, $10, $40  // 4.0
           @Exit:
  end;
       

[marcov]

You don't need to stuff it in a procedure, just use a const.

Code: [Select]

const
      splitsh6 :  array[0..31] of byte = (  $00,$04,$08,$0C,$01,$05,$09,$0d,
                                            $02,$06,$0A,$0E,$03,$07,$0B,$0F,
                                            $00,$04,$08,$0C,$01,$05,$09,$0d,
                                            $02,$06,$0A,$0E,$03,$07,$0B,$0F);


procedure xx;
asm
 ...
       // in FPC code, loads of constants should now be 32-byte aligned (constmin)
       vmovdqa  ymm9, [rip+splitsh6]
...
end;

[/code]