/// <summary>
/// Create native code of cpuid checking function and wrap it into <see cref="NativeMethodHandle{TDelegate}"/>.
/// </summary>
/// <returns>cpuid support checking method cache.</returns>
/// <exception cref="PlatformNotSupportedException">Throw when processor architecture is not x64 nor x86.</exception>
public static NativeMethodHandle <IsCpuIdSupportedDelegate> CreateIsCpuIdSupportedMethodHandle()
{
if (!IsSupportedArchitecture())
{
ThrowPlatformNotSupportedException("cpuid support checking function is not supported on this architecture");
}
return(NativeMethodHandle.Create <IsCpuIdSupportedDelegate>(Environment.Is64BitProcess
// bool __stdcall isCpuIdSupported(void); // The return value type may be 64bit value such as long long int
? new byte[]
{
0x9c, // pushfq
0x9c, // pushfq
0x58, // pop rax
0x48, 0x89, 0xc1, // mov rcx,rax
0x48, 0x35, 0x00, 0x00, 0x20, 0x00, // xor rax,200000h
0x50, // push rax
0x9d, // popfq
0x9c, // pushfq
0x58, // pop rax
0x48, 0x31, 0xc8, // xor rax,rcx
0x48, 0xc1, 0xe8, 0x15, // shr rax,21
0x9d, // popfq
0xc3 // ret
}
// bool __stdcall isCpuIdSupported(void); // The return value type may be 32bit value such as int
: new byte[]
{
0x9c, // pushf
0x9c, // pushf
0x58, // pop eax
0x89, 0xc1, // mov ecx,eax
0x35, 0x00, 0x00, 0x20, 0x00, // xor eax,200000h
0x50, // push eax
0x9d, // popf
0x9c, // pushf
0x58, // pop eax
0x31, 0xc8, // xor eax,ecx
0xc1, 0xe8, 0x15, // shr eax,21
0x9d, // popf
0xc3, // ret
}));
}
/// <summary>
/// Create native code of cpuid and wrap it into <see cref="NativeMethodHandle{TDelegate}"/>.
/// </summary>
/// <typeparam name="TDelegate">Delegate for native method of cpuid.</typeparam>
/// <returns>Method handle of cpuid.</returns>
/// <exception cref="PlatformNotSupportedException">Throw when processor architecture is not x64 nor x86.</exception>
private static NativeMethodHandle <TDelegate> CreateCpuIdMethodHandle <TDelegate>()
where TDelegate : Delegate
{
if (!IsSupportedArchitecture())
{
ThrowPlatformNotSupportedException("cpuid is not supported on this architecture");
}
return(NativeMethodHandle.Create <TDelegate>(Environment.Is64BitProcess
// void __stdcall cpuid(int* cpuInfo, int eax, int ecx);
? new byte[]
{
0x53, // push rbx
0x49, 0x89, 0xc9, // mov r9,rcx
0x89, 0xd0, // mov eax,edx
0x44, 0x89, 0xc1, // mov ecx,r8d
0x0f, 0xa2, // cpuid
0x41, 0x89, 0x01, // mov dword ptr [r9],eax
0x41, 0x89, 0x59, 0x04, // mov dword ptr [r9 + 04h],ebx
0x41, 0x89, 0x49, 0x08, // mov dword ptr [r9 + 08h],ecx
0x41, 0x89, 0x51, 0x0c, // mov dword ptr [r9 + 0ch],edx
0x5b, // pop rbx
0xc3 // ret
}
// void __stdcall cpuid(int* cpuInfo, int eax, int ecx);
: new byte[]
{
0x56, // push esi
0x53, // push ebx
0x8b, 0x74, 0x24, 0x0c, // mov esi,dword ptr [esp + 0Ch]
0x8b, 0x44, 0x24, 0x10, // mov eax,dword ptr [esp + 10h]
0x8b, 0x4c, 0x24, 0x14, // mov ecx,dword ptr [esp + 14h]
0x0f, 0xa2, // cpuid
0x89, 0x06, // mov dword ptr [esi],eax
0x89, 0x5e, 0x04, // mov dword ptr [esi + 04h],ebx
0x89, 0x4e, 0x08, // mov dword ptr [esi + 08h],ecx
0x89, 0x56, 0x0c, // mov dword ptr [esi + 0Ch],edx
0x5b, // pop ebx
0x5e, // pop esi
0xc2, 0x0c, 0x00 // ret 0Ch
}));
}
/// <summary>
/// Create native method handle of memory compare function using AVX2.
/// </summary>
/// <returns>Created native method handle.</returns>
private static NativeMethodHandle <CompareMemoryDelegate> CreateCompareMemoryAvx2MethodHandle()
{
return(NativeMethodHandle.Create <CompareMemoryDelegate>(Environment.Is64BitProcess ? new byte[]
{
0x4d, 0x89, 0xc1, // mov r9,r8
0x49, 0x83, 0xe1, 0xe0, // and r9,0xffffffffffffffe0
0x74, 0x43, // je L4
0x45, 0x31, 0xd2, // xor r10d,r10d
0x31, 0xc0, // xor eax,eax
0xeb, 0x0c, // jmp L2
// L1:
0x41, 0x83, 0xc2, 0x20, // add r10d,0x20
0x49, 0x63, 0xc2, // movsxd rax,r10d
0x4c, 0x39, 0xc8, // cmp rax,r9
0x73, 0x2d, // jae L3
// L2:
0xc5, 0xfa, 0x6f, 0x14, 0x02, // vmovdqu xmm2,XMMWORD PTR [rdx+rax*1]
0xc5, 0xfa, 0x6f, 0x1c, 0x01, // vmovdqu xmm3,XMMWORD PTR [rcx+rax*1]
0xc4, 0xe3, 0x6d, 0x38, 0x44, 0x02, 0x10, // vinserti128 ymm0,ymm2,XMMWORD PTR [rdx+rax*1+0x10],0x1
0x01,
0xc4, 0xe3, 0x65, 0x38, 0x4c, 0x01, 0x10, // vinserti128 ymm1,ymm3,XMMWORD PTR [rcx+rax*1+0x10],0x1
0x01,
0xc5, 0xfd, 0x74, 0xc1, // vpcmpeqb ymm0,ymm0,ymm1
0xc5, 0xfd, 0xd7, 0xc0, // vpmovmskb eax,ymm0
0x83, 0xf8, 0xff, // cmp eax,0xffffffff
0x74, 0xcd, // je L1
0x31, 0xc0, // xor eax,eax
0xc5, 0xf8, 0x77, // vzeroupper
0xc3, // ret
// L3:
0xc5, 0xf8, 0x77, // vzeroupper
// L4:
0x49, 0x63, 0xc1, // movsxd rax,r9d
0x49, 0x39, 0xc0, // cmp r8,rax
0x77, 0x0f, // ja L7
// L5:
0xb8, 0x01, 0x00, 0x00, 0x00, // mov eax,0x1
0xc3, // ret
// L6:
0x48, 0x83, 0xc0, 0x01, // add rax,0x1
0x49, 0x39, 0xc0, // cmp r8,rax
0x76, 0xf1, // jbe L5
// L7:
0x44, 0x0f, 0xb6, 0x1c, 0x02, // movzx r11d,BYTE PTR [rdx+rax*1]
0x44, 0x38, 0x1c, 0x01, // cmp BYTE PTR [rcx+rax*1],r11b
0x74, 0xec, // je L6
0x31, 0xc0, // xor eax,eax
0xc3 // ret
} : new byte[]
{
0x55, // push ebp
0x89, 0xe5, // mov ebp,esp
0x57, // push edi
0x56, // push esi
0x8b, 0x75, 0x10, // mov esi,DWORD PTR [ebp+0x10]
0x53, // push ebx
0x8b, 0x4d, 0x08, // mov ecx,DWORD PTR [ebp+0x8]
0x89, 0xf7, // mov edi,esi
0x8b, 0x5d, 0x0c, // mov ebx,DWORD PTR [ebp+0xc]
0x83, 0xe7, 0xe0, // and edi,0xffffffe0
0x74, 0x3f, // je L4
0x31, 0xd2, // xor edx,edx
0xeb, 0x07, // jmp L2
// L1:
0x83, 0xc2, 0x20, // add edx,0x20
0x39, 0xd7, // cmp edi,edx
0x76, 0x31, // jbe L3
// L2:
0xc5, 0xfa, 0x6f, 0x14, 0x13, // vmovdqu xmm2,XMMWORD PTR [ebx+edx*1]
0xc5, 0xfa, 0x6f, 0x1c, 0x11, // vmovdqu xmm3,XMMWORD PTR [ecx+edx*1]
0xc4, 0xe3, 0x6d, 0x38, 0x44, 0x13, 0x10, // vinserti128 ymm0,ymm2,XMMWORD PTR [ebx+edx*1+0x10],0x1
0x01,
0xc4, 0xe3, 0x65, 0x38, 0x4c, 0x11, 0x10, // vinserti128 ymm1,ymm3,XMMWORD PTR [ecx+edx*1+0x10],0x1
0x01,
0xc5, 0xfd, 0x74, 0xc1, // vpcmpeqb ymm0,ymm0,ymm1
0xc5, 0xfd, 0xd7, 0xc0, // vpmovmskb eax,ymm0
0x83, 0xf8, 0xff, // cmp eax,0xffffffff
0x74, 0xd2, // je L1
0x31, 0xc0, // xor eax,eax
0xc5, 0xf8, 0x77, // vzeroupper
0x5b, // pop ebx
0x5e, // pop esi
0x5f, // pop edi
0x5d, // pop ebp
0xc3, // ret
// L3:
0xc5, 0xf8, 0x77, // vzeroupper
// L4:
0x89, 0xf8, // mov eax,edi
0x39, 0xfe, // cmp esi,edi
0x77, 0x11, // ja L7
// L5:
0x5b, // pop ebx
0xb8, 0x01, 0x00, 0x00, 0x00, // mov eax,0x1
0x5e, // pop esi
0x5f, // pop edi
0x5d, // pop ebp
0xc3, // ret
// L6:
0x83, 0xc0, 0x01, // add eax,0x1
0x39, 0xc6, // cmp esi,eax
0x76, 0xef, // jbe L5
// L7:
0x0f, 0xb6, 0x14, 0x03, // movzx edx,BYTE PTR [ebx+eax*1]
0x38, 0x14, 0x01, // cmp BYTE PTR [ecx+eax*1],dl
0x74, 0xf0, // je L6
0x5b, // pop ebx
0x31, 0xc0, // xor eax,eax
0x5e, // pop esi
0x5f, // pop edi
0x5d, // pop ebp
0xc3 // ret
}));
}
/// <summary>
/// Create native code method that calculate inner product of two <see cref="float"/> arrays.
/// </summary>
/// <returns>Native code method of inner product calculation</returns>
private NativeMethodHandle <InnerProductDelegate> CreateInnerProductMethodHandle()
{
// void __stdcall
// inner_product(
// float* dst,
// const float* src1,
// const float* src2,
// int size) noexcept
// {
// constexpr auto stride = static_cast<int>(sizeof(__m128) / sizeof(float));
//
// for (int i = 0, im = size - stride; i <= im; i += stride) {
// _mm_storeu_ps(
// &dst[i],
// _mm_mul_ps(
// _mm_loadu_ps(&src1[i]),
// _mm_loadu_ps(&src2[i])));
// }
//
// const auto remSize = size % stride;
// if (remSize == 0) {
// return;
// }
//
// const auto offset = size - remSize;
// for (int i = offset; i < size; i++) {
// dst[i] = src1[i] * src2[i];
// }
// }
return(NativeMethodHandle.Create <InnerProductDelegate>(Environment.Is64BitProcess ? new byte[]
{
0x41, 0x83, 0xf9, 0x03, // cmp r9d,0x3
0x7e, 0x33, // jle L2
0x45, 0x8d, 0x51, 0xfc, // lea r10d,[r9-0x4]
0x31, 0xc0, // xor eax,eax
0x41, 0xc1, 0xea, 0x02, // shr r10d,0x2
0x49, 0x83, 0xc2, 0x01, // add r10,0x1
0x49, 0xc1, 0xe2, 0x04, // shl r10,0x4
0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, // nop DWORD PTR [rax+rax*1+0x0]
0x00,
// L1:
0x41, 0x0f, 0x10, 0x04, 0x00, // movups xmm0,XMMWORD PTR [r8+rax*1]
0x0f, 0x10, 0x0c, 0x02, // movups xmm1,XMMWORD PTR [rdx+rax*1]
0x0f, 0x59, 0xc1, // mulps xmm0,xmm1
0x0f, 0x11, 0x04, 0x01, // movups XMMWORD PTR [rcx+rax*1],xmm0
0x48, 0x83, 0xc0, 0x10, // add rax,0x10
0x49, 0x39, 0xc2, // cmp r10,rax
0x75, 0xe7, // jne L1
// L2:
0x45, 0x89, 0xca, // mov r10d,r9d
0x41, 0xc1, 0xfa, 0x1f, // sar r10d,0x1f
0x41, 0xc1, 0xea, 0x1e, // shr r10d,0x1e
0x43, 0x8d, 0x04, 0x11, // lea eax,[r9+r10*1]
0x83, 0xe0, 0x03, // and eax,0x3
0x44, 0x29, 0xd0, // sub eax,r10d
0x74, 0x49, // je L4
0x45, 0x89, 0xca, // mov r10d,r9d
0x41, 0x29, 0xc2, // sub r10d,eax
0x45, 0x39, 0xd1, // cmp r9d,r10d
0x7e, 0x3e, // jle L4
0x41, 0x83, 0xe9, 0x01, // sub r9d,0x1
0x4d, 0x63, 0xda, // movsxd r11,r10d
0x45, 0x29, 0xd1, // sub r9d,r10d
0x4a, 0x8d, 0x04, 0x9d, 0x00, 0x00, 0x00, // lea rax,[r11*4+0x0]
0x00,
0x4f, 0x8d, 0x4c, 0x0b, 0x01, // lea r9,[r11+r9*1+0x1]
0x49, 0xc1, 0xe1, 0x02, // shl r9,0x2
0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, // nop WORD PTR cs:[rax+rax*1+0x0]
0x00, 0x00, 0x00,
// L3:
0xf3, 0x0f, 0x10, 0x04, 0x02, // movss xmm0,DWORD PTR [rdx+rax*1]
0xf3, 0x41, 0x0f, 0x59, 0x04, 0x00, // mulss xmm0,DWORD PTR [r8+rax*1]
0xf3, 0x0f, 0x11, 0x04, 0x01, // movss DWORD PTR [rcx+rax*1],xmm0
0x48, 0x83, 0xc0, 0x04, // add rax,0x4
0x4c, 0x39, 0xc8, // cmp rax,r9
0x75, 0xe7, // jne L3
// L4:
0xc3, // ret
} : new byte[]
{
0x55, // push ebp
0x89, 0xe5, // mov ebp,esp
0x57, // push edi
0x56, // push esi
0x8b, 0x75, 0x14, // mov esi,DWORD PTR [ebp+0x14]
0x53, // push ebx
0x8b, 0x55, 0x08, // mov edx,DWORD PTR [ebp+0x8]
0x8b, 0x5d, 0x0c, // mov ebx,DWORD PTR [ebp+0xc]
0x83, 0xe4, 0xf0, // and esp,0xfffffff0
0x8b, 0x4d, 0x10, // mov ecx,DWORD PTR [ebp+0x10]
0x83, 0xfe, 0x03, // cmp esi,0x3
0x7e, 0x21, // jle L2
0x8d, 0x7e, 0xfc, // lea edi,[esi-0x4]
0x31, 0xc0, // xor eax,eax
0x83, 0xe7, 0xfc, // and edi,0xfffffffc
0x83, 0xc7, 0x04, // add edi,0x4
// L1:
0x0f, 0x10, 0x04, 0x81, // movups xmm0,XMMWORD PTR [ecx+eax*4]
0x0f, 0x10, 0x0c, 0x83, // movups xmm1,XMMWORD PTR [ebx+eax*4]
0x0f, 0x59, 0xc1, // mulps xmm0,xmm1
0x0f, 0x11, 0x04, 0x82, // movups XMMWORD PTR [edx+eax*4],xmm0
0x83, 0xc0, 0x04, // add eax,0x4
0x39, 0xf8, // cmp eax,edi
0x75, 0xea, // jne L1
// L2:
0x89, 0xf7, // mov edi,esi
0xc1, 0xff, 0x1f, // sar edi,0x1f
0xc1, 0xef, 0x1e, // shr edi,0x1e
0x8d, 0x04, 0x3e, // lea eax,[esi+edi*1]
0x83, 0xe0, 0x03, // and eax,0x3
0x29, 0xf8, // sub eax,edi
0x74, 0x2a, // je L4
0x89, 0xf7, // mov edi,esi
0x29, 0xc7, // sub edi,eax
0x39, 0xfe, // cmp esi,edi
0x7e, 0x22, // jle L4
0xc1, 0xe7, 0x02, // shl edi,0x2
0x8d, 0x04, 0x3b, // lea eax,[ebx+edi*1]
0x8d, 0x1c, 0xb3, // lea ebx,[ebx+esi*4]
0x01, 0xf9, // add ecx,edi
0x01, 0xfa, // add edx,edi
// L3:
0xd9, 0x00, // fld DWORD PTR [eax]
0x83, 0xc0, 0x04, // add eax,0x4
0x83, 0xc1, 0x04, // add ecx,0x4
0x83, 0xc2, 0x04, // add edx,0x4
0xd8, 0x49, 0xfc, // fmul DWORD PTR [ecx-0x4]
0xd9, 0x5a, 0xfc, // fstp DWORD PTR [edx-0x4]
0x39, 0xd8, // cmp eax,ebx
0x75, 0xeb, // jne L3
// L4:
0x8d, 0x65, 0xf4, // lea esp,[ebp-0xc]
0x5b, // pop ebx
0x5e, // pop esi
0x5f, // pop edi
0x5d, // pop ebp
0xc2, 0x10, 0x00, // ret 0x10
}));
}