Delphi update for FMA instruction sets

-> updated opcode to dB statements project (AVXPortToDelphi.dpr) for this FMA stuff
mikerabat · Mar 1, 2018 · 26611de · 26611de
1 parent 6e346d7
commit 26611de
Show file tree

Hide file tree

Showing 14 changed files with 278 additions and 245 deletions.
diff --git a/AVXMatrixAbsOperationsx64.pas b/AVXMatrixAbsOperationsx64.pas
@@ -120,7 +120,7 @@ procedure AVXMatrixAbsAligned(Dest : PDouble; const LineWidth, Width, Height : T
        jz @nextLine;
 
        {$IFDEF FPC}vmovsd xmm1, [rcx + rax];{$ELSE}db $C5,$FB,$10,$0C,$01;{$ENDIF} 
-       {$IFDEF FPC}vandpd xmm1, xmm1, xmm0;{$ELSE}db $C5,$F1,$54,$C8;{$ENDIF}  
+       {$IFDEF FPC}vandpd xmm1, xmm1, xmm0;{$ELSE}db $C5,$F1,$54,$C8;{$ENDIF} 
        {$IFDEF FPC}vmovsd [rcx + rax], xmm1;{$ELSE}db $C5,$FB,$11,$0C,$01;{$ENDIF} 
 
        @nextLine:

diff --git a/AVXMatrixMultOperationsx64.pas b/AVXMatrixMultOperationsx64.pas
@@ -567,7 +567,7 @@ procedure AVXMtxMultTria2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt2 :
       // for y := 0 to height1 - 1
       @@foryloop:
          // tmp := 0;
-         vxorpd xmm0, xmm0, xmm0;
+         {$IFDEF FPC}vxorpd xmm0, xmm0, xmm0;{$ELSE}db $C5,$F9,$57,$C0;{$ENDIF} 
 
          // r8, mt2
          mov rbx, r8;
@@ -580,13 +580,13 @@ procedure AVXMtxMultTria2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt2 :
       jz @@foridxloopend;
 
       @@foridxloop:
-         vmovsd xmm1, [rbx];
-         vmovsd xmm2, [rax + rsi];
+         {$IFDEF FPC}vmovsd xmm1, [rbx];{$ELSE}db $C5,$FB,$10,$0B;{$ENDIF} 
+         {$IFDEF FPC}vmovsd xmm2, [rax + rsi];{$ELSE}db $C5,$FB,$10,$14,$30;{$ENDIF} 
 
          add rbx, r9;  // + linewidth2
 
-         vmulsd xmm1, xmm1, xmm2;
-         vaddsd xmm0, xmm0, xmm1;
+         {$IFDEF FPC}vmulsd xmm1, xmm1, xmm2;{$ELSE}db $C5,$F3,$59,$CA;{$ENDIF} 
+         {$IFDEF FPC}vaddsd xmm0, xmm0, xmm1;{$ELSE}db $C5,$FB,$58,$C1;{$ENDIF} 
 
          add rsi, 8;
       jnz @@foridxloop;
@@ -598,7 +598,7 @@ procedure AVXMtxMultTria2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt2 :
       add rbx, r10;
       mov rsi, r12; // r12 = width2
       dec rsi;
-      vmovsd [rbx + 8*rsi], xmm0;
+      {$IFDEF FPC}vmovsd [rbx + 8*rsi], xmm0;{$ELSE}db $C5,$FB,$11,$04,$F3;{$ENDIF} 
 
       // inc(PByte(pmT1), LineWidth1);
       add rax, rdx;
@@ -616,7 +616,7 @@ procedure AVXMtxMultTria2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt2 :
    jnz @@forxloop;
 
    // cleanup stack
-   vzeroupper;
+   {$IFDEF FPC}vzeroupper;{$ELSE}db $C5,$F8,$77;{$ENDIF} 
 
    mov rbx, iRBX;
    mov rdi, iRDI;

diff --git a/AVXPrecompiled/AVXPortToDelphi.dpr b/AVXPrecompiled/AVXPortToDelphi.dpr
@@ -83,7 +83,7 @@ begin
      for i := 0 to slIn.Count - 1 do
      begin
           // AVX in the first few characters hints for a procedure/function name
-          if Pos('AVX', slIn[i] ) = 1 then
+          if (Pos('AVX', slIn[i] ) = 1) or (Pos('FMA', slIn[i]) = 1) then
           begin
                // search for the function name in the output file
                s := slIn[i];
@@ -139,11 +139,11 @@ begin
                                    idx := Pos('//', outLine);
                                    cmt := '';
                                    if idx > 0 then
-                                   begin  
+                                   begin
                                         cmt := Copy(outLine, idx, Length(outLine));
                                         outLine := Copy(outLine, 1, idx - 1);
                                    end;
-                                   
+
 
                                    idx := Pos('{$IFDEF FPC}', outLine);
                                    if idx > 0 then

diff --git a/AVXPrecompiled/processPasFiles.bat b/AVXPrecompiled/processPasFiles.bat
@@ -6,7 +6,9 @@ mkdir oFiles
 echo Copy all object files
 
 copy ..\Test\lib\i386-win32\AVX*.o .\oFiles\
+copy ..\Test\lib\i386-win32\FMA*.o .\oFiles\
 copy ..\Test\lib\x86_64-win64\AVX*x64.o .\oFiles\
+copy ..\Test\lib\x86_64-win64\FMA*x64.o .\oFiles\
 
 
 echo 

diff --git a/FMAMatrixMultOperations.pas b/FMAMatrixMultOperations.pas
@@ -144,8 +144,8 @@ procedure FMAMatrixMultAligned(dest : PDouble; const destLineWidth : TASMNativeI
              {$IFDEF FPC}vinsertf128 ymm3, ymm3, xmm5, 1;{$ELSE}db $C4,$E3,$65,$18,$DD,$01;{$ENDIF} 
 
              // now multiply and add
-             vfmadd231pd ymm0, ymm2, ymm7;
-             vfmadd231pd ymm1, ymm3, ymm7;
+             {$IFDEF FPC}vfmadd231pd ymm0, ymm2, ymm7;{$ELSE}db $C4,$E2,$ED,$B8,$C7;{$ENDIF} 
+             {$IFDEF FPC}vfmadd231pd ymm1, ymm3, ymm7;{$ELSE}db $C4,$E2,$E5,$B8,$CF;{$ENDIF} 
           add eax, 32;
           jl @@InnerLoop1;
 
@@ -329,8 +329,8 @@ procedure FMAMatrixMultUnAligned(dest : PDouble; const destLineWidth : TASMNativ
              {$IFDEF FPC}vinsertf128 ymm3, ymm3, xmm5, 1;{$ELSE}db $C4,$E3,$65,$18,$DD,$01;{$ENDIF} 
 
              // now multiply and add
-             vfmadd231pd ymm0, ymm2, ymm7;
-             vfmadd231pd ymm1, ymm3, ymm7;
+             {$IFDEF FPC}vfmadd231pd ymm0, ymm2, ymm7;{$ELSE}db $C4,$E2,$ED,$B8,$C7;{$ENDIF} 
+             {$IFDEF FPC}vfmadd231pd ymm1, ymm3, ymm7;{$ELSE}db $C4,$E2,$E5,$B8,$CF;{$ENDIF} 
 
           add eax, 32;
           jl @@InnerLoop1;
@@ -495,8 +495,8 @@ procedure FMAMtxMultTria2T1StoreT1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt
             jg @@foriloopFMAend;
 
             {$IFDEF FPC}vmovupd ymm1, [ecx + eax - 32];{$ELSE}db $C5,$FD,$10,$4C,$01,$E0;{$ENDIF} 
-            {$IFDEF FPC}vmovupd ymm2, [ebx + eax - 32];{$ELSE}db $C5,$FD,$10,$54,$03,$E0;{$ENDIF}
-            vfmadd231pd ymm0, ymm1, ymm2;
+            {$IFDEF FPC}vmovupd ymm2, [ebx + eax - 32];{$ELSE}db $C5,$FD,$10,$54,$03,$E0;{$ENDIF} 
+            {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
 
          jmp @@foriloopFMA;
 
@@ -513,7 +513,7 @@ procedure FMAMtxMultTria2T1StoreT1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt
             // two elements at a time:
             {$IFDEF FPC}vmovupd xmm1, [ecx + eax];{$ELSE}db $C5,$F9,$10,$0C,$01;{$ENDIF} 
             {$IFDEF FPC}vmovupd xmm2, [ebx + eax];{$ELSE}db $C5,$F9,$10,$14,$03;{$ENDIF} 
-            vfmadd231pd xmm0, xmm1, xmm2;
+            {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
 
             add eax, 16;
          jnz @@foriloop;
@@ -629,7 +629,7 @@ procedure FMAMtxMultTria2TUpperUnit(dest : PDouble; LineWidthDest : TASMNativeIn
              // 4 elements at a time:
              {$IFDEF FPC}vmovupd ymm1, [edx + eax - 32];{$ELSE}db $C5,$FD,$10,$4C,$02,$E0;{$ENDIF} 
              {$IFDEF FPC}vmovupd ymm2, [ebx + eax - 32];{$ELSE}db $C5,$FD,$10,$54,$03,$E0;{$ENDIF} 
-             vfmadd231pd ymm0, ymm1, ymm2;
+             {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
          jmp @@foriloop;
 
          @@foriloopend:
@@ -643,7 +643,7 @@ procedure FMAMtxMultTria2TUpperUnit(dest : PDouble; LineWidthDest : TASMNativeIn
          // need to process two more elements:
          {$IFDEF FPC}vmovupd xmm1, [edx + eax];{$ELSE}db $C5,$F9,$10,$0C,$02;{$ENDIF} 
          {$IFDEF FPC}vmovupd xmm2, [ebx + eax];{$ELSE}db $C5,$F9,$10,$14,$03;{$ENDIF} 
-         vfmadd231pd xmm0, xmm1, xmm2;
+         {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
 
          @@finalizeloop:
 
@@ -774,7 +774,7 @@ procedure FMAMtxMultTria2T1(dest : PDouble; LineWidthDest : TASMNativeInt; mt1 :
                {$IFDEF FPC}vmovddup xmm1, [esi];{$ELSE}db $C5,$FB,$12,$0E;{$ENDIF} 
                {$IFDEF FPC}vmovupd xmm2, [edi];{$ELSE}db $C5,$F9,$10,$17;{$ENDIF} 
 
-               vfmadd231pd xmm0, xmm1, xmm2;
+               {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
 
                //inc(PByte(valCounter1), LineWidth1);
                //inc(PByte(valCounter2), LineWidth2);
@@ -896,11 +896,11 @@ procedure FMAMtxMultLowTria2T2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt;
          @@foridxlongloop:
             {$IFDEF FPC}vmovupd ymm1, [eax + edi - 64];{$ELSE}db $C5,$FD,$10,$4C,$38,$C0;{$ENDIF} 
             {$IFDEF FPC}vmovupd ymm2, [edx + edi - 64];{$ELSE}db $C5,$FD,$10,$54,$3A,$C0;{$ENDIF} 
-            vfmadd231pd ymm0, ymm1, ymm2;
+            {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
 
             {$IFDEF FPC}vmovupd ymm1, [eax + edi - 32];{$ELSE}db $C5,$FD,$10,$4C,$38,$E0;{$ENDIF} 
             {$IFDEF FPC}vmovupd ymm2, [edx + edi - 32];{$ELSE}db $C5,$FD,$10,$54,$3A,$E0;{$ENDIF} 
-            vfmadd231pd ymm0, ymm1, ymm2;
+            {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
          add edi, 64;
          jl @@foridxlongloop;
 
@@ -915,7 +915,7 @@ procedure FMAMtxMultLowTria2T2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt;
          @@foridxSSEloop:
             {$IFDEF FPC}vmovupd xmm1, [eax + edi - 16];{$ELSE}db $C5,$F9,$10,$4C,$38,$F0;{$ENDIF} 
             {$IFDEF FPC}vmovupd xmm2, [edx + edi - 16];{$ELSE}db $C5,$F9,$10,$54,$3A,$F0;{$ENDIF} 
-            vfmadd231pd xmm0, xmm1, xmm2;
+            {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
          add edi, 16;
          jl @@foridxSSEloop;
 

diff --git a/FMAMatrixMultOperationsx64.pas b/FMAMatrixMultOperationsx64.pas
@@ -164,8 +164,8 @@ procedure FMAMatrixMultAligned(dest : PDouble; const destLineWidth : TASMNativeI
              {$IFDEF FPC}vinsertf128 ymm3, ymm3, xmm5, 1;{$ELSE}db $C4,$E3,$65,$18,$DD,$01;{$ENDIF} 
 
              // now multiply and add
-             vfmadd231pd ymm0, ymm7, ymm2;
-             vfmadd231pd ymm1, ymm7, ymm3;
+             {$IFDEF FPC}vfmadd231pd ymm0, ymm7, ymm2;{$ELSE}db $C4,$E2,$C5,$B8,$C2;{$ENDIF} 
+             {$IFDEF FPC}vfmadd231pd ymm1, ymm7, ymm3;{$ELSE}db $C4,$E2,$C5,$B8,$CB;{$ENDIF} 
           add rax, 32;
           jl @@InnerLoop1;
 
@@ -380,8 +380,8 @@ procedure FMAMatrixMultUnAligned(dest : PDouble; const destLineWidth : TASMNativ
               {$IFDEF FPC}vinsertf128 ymm3, ymm3, xmm5, 1;{$ELSE}db $C4,$E3,$65,$18,$DD,$01;{$ENDIF} 
 
               // now multiply and add
-              vfmadd231pd ymm0, ymm7, ymm2;
-              vfmadd231pd ymm1, ymm7, ymm3;
+              {$IFDEF FPC}vfmadd231pd ymm0, ymm7, ymm2;{$ELSE}db $C4,$E2,$C5,$B8,$C2;{$ENDIF} 
+              {$IFDEF FPC}vfmadd231pd ymm1, ymm7, ymm3;{$ELSE}db $C4,$E2,$C5,$B8,$CB;{$ENDIF} 
            add rax, 32;
            jl @@InnerLoop1;
 
@@ -563,7 +563,7 @@ procedure FMAMtxMultTria2T1StoreT1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt
 
             {$IFDEF FPC}vmovupd ymm1, [rcx + rax - 32];{$ELSE}db $C5,$FD,$10,$4C,$01,$E0;{$ENDIF} 
             {$IFDEF FPC}vmovupd ymm2, [rbx + rax - 32];{$ELSE}db $C5,$FD,$10,$54,$03,$E0;{$ENDIF} 
-            vfmadd231pd ymm0, ymm1, ymm2;
+            {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
          jmp @@foriloopFMA;
 
          @@foriloopFMAend:
@@ -579,7 +579,7 @@ procedure FMAMtxMultTria2T1StoreT1(mt1 : PDouble; LineWidth1 : TASMNativeInt; mt
             // two elements at a time:
             {$IFDEF FPC}vmovupd xmm1, [rcx + rax];{$ELSE}db $C5,$F9,$10,$0C,$01;{$ENDIF} 
             {$IFDEF FPC}vmovupd xmm2, [rbx + rax];{$ELSE}db $C5,$F9,$10,$14,$03;{$ENDIF} 
-            vfmadd231pd xmm0, xmm1, xmm2;
+            {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
 
             add rax, 16;
          jnz @@foriloop;
@@ -712,7 +712,7 @@ procedure FMAMtxMultTria2TUpperUnit(dest : PDouble; LineWidthDest : TASMNativeIn
              // two elements at a time:
              {$IFDEF FPC}vmovupd ymm1, [r8 + rax - 32];{$ELSE}db $C4,$C1,$7D,$10,$4C,$00,$E0;{$ENDIF} 
              {$IFDEF FPC}vmovupd ymm2, [rbx + rax - 32];{$ELSE}db $C5,$FD,$10,$54,$03,$E0;{$ENDIF} 
-             vfmadd231pd ymm0, ymm1, ymm2;
+             {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
          jmp @@foriloop;
 
          @@foriloopend:
@@ -726,7 +726,7 @@ procedure FMAMtxMultTria2TUpperUnit(dest : PDouble; LineWidthDest : TASMNativeIn
          // need to process two more elements:
          {$IFDEF FPC}vmovupd xmm1, [r8 + rax];{$ELSE}db $C4,$C1,$79,$10,$0C,$00;{$ENDIF} 
          {$IFDEF FPC}vmovupd xmm2, [rbx + rax];{$ELSE}db $C5,$F9,$10,$14,$03;{$ENDIF} 
-         vfmadd231pd xmm0, xmm1, xmm2;
+         {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
 
          @@finalizeloop:
 
@@ -884,7 +884,7 @@ procedure FMAMtxMultTria2T1(dest : PDouble; LineWidthDest : TASMNativeInt; mt1 :
              {$IFDEF FPC}vmovddup xmm1, [rsi];{$ELSE}db $C5,$FB,$12,$0E;{$ENDIF} 
              {$IFDEF FPC}vmovupd xmm2, [rdi];{$ELSE}db $C5,$F9,$10,$17;{$ENDIF} 
 
-             vfmadd231pd xmm0, xmm1, xmm2;
+             {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
 
              //inc(PByte(valCounter1), LineWidth1);
              //inc(PByte(valCounter2), LineWidth2);
@@ -1024,11 +1024,11 @@ procedure FMAMtxMultLowTria2T2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt;
          @@foridxlongloop:
             {$IFDEF FPC}vmovupd ymm1, [rax + rdi - 64];{$ELSE}db $C5,$FD,$10,$4C,$38,$C0;{$ENDIF} 
             {$IFDEF FPC}vmovupd ymm2, [r8 + rdi - 64];{$ELSE}db $C4,$C1,$7D,$10,$54,$38,$C0;{$ENDIF} 
-            vfmadd231pd ymm0, ymm1, ymm2;
+            {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
 
             {$IFDEF FPC}vmovupd ymm1, [rax + rdi - 32];{$ELSE}db $C5,$FD,$10,$4C,$38,$E0;{$ENDIF} 
             {$IFDEF FPC}vmovupd ymm2, [r8 + rdi - 32];{$ELSE}db $C4,$C1,$7D,$10,$54,$38,$E0;{$ENDIF} 
-            vfmadd231pd ymm0, ymm1, ymm2;
+            {$IFDEF FPC}vfmadd231pd ymm0, ymm1, ymm2;{$ELSE}db $C4,$E2,$F5,$B8,$C2;{$ENDIF} 
          add rdi, 64;
          jl @@foridxlongloop;
 
@@ -1042,7 +1042,7 @@ procedure FMAMtxMultLowTria2T2Store1(mt1 : PDouble; LineWidth1 : TASMNativeInt;
          @@foridxSSEloop:
             {$IFDEF FPC}vmovupd xmm1, [rax + rdi - 16];{$ELSE}db $C5,$F9,$10,$4C,$38,$F0;{$ENDIF} 
             {$IFDEF FPC}vmovupd xmm2, [r8 + rdi - 16];{$ELSE}db $C4,$C1,$79,$10,$54,$38,$F0;{$ENDIF} 
-            vfmadd231pd xmm0, xmm1, xmm2;
+            {$IFDEF FPC}vfmadd231pd xmm0, xmm1, xmm2;{$ELSE}db $C4,$E2,$F1,$B8,$C2;{$ENDIF} 
          add rdi, 16;
          jl @@foridxSSEloop;