Some experiments with asmjityouki/jit-experiments

7 files changed, 94 insertions, 94 deletions

diff --git a/src/GenerateKernelU8S8S32ACC16.cc b/src/GenerateKernelU8S8S32ACC16.cc
index 205af14..6fbe1fb 100644
--- a/src/GenerateKernelU8S8S32ACC16.cc
+++ b/src/GenerateKernelU8S8S32ACC16.cc
@@ -105,11 +105,11 @@ void CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::storeCRegs<
             extractDest128, CRegs(i * leadingDimCReg + j), idx);
         a->vpmovsxwd(extractDest256, extractDest128);
         x86::Mem destAddr = x86::dword_ptr(
-#ifdef _MSC_VER
-            a->gpz(9), C_Offset, 0, (j * 2 + idx) * 8 * sizeof(int32_t));
-#else
+//#ifdef _MSC_VER
+            //a->gpz(9), C_Offset, 0, (j * 2 + idx) * 8 * sizeof(int32_t));
+//#else
             a->zcx(), C_Offset, 0, (j * 2 + idx) * 8 * sizeof(int32_t));
-#endif
+//#endif
         if (accum) {
           a->vpaddd(extractDest256, extractDest256, destAddr);
         }
@@ -195,21 +195,21 @@ CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx2>(
     //"nc must be equal to the number of register blocks");
 
     // arguments to the function created
-#ifdef _MSC_VER
-    x86::Gp buffer_A = a->zcx();
-    x86::Gp buffer_B = a->zdx();
-    x86::Gp B_pf = a->gpz(8);
-    x86::Gp CBase = a->gpz(9);
-    x86::Gp kSize = a->zdi();
-    x86::Gp ldcReg = a->zsi();
-#else
+//#ifdef _MSC_VER
+//    x86::Gp buffer_A = a->zcx();
+//    x86::Gp buffer_B = a->zdx();
+//    x86::Gp B_pf = a->gpz(8);
+//    x86::Gp CBase = a->gpz(9);
+//    x86::Gp kSize = a->zdi();
+//    x86::Gp ldcReg = a->zsi();
+//#else
     x86::Gp buffer_A = a->zdi();
     x86::Gp buffer_B = a->zsi();
     x86::Gp B_pf = a->zdx();
     x86::Gp CBase = a->zcx();
     x86::Gp kSize = a->gpz(8);
     x86::Gp ldcReg = a->gpz(9);
-#endif
+//#endif
 
     asmjit::FuncDetail func;
     func.init(
diff --git a/src/GenerateKernelU8S8S32ACC16Avx512.cc b/src/GenerateKernelU8S8S32ACC16Avx512.cc
index 819f33b..c3a7b7b 100644
--- a/src/GenerateKernelU8S8S32ACC16Avx512.cc
+++ b/src/GenerateKernelU8S8S32ACC16Avx512.cc
@@ -112,11 +112,11 @@ void CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::storeCRegs<
             extractDest256, CRegs(i * leadingDimCReg + j), idx);
         a->vpmovsxwd(extractDest512, extractDest256);
         x86::Mem destAddr = x86::dword_ptr(
-#ifdef _MSC_VER
-            a->gpz(9), C_Offset, 0, (j * 2 + idx) * 16 * sizeof(int32_t));
-#else
+//#ifdef _MSC_VER
+//            a->gpz(9), C_Offset, 0, (j * 2 + idx) * 16 * sizeof(int32_t));
+//#else
             a->zcx(), C_Offset, 0, (j * 2 + idx) * 16 * sizeof(int32_t));
-#endif
+//#endif
         if (accum) {
           a->vpaddd(extractDest512, extractDest512, destAddr);
         }
@@ -207,21 +207,21 @@ CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx512>(
     int mRegBlocksRem = mc % mRegBlockSize;
 
     // arguments to the function created
-#ifdef _MSC_VER
-    x86::Gp buffer_A = a->zcx();
-    x86::Gp buffer_B = a->zdx();
-    x86::Gp B_pf = a->gpz(8);
-    x86::Gp CBase = a->gpz(9);
-    x86::Gp kSize = a->zdi();
-    x86::Gp ldcReg = a->zsi();
-#else
+//#ifdef _MSC_VER
+//    x86::Gp buffer_A = a->zcx();
+//    x86::Gp buffer_B = a->zdx();
+//    x86::Gp B_pf = a->gpz(8);
+//    x86::Gp CBase = a->gpz(9);
+//    x86::Gp kSize = a->zdi();
+//    x86::Gp ldcReg = a->zsi();
+//#else
     x86::Gp buffer_A = a->zdi();
     x86::Gp buffer_B = a->zsi();
     x86::Gp B_pf = a->zdx();
     x86::Gp CBase = a->zcx();
     x86::Gp kSize = a->gpz(8);
     x86::Gp ldcReg = a->gpz(9);
-#endif
+//#endif
 
     asmjit::FuncDetail func;
     func.init(
diff --git a/src/GenerateKernelU8S8S32ACC32.cc b/src/GenerateKernelU8S8S32ACC32.cc
index dc9c534..977f635 100644
--- a/src/GenerateKernelU8S8S32ACC32.cc
+++ b/src/GenerateKernelU8S8S32ACC32.cc
@@ -107,18 +107,18 @@ void CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::storeCRegs<
         a->vpaddd(
             CRegs(i * leadingDimCReg + j),
             CRegs(i * leadingDimCReg + j),
-#ifdef _MSC_VER
-            x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 8 * sizeof(int32_t)));
-#else
+//#ifdef _MSC_VER
+//            x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 8 * sizeof(int32_t)));
+//#else
             x86::dword_ptr(a->zcx(), C_Offset, 0, j * 8 * sizeof(int32_t)));
-#endif
+//#endif
       }
       a->vmovups(
-#ifdef _MSC_VER
-          x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 8 * sizeof(int32_t)),
-#else
+//#ifdef _MSC_VER
+//          x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 8 * sizeof(int32_t)),
+//#else
           x86::dword_ptr(a->zcx(), C_Offset, 0, j * 8 * sizeof(int32_t)),
-#endif
+//#endif
           CRegs(i * leadingDimCReg + j));
     }
   }
@@ -196,21 +196,21 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<inst_set_t::avx2>(
     int mRegBlocksRem = mc % mRegBlockSize;
 
     // arguments to the function created
-#ifdef _MSC_VER
-    x86::Gp buffer_A = a->zcx();
-    x86::Gp buffer_B = a->zdx();
-    x86::Gp B_pf = a->gpz(8);
-    x86::Gp CBase = a->gpz(9);
-    x86::Gp kSize = a->zdi();
-    x86::Gp ldcReg = a->zsi();
-#else
+//#ifdef _MSC_VER
+//    x86::Gp buffer_A = a->zcx();
+//    x86::Gp buffer_B = a->zdx();
+//    x86::Gp B_pf = a->gpz(8);
+//    x86::Gp CBase = a->gpz(9);
+//    x86::Gp kSize = a->zdi();
+//    x86::Gp ldcReg = a->zsi();
+//#else
     x86::Gp buffer_A = a->zdi();
     x86::Gp buffer_B = a->zsi();
     x86::Gp B_pf = a->zdx();
     x86::Gp CBase = a->zcx();
     x86::Gp kSize = a->gpz(8);
     x86::Gp ldcReg = a->gpz(9);
-#endif
+//#endif
 
     asmjit::FuncDetail func;
     func.init(
diff --git a/src/GenerateKernelU8S8S32ACC32Avx512.cc b/src/GenerateKernelU8S8S32ACC32Avx512.cc
index 5037292..15b6fd4 100644
--- a/src/GenerateKernelU8S8S32ACC32Avx512.cc
+++ b/src/GenerateKernelU8S8S32ACC32Avx512.cc
@@ -108,18 +108,18 @@ void CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::storeCRegs<
         a->vpaddd(
             CRegs(i * leadingDimCReg + j),
             CRegs(i * leadingDimCReg + j),
-#ifdef _MSC_VER
-            x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)));
-#else
+//#ifdef _MSC_VER
+//            x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)));
+//#else
             x86::dword_ptr(a->zcx(), C_Offset, 0, j * 16 * sizeof(int32_t)));
-#endif
+//#endif
       }
       a->vmovups(
-#ifdef _MSC_VER
-          x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)),
-#else
+//#ifdef _MSC_VER
+//          x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)),
+//#else
           x86::dword_ptr(a->zcx(), C_Offset, 0, j * 16 * sizeof(int32_t)),
-#endif
+//#endif
           CRegs(i * leadingDimCReg + j));
     }
   }
@@ -204,21 +204,21 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<inst_set_t::avx512>(
     int mRegBlocksRem = mc % mRegBlockSize;
 
     // arguments to the function created
-#ifdef _MSC_VER
-    x86::Gp buffer_A = a->zcx();
-    x86::Gp buffer_B = a->zdx();
-    x86::Gp B_pf = a->gpz(8);
-    x86::Gp CBase = a->gpz(9);
-    x86::Gp kSize = a->zdi();
-    x86::Gp ldcReg = a->zsi();
-#else
+//#ifdef _MSC_VER
+//    x86::Gp buffer_A = a->zcx();
+//    x86::Gp buffer_B = a->zdx();
+//    x86::Gp B_pf = a->gpz(8);
+//    x86::Gp CBase = a->gpz(9);
+//    x86::Gp kSize = a->zdi();
+//    x86::Gp ldcReg = a->zsi();
+//#else
     x86::Gp buffer_A = a->zdi();
     x86::Gp buffer_B = a->zsi();
     x86::Gp B_pf = a->zdx();
     x86::Gp CBase = a->zcx();
     x86::Gp kSize = a->gpz(8);
     x86::Gp ldcReg = a->gpz(9);
-#endif
+//#endif
 
     asmjit::FuncDetail func;
     func.init(
diff --git a/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc b/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc
index bd8be1f..102083f 100644
--- a/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc
+++ b/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc
@@ -98,18 +98,18 @@ void CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::storeCRegs<
         a->vpaddd(
             CRegs(i * leadingDimCReg + j),
             CRegs(i * leadingDimCReg + j),
-#ifdef _MSC_VER
-            x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)));
-#else
+//#ifdef _MSC_VER
+//            x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)));
+//#else
             x86::dword_ptr(a->zcx(), C_Offset, 0, j * 16 * sizeof(int32_t)));
-#endif
+//#endif
       }
       a->vmovups(
-#ifdef _MSC_VER
-          x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)),
-#else
+//#ifdef _MSC_VER
+//          x86::dword_ptr(a->gpz(9), C_Offset, 0, j * 16 * sizeof(int32_t)),
+//#else
           x86::dword_ptr(a->zcx(), C_Offset, 0, j * 16 * sizeof(int32_t)),
-#endif
+//#endif
           CRegs(i * leadingDimCReg + j));
     }
   }
@@ -198,21 +198,21 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<
     int mRegBlocksRem = mc % mRegBlockSize;
 
     // arguments to the function created
-#ifdef _MSC_VER
-    x86::Gp buffer_A = a->zcx();
-    x86::Gp buffer_B = a->zdx();
-    x86::Gp B_pf = a->gpz(8);
-    x86::Gp CBase = a->gpz(9);
-    x86::Gp kSize = a->zdi();
-    x86::Gp ldcReg = a->zsi();
-#else
+//#ifdef _MSC_VER
+//    x86::Gp buffer_A = a->zcx();
+//    x86::Gp buffer_B = a->zdx();
+//    x86::Gp B_pf = a->gpz(8);
+//    x86::Gp CBase = a->gpz(9);
+//    x86::Gp kSize = a->zdi();
+//    x86::Gp ldcReg = a->zsi();
+//#else
     x86::Gp buffer_A = a->zdi();
     x86::Gp buffer_B = a->zsi();
     x86::Gp B_pf = a->zdx();
     x86::Gp CBase = a->zcx();
     x86::Gp kSize = a->gpz(8);
     x86::Gp ldcReg = a->gpz(9);
-#endif
+//#endif
 
     asmjit::FuncDetail func;
     func.init(
diff --git a/src/GroupwiseConvAcc32Avx2.cc b/src/GroupwiseConvAcc32Avx2.cc
index 396e792..dd309c3 100644
--- a/src/GroupwiseConvAcc32Avx2.cc
+++ b/src/GroupwiseConvAcc32Avx2.cc
@@ -1010,21 +1010,21 @@ jit_conv_kernel_fp GenConvKernel<2, int32_t>::getOrCreate<inst_set_t::avx2>(
 #endif
 
   // arguments to the function created
-#ifdef _MSC_VER
-  in_acts_R_ = a->zcx();
-  wghts_R_ = a->zdx();
-  out_acts_R_ = a->gpz(8);
-  a_zero_pt_R_ = a->gpz(9);
-  H_R_ = a->zdi();
-  W_R_ = a->zsi();
-#else
+//#ifdef _MSC_VER
+//  in_acts_R_ = a->zcx();
+//  wghts_R_ = a->zdx();
+//  out_acts_R_ = a->gpz(8);
+//  a_zero_pt_R_ = a->gpz(9);
+//  H_R_ = a->zdi();
+//  W_R_ = a->zsi();
+//#else
   in_acts_R_ = a->zdi();
   wghts_R_ = a->zsi();
   out_acts_R_ = a->zdx();
   a_zero_pt_R_ = a->zcx();
   H_R_ = a->gpz(8);
   W_R_ = a->gpz(9);
-#endif
+//#endif
   row_offset_R_ = a->gpz(10);
 
   // register for temporary use
@@ -1501,19 +1501,19 @@ GenConvKernel<2, int32_t>::getOrCreateRowOffset<inst_set_t::avx2>(
 #endif
 
   // arguments to the function created
-#ifdef _MSC_VER
-  in_acts_R_ = a->zcx();
-  a_zero_pt_R_ = a->zdx();
-  H_R_ = a->gpz(8);
-  W_R_ = a->gpz(9);
-  row_offset_R_ = a->zdi();
-#else
+//#ifdef _MSC_VER
+//  in_acts_R_ = a->zcx();
+//  a_zero_pt_R_ = a->zdx();
+//  H_R_ = a->gpz(8);
+//  W_R_ = a->gpz(9);
+//  row_offset_R_ = a->zdi();
+//#else
   in_acts_R_ = a->zdi();
   a_zero_pt_R_ = a->zsi();
   H_R_ = a->zdx();
   W_R_ = a->zcx();
   row_offset_R_ = a->gpz(8);
-#endif
+//#endif
 
   // register for temporary use
   scratchReg1_ = a->gpz(12);
diff --git a/third_party/asmjit b/third_party/asmjit
-Subproject 4da474ac9aa2689e88d5e40a2f37628f302d7e3
+Subproject ac77dfcd75f043e2fe317133a971040e5b99991