; Listing generated by Microsoft (R) Optimizing Compiler Version 19.37.32824.0 include listing.inc INCLUDELIB MSVCRT INCLUDELIB OLDNAMES _DATA SEGMENT COMM report:QWORD COMM infolog:QWORD:0200H _DATA ENDS msvcjmc SEGMENT __0ED14FEE_winpackagefamily@h DB 01H __F92EF742_winapifamily@h DB 01H __22E515EB_sdkddkver@h DB 01H __9C03CD75_concurrencysal@h DB 01H __CA64BAF1_sal@h DB 01H __95655E73_vadefs@h DB 01H __1D560A45_vcruntime@h DB 01H __F214EFBF_excpt@h DB 01H __D9B78EDA_stdarg@h DB 01H __29F49F40_specstrings_undef@h DB 01H __D0F42932_specstrings_strict@h DB 01H __F00405FD_sdv_driverspecs@h DB 01H __173E09CA_driverspecs@h DB 01H __9A452361_specstrings@h DB 01H __1B0272D9_corecrt@h DB 01H __74484FB3_corecrt_wctype@h DB 01H __1EC467ED_ctype@h DB 01H __98517906_kernelspecs@h DB 01H __D3497022_basetsd@h DB 01H __61CC25A1_errno@h DB 01H __7061C667_vcruntime_string@h DB 01H __E4C945B4_corecrt_memcpy_s@h DB 01H __8EAFF484_corecrt_memory@h DB 01H __6490AEC5_corecrt_wstring@h DB 01H __E8295E1E_string@h DB 01H __7FCB5136_guiddef@h DB 01H __FF8A9E11_pshpack4@h DB 01H __D5438902_poppack@h DB 01H __FB07E2A3_pshpack2@h DB 01H __F6906775_pshpack8@h DB 01H __F9415CFA_pshpack1@h DB 01H __E58C6F1C_apiset@h DB 01H __A5C52262_ktmtypes@h DB 01H __FBBA8F9D_winnt@h DB 01H __A32D3D2C_minwindef@h DB 01H __7D798299_windef@h DB 01H __9AE108E9_apisetcconv@h DB 01H __9CE53D2C_minwinbase@h DB 01H __5D0AA210_apiquery2@h DB 01H __4DD4E726_processenv@h DB 01H __3575CC96_fileapi@h DB 01H __8EA220E6_fileapifromapp@h DB 01H __BF9B31FE_debugapi@h DB 01H __C3AFE134_utilapiset@h DB 01H __76F69E51_handleapi@h DB 01H __7E63D784_errhandlingapi@h DB 01H __FFA7003C_fibersapi@h DB 01H __8971CECF_namedpipeapi@h DB 01H __46C9D8A6_profileapi@h DB 01H __98D4F0A3_heapapi@h DB 01H __50EF2A88_ioapiset@h DB 01H __92EA1464_synchapi@h DB 01H __A53D8AED_interlockedapi@h DB 01H __B506CAAF_processthreadsapi@h DB 01H __FF735694_sysinfoapi@h DB 01H __29589789_memoryapi@h DB 01H __E9C7A663_enclaveapi@h DB 01H __D035F9F7_threadpoollegacyapiset@h DB 01H __C899130C_threadpoolapiset@h DB 01H __E0441EE7_jobapi@h DB 01H __CEF86838_jobapi2@h DB 01H __35AC6B9A_wow64apiset@h DB 01H __C8AE41A5_libloaderapi@h DB 01H __CC9E232D_securitybaseapi@h DB 01H __C9D0C9B8_namespaceapi@h DB 01H __30FF9CAD_systemtopologyapi@h DB 01H __4D358795_processtopologyapi@h DB 01H __40446D10_securityappcontainer@h DB 01H __3A952EDF_realtimeapiset@h DB 01H __F4865F4B_winerror@h DB 01H __53C1BD93_timezoneapi@h DB 01H __1FFBEDBF_winbase@h DB 01H __43D3FA34_wingdi@h DB 01H __D24A7391_tvout@h DB 01H __7713A6C8_winuser@h DB 01H __0B345F71_datetimeapi@h DB 01H __9AC6960C_winnls@h DB 01H __973A40A8_stringapiset@h DB 01H __CC46FEE2_wincontypes@h DB 01H __97D70BE1_consoleapi@h DB 01H __27EC5E18_consoleapi2@h DB 01H __262E342F_consoleapi3@h DB 01H __641DCD70_wincon@h DB 01H __5A96C7A7_verrsrc@h DB 01H __B69C6FF2_winver@h DB 01H __7AEB8C37_reason@h DB 01H __59F1A8A9_winreg@h DB 01H __EE0E2B7F_wnnc@h DB 01H __DCDB7054_winnetwk@h DB 01H __81DDEDE3_stralign@h DB 01H __2134D1B4_winsvc@h DB 01H __11C07376_mcx@h DB 01H __577CE3E7_ime_cmodes@h DB 01H __1E72C31E_imm@h DB 01H __E15BF856_windows@h DB 01H __F7492337_framework@h DB 01H __ED683015_stdint@h DB 01H __018B5678_khrplatform@h DB 01H __79FC18F0_glad@h DB 01H __13F5961F_stddef@h DB 01H __986CAB3E_glfw3@h DB 01H __11E77DD3_mmintrin@h DB 01H __45DED006_corecrt_malloc@h DB 01H __FA77ED4C_malloc@h DB 01H __9D836611_xmmintrin@h DB 01H __F8E1EFDD_emmintrin@h DB 01H __27408754_pmmintrin@h DB 01H __9799F4D6_tmmintrin@h DB 01H __BEA2E155_smmintrin@h DB 01H __DBC06899_nmmintrin@h DB 01H __0E7B92D7_wmmintrin@h DB 01H __C5EFDFD0_zmmintrin@h DB 01H __F2FB7D1A_immintrin@h DB 01H __C508448F_StgSharpGraphic@h DB 01H __44420885_corecrt_stdio_config@h DB 01H __AB601D8D_corecrt_wstdio@h DB 01H __C1EC35D3_stdio@h DB 01H __68875EC1_mat3@c DB 01H msvcjmc ENDS PUBLIC transpose3to2 PUBLIC transpose3to3 PUBLIC transpose3to4 PUBLIC det_mat3 PUBLIC det_mat3_internal PUBLIC __JustMyCode_Default EXTRN __CheckForDebuggerJustMyCode:PROC EXTRN defaultVecPtr:QWORD EXTRN _fltused:DWORD ; COMDAT pdata pdata SEGMENT $pdata$transpose3to2 DD imagerel $LN3 DD imagerel $LN3+185 DD imagerel $unwind$transpose3to2 pdata ENDS ; COMDAT pdata pdata SEGMENT $pdata$transpose3to3 DD imagerel $LN3 DD imagerel $LN3+333 DD imagerel $unwind$transpose3to3 pdata ENDS ; COMDAT pdata pdata SEGMENT $pdata$transpose3to4 DD imagerel $LN3 DD imagerel $LN3+362 DD imagerel $unwind$transpose3to4 pdata ENDS ; COMDAT pdata pdata SEGMENT $pdata$det_mat3 DD imagerel $LN3 DD imagerel $LN3+125 DD imagerel $unwind$det_mat3 pdata ENDS ; COMDAT pdata pdata SEGMENT $pdata$det_mat3_internal DD imagerel $LN3 DD imagerel $LN3+261 DD imagerel $unwind$det_mat3_internal pdata ENDS ; COMDAT xdata xdata SEGMENT $unwind$det_mat3_internal DD 025041c01H DD 0117231cH DD 05010001aH xdata ENDS ; COMDAT xdata xdata SEGMENT $unwind$det_mat3 DD 025041201H DD 010d2312H DD 050060012H xdata ENDS ; COMDAT xdata xdata SEGMENT $unwind$transpose3to4 DD 025041701H DD 01122317H DD 0500b0014H xdata ENDS ; COMDAT xdata xdata SEGMENT $unwind$transpose3to3 DD 025041701H DD 01122317H DD 0500b0014H xdata ENDS ; COMDAT xdata xdata SEGMENT $unwind$transpose3to2 DD 025031701H DD 0f2122317H DD 0500bH xdata ENDS ; Function compile flags: /Odt ; COMDAT __JustMyCode_Default _TEXT SEGMENT __JustMyCode_Default PROC ; COMDAT 00000 c2 00 00 ret 0 __JustMyCode_Default ENDP _TEXT ENDS ; Function compile flags: /Odtp /ZI ; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\源代码\StgShqrpGraphic\mat3.c ; COMDAT det_mat3_internal _TEXT SEGMENT addc1$ = 0 addc2$ = 16 subc1$ = 32 subc2$ = 48 cache$ = 64 sub$ = 80 ret$ = 96 c0$ = 192 c1$ = 200 c2$ = 208 det_mat3_internal PROC ; COMDAT ; 40 : { $LN3: 00000 4c 89 44 24 18 mov QWORD PTR [rsp+24], r8 00005 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx 0000a 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx 0000f 55 push rbp 00010 48 81 ec d0 00 00 00 sub rsp, 208 ; 000000d0H 00017 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32] 0001c 48 8d 0d 00 00 00 00 lea rcx, OFFSET FLAT:__68875EC1_mat3@c 00023 e8 00 00 00 00 call __CheckForDebuggerJustMyCode ; 41 : ; 42 : //let adding row aligned ; 43 : __m128 addc1 = _mm_shuffle_ps(c1, c1, _MM_SHUFFLE(3, 0, 2, 1)); 00028 48 8b 85 c8 00 00 00 mov rax, QWORD PTR c1$[rbp] 0002f 48 8b 8d c8 00 00 00 mov rcx, QWORD PTR c1$[rbp] 00036 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00039 0f c6 01 c9 shufps xmm0, XMMWORD PTR [rcx], 201 ; 000000c9H 0003d 0f 29 45 00 movaps XMMWORD PTR addc1$[rbp], xmm0 ; 44 : __m128 addc2 = _mm_shuffle_ps(c2, c2, _MM_SHUFFLE(3, 1, 0, 2)); 00041 48 8b 85 d0 00 00 00 mov rax, QWORD PTR c2$[rbp] 00048 48 8b 8d d0 00 00 00 mov rcx, QWORD PTR c2$[rbp] 0004f 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00052 0f c6 01 d2 shufps xmm0, XMMWORD PTR [rcx], 210 ; 000000d2H 00056 0f 29 45 10 movaps XMMWORD PTR addc2$[rbp], xmm0 ; 45 : ; 46 : //let sub row aligned ; 47 : __m128 subc1 = _mm_shuffle_ps(c1, c1, _MM_SHUFFLE(3, 1, 0, 2)); 0005a 48 8b 85 c8 00 00 00 mov rax, QWORD PTR c1$[rbp] 00061 48 8b 8d c8 00 00 00 mov rcx, QWORD PTR c1$[rbp] 00068 0f 10 00 movups xmm0, XMMWORD PTR [rax] 0006b 0f c6 01 d2 shufps xmm0, XMMWORD PTR [rcx], 210 ; 000000d2H 0006f 0f 29 45 20 movaps XMMWORD PTR subc1$[rbp], xmm0 ; 48 : __m128 subc2 = _mm_shuffle_ps(c2, c2, _MM_SHUFFLE(3, 0, 2, 1)); 00073 48 8b 85 d0 00 00 00 mov rax, QWORD PTR c2$[rbp] 0007a 48 8b 8d d0 00 00 00 mov rcx, QWORD PTR c2$[rbp] 00081 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00084 0f c6 01 c9 shufps xmm0, XMMWORD PTR [rcx], 201 ; 000000c9H 00088 0f 29 45 30 movaps XMMWORD PTR subc2$[rbp], xmm0 ; 49 : ; 50 : //multple all add row ; 51 : __m128 cache = _mm_mul_ps(c0,addc1); 0008c 48 8b 85 c0 00 00 00 mov rax, QWORD PTR c0$[rbp] 00093 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00096 0f 59 45 00 mulps xmm0, XMMWORD PTR addc1$[rbp] 0009a 0f 29 45 40 movaps XMMWORD PTR cache$[rbp], xmm0 ; 52 : cache = _mm_mul_ps(cache,addc2); 0009e 0f 28 45 40 movaps xmm0, XMMWORD PTR cache$[rbp] 000a2 0f 59 45 10 mulps xmm0, XMMWORD PTR addc2$[rbp] 000a6 0f 29 45 40 movaps XMMWORD PTR cache$[rbp], xmm0 ; 53 : ; 54 : //multiple all sub row ; 55 : __m128 sub = _mm_mul_ps(c0, subc1); 000aa 48 8b 85 c0 00 00 00 mov rax, QWORD PTR c0$[rbp] 000b1 0f 10 00 movups xmm0, XMMWORD PTR [rax] 000b4 0f 59 45 20 mulps xmm0, XMMWORD PTR subc1$[rbp] 000b8 0f 29 45 50 movaps XMMWORD PTR sub$[rbp], xmm0 ; 56 : sub = _mm_mul_ps(sub, subc2); 000bc 0f 28 45 50 movaps xmm0, XMMWORD PTR sub$[rbp] 000c0 0f 59 45 30 mulps xmm0, XMMWORD PTR subc2$[rbp] 000c4 0f 29 45 50 movaps XMMWORD PTR sub$[rbp], xmm0 ; 57 : ; 58 : //main-corner sub sub-corner ; 59 : cache = _mm_sub_ps(cache,sub); 000c8 0f 28 45 40 movaps xmm0, XMMWORD PTR cache$[rbp] 000cc 0f 5c 45 50 subps xmm0, XMMWORD PTR sub$[rbp] 000d0 0f 29 45 40 movaps XMMWORD PTR cache$[rbp], xmm0 ; 60 : ; 61 : __m128 ret = _mm_hadd_ps(cache,cache); //element 1 add element 2 000d4 0f 28 45 40 movaps xmm0, XMMWORD PTR cache$[rbp] 000d8 f2 0f 7c 45 40 haddps xmm0, XMMWORD PTR cache$[rbp] 000dd 0f 29 45 60 movaps XMMWORD PTR ret$[rbp], xmm0 ; 62 : ret = _mm_add_ps(ret, cache); //(e1 add e2) add element 3 000e1 0f 28 45 60 movaps xmm0, XMMWORD PTR ret$[rbp] 000e5 0f 58 45 40 addps xmm0, XMMWORD PTR cache$[rbp] 000e9 0f 29 45 60 movaps XMMWORD PTR ret$[rbp], xmm0 ; 63 : ; 64 : ; 65 : return ret.m128_f32[2]; 000ed b8 04 00 00 00 mov eax, 4 000f2 48 6b c0 02 imul rax, rax, 2 000f6 f3 0f 10 44 05 60 movss xmm0, DWORD PTR ret$[rbp+rax] ; 66 : } 000fc 48 8d a5 b0 00 00 00 lea rsp, QWORD PTR [rbp+176] 00103 5d pop rbp 00104 c3 ret 0 det_mat3_internal ENDP _TEXT ENDS ; Function compile flags: /Odtp /ZI ; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\源代码\StgShqrpGraphic\mat3.c ; COMDAT det_mat3 _TEXT SEGMENT $T1 = 64 $T2 = 80 $T3 = 96 mat$ = 128 det_mat3 PROC ; COMDAT ; 69 : { $LN3: 00000 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx 00005 55 push rbp 00006 48 81 ec 90 00 00 00 sub rsp, 144 ; 00000090H 0000d 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32] 00012 48 8d 0d 00 00 00 00 lea rcx, OFFSET FLAT:__68875EC1_mat3@c 00019 e8 00 00 00 00 call __CheckForDebuggerJustMyCode ; 70 : return det_mat3_internal( 0001e b8 10 00 00 00 mov eax, 16 00023 48 6b c0 02 imul rax, rax, 2 00027 48 8b 8d 80 00 00 00 mov rcx, QWORD PTR mat$[rbp] 0002e 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax] 00032 0f 11 45 60 movups XMMWORD PTR $T3[rbp], xmm0 00036 b8 10 00 00 00 mov eax, 16 0003b 48 6b c0 01 imul rax, rax, 1 0003f 48 8b 8d 80 00 00 00 mov rcx, QWORD PTR mat$[rbp] 00046 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax] 0004a 0f 11 45 50 movups XMMWORD PTR $T2[rbp], xmm0 0004e b8 10 00 00 00 mov eax, 16 00053 48 6b c0 00 imul rax, rax, 0 00057 48 8b 8d 80 00 00 00 mov rcx, QWORD PTR mat$[rbp] 0005e 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax] 00062 0f 11 45 40 movups XMMWORD PTR $T1[rbp], xmm0 00066 4c 8d 45 60 lea r8, QWORD PTR $T3[rbp] 0006a 48 8d 55 50 lea rdx, QWORD PTR $T2[rbp] 0006e 48 8d 4d 40 lea rcx, QWORD PTR $T1[rbp] 00072 e8 00 00 00 00 call det_mat3_internal ; 71 : mat->colum[0], ; 72 : mat->colum[1], ; 73 : mat->colum[2] ; 74 : ); ; 75 : } 00077 48 8d 65 70 lea rsp, QWORD PTR [rbp+112] 0007b 5d pop rbp 0007c c3 ret 0 det_mat3 ENDP _TEXT ENDS ; Function compile flags: /Odtp /ZI ; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\源代码\StgShqrpGraphic\mat3.c ; COMDAT transpose3to4 _TEXT SEGMENT t0$ = 0 t1$ = 16 t2$ = 32 t3$ = 48 source$ = 144 target$ = 152 transpose3to4 PROC ; COMDAT ; 27 : { $LN3: 00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx 00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx 0000a 55 push rbp 0000b 48 81 ec a0 00 00 00 sub rsp, 160 ; 000000a0H 00012 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32] 00017 48 8d 0d 00 00 00 00 lea rcx, OFFSET FLAT:__68875EC1_mat3@c 0001e e8 00 00 00 00 call __CheckForDebuggerJustMyCode ; 28 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->colum[0]), ALIGN(source->colum[1]), _MM_SHUFFLE(1, 0, 1, 0)); 00023 b8 10 00 00 00 mov eax, 16 00028 48 6b c0 01 imul rax, rax, 1 0002c 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 00033 48 03 c8 add rcx, rax 00036 48 8b c1 mov rax, rcx 00039 0f 10 00 movups xmm0, XMMWORD PTR [rax] 0003c b8 10 00 00 00 mov eax, 16 00041 48 6b c0 00 imul rax, rax, 0 00045 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 0004c 48 03 c8 add rcx, rax 0004f 48 8b c1 mov rax, rcx 00052 0f 10 08 movups xmm1, XMMWORD PTR [rax] 00055 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H 00059 0f 28 c1 movaps xmm0, xmm1 0005c 0f 29 45 00 movaps XMMWORD PTR t0$[rbp], xmm0 ; 29 : __m128 t1 = _mm_shuffle_ps(ALIGN(source->colum[0]), ALIGN(source->colum[1]), _MM_SHUFFLE(3, 2, 3, 2)); 00060 b8 10 00 00 00 mov eax, 16 00065 48 6b c0 01 imul rax, rax, 1 00069 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 00070 48 03 c8 add rcx, rax 00073 48 8b c1 mov rax, rcx 00076 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00079 b8 10 00 00 00 mov eax, 16 0007e 48 6b c0 00 imul rax, rax, 0 00082 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 00089 48 03 c8 add rcx, rax 0008c 48 8b c1 mov rax, rcx 0008f 0f 10 08 movups xmm1, XMMWORD PTR [rax] 00092 0f c6 c8 ee shufps xmm1, xmm0, 238 ; 000000eeH 00096 0f 28 c1 movaps xmm0, xmm1 00099 0f 29 45 10 movaps XMMWORD PTR t1$[rbp], xmm0 ; 30 : __m128 t2 = _mm_shuffle_ps(ALIGN(source->colum[2]), *defaultVecPtr, _MM_SHUFFLE(1, 0, 1, 0)); 0009d b8 10 00 00 00 mov eax, 16 000a2 48 6b c0 02 imul rax, rax, 2 000a6 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 000ad 48 03 c8 add rcx, rax 000b0 48 8b c1 mov rax, rcx 000b3 0f 10 00 movups xmm0, XMMWORD PTR [rax] 000b6 48 8b 05 00 00 00 00 mov rax, QWORD PTR defaultVecPtr 000bd 0f c6 00 44 shufps xmm0, XMMWORD PTR [rax], 68 ; 00000044H 000c1 0f 29 45 20 movaps XMMWORD PTR t2$[rbp], xmm0 ; 31 : __m128 t3 = _mm_shuffle_ps(ALIGN(source->colum[2]), *defaultVecPtr, _MM_SHUFFLE(3, 2, 3, 2)); 000c5 b8 10 00 00 00 mov eax, 16 000ca 48 6b c0 02 imul rax, rax, 2 000ce 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 000d5 48 03 c8 add rcx, rax 000d8 48 8b c1 mov rax, rcx 000db 0f 10 00 movups xmm0, XMMWORD PTR [rax] 000de 48 8b 05 00 00 00 00 mov rax, QWORD PTR defaultVecPtr 000e5 0f c6 00 ee shufps xmm0, XMMWORD PTR [rax], 238 ; 000000eeH 000e9 0f 29 45 30 movaps XMMWORD PTR t3$[rbp], xmm0 ; 32 : ; 33 : target->colum[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0)); 000ed 0f 28 45 00 movaps xmm0, XMMWORD PTR t0$[rbp] 000f1 0f c6 45 20 88 shufps xmm0, XMMWORD PTR t2$[rbp], 136 ; 00000088H 000f6 b8 10 00 00 00 mov eax, 16 000fb 48 6b c0 00 imul rax, rax, 0 000ff 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 00106 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 34 : target->colum[1] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 1, 3, 1)); 0010a 0f 28 45 00 movaps xmm0, XMMWORD PTR t0$[rbp] 0010e 0f c6 45 20 dd shufps xmm0, XMMWORD PTR t2$[rbp], 221 ; 000000ddH 00113 b8 10 00 00 00 mov eax, 16 00118 48 6b c0 01 imul rax, rax, 1 0011c 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 00123 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 35 : target->colum[2] = _mm_shuffle_ps(t1, t3, _MM_SHUFFLE(2, 0, 2, 0)); 00127 0f 28 45 10 movaps xmm0, XMMWORD PTR t1$[rbp] 0012b 0f c6 45 30 88 shufps xmm0, XMMWORD PTR t3$[rbp], 136 ; 00000088H 00130 b8 10 00 00 00 mov eax, 16 00135 48 6b c0 02 imul rax, rax, 2 00139 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 00140 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 36 : target->colum[3] = _mm_shuffle_ps(t1, t3, _MM_SHUFFLE(3, 1, 3, 1)); 00144 0f 28 45 10 movaps xmm0, XMMWORD PTR t1$[rbp] 00148 0f c6 45 30 dd shufps xmm0, XMMWORD PTR t3$[rbp], 221 ; 000000ddH 0014d b8 10 00 00 00 mov eax, 16 00152 48 6b c0 03 imul rax, rax, 3 00156 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 0015d 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 37 : } 00161 48 8d a5 80 00 00 00 lea rsp, QWORD PTR [rbp+128] 00168 5d pop rbp 00169 c3 ret 0 transpose3to4 ENDP _TEXT ENDS ; Function compile flags: /Odtp /ZI ; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\源代码\StgShqrpGraphic\mat3.c ; COMDAT transpose3to3 _TEXT SEGMENT t0$ = 0 t1$ = 16 t2$ = 32 t3$ = 48 source$ = 144 target$ = 152 transpose3to3 PROC ; COMDAT ; 14 : { $LN3: 00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx 00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx 0000a 55 push rbp 0000b 48 81 ec a0 00 00 00 sub rsp, 160 ; 000000a0H 00012 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32] 00017 48 8d 0d 00 00 00 00 lea rcx, OFFSET FLAT:__68875EC1_mat3@c 0001e e8 00 00 00 00 call __CheckForDebuggerJustMyCode ; 15 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->colum[0]), ALIGN(source->colum[1]), _MM_SHUFFLE(1, 0, 1, 0)); 00023 b8 10 00 00 00 mov eax, 16 00028 48 6b c0 01 imul rax, rax, 1 0002c 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 00033 48 03 c8 add rcx, rax 00036 48 8b c1 mov rax, rcx 00039 0f 10 00 movups xmm0, XMMWORD PTR [rax] 0003c b8 10 00 00 00 mov eax, 16 00041 48 6b c0 00 imul rax, rax, 0 00045 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 0004c 48 03 c8 add rcx, rax 0004f 48 8b c1 mov rax, rcx 00052 0f 10 08 movups xmm1, XMMWORD PTR [rax] 00055 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H 00059 0f 28 c1 movaps xmm0, xmm1 0005c 0f 29 45 00 movaps XMMWORD PTR t0$[rbp], xmm0 ; 16 : __m128 t1 = _mm_shuffle_ps(ALIGN(source->colum[0]), ALIGN(source->colum[1]), _MM_SHUFFLE(3, 2, 3, 2)); 00060 b8 10 00 00 00 mov eax, 16 00065 48 6b c0 01 imul rax, rax, 1 00069 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 00070 48 03 c8 add rcx, rax 00073 48 8b c1 mov rax, rcx 00076 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00079 b8 10 00 00 00 mov eax, 16 0007e 48 6b c0 00 imul rax, rax, 0 00082 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 00089 48 03 c8 add rcx, rax 0008c 48 8b c1 mov rax, rcx 0008f 0f 10 08 movups xmm1, XMMWORD PTR [rax] 00092 0f c6 c8 ee shufps xmm1, xmm0, 238 ; 000000eeH 00096 0f 28 c1 movaps xmm0, xmm1 00099 0f 29 45 10 movaps XMMWORD PTR t1$[rbp], xmm0 ; 17 : __m128 t2 = _mm_shuffle_ps(ALIGN(source->colum[2]), *defaultVecPtr, _MM_SHUFFLE(1, 0, 1, 0)); 0009d b8 10 00 00 00 mov eax, 16 000a2 48 6b c0 02 imul rax, rax, 2 000a6 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 000ad 48 03 c8 add rcx, rax 000b0 48 8b c1 mov rax, rcx 000b3 0f 10 00 movups xmm0, XMMWORD PTR [rax] 000b6 48 8b 05 00 00 00 00 mov rax, QWORD PTR defaultVecPtr 000bd 0f c6 00 44 shufps xmm0, XMMWORD PTR [rax], 68 ; 00000044H 000c1 0f 29 45 20 movaps XMMWORD PTR t2$[rbp], xmm0 ; 18 : __m128 t3 = _mm_shuffle_ps(ALIGN(source->colum[2]), *defaultVecPtr, _MM_SHUFFLE(3, 2, 3, 2)); 000c5 b8 10 00 00 00 mov eax, 16 000ca 48 6b c0 02 imul rax, rax, 2 000ce 48 8b 8d 90 00 00 00 mov rcx, QWORD PTR source$[rbp] 000d5 48 03 c8 add rcx, rax 000d8 48 8b c1 mov rax, rcx 000db 0f 10 00 movups xmm0, XMMWORD PTR [rax] 000de 48 8b 05 00 00 00 00 mov rax, QWORD PTR defaultVecPtr 000e5 0f c6 00 ee shufps xmm0, XMMWORD PTR [rax], 238 ; 000000eeH 000e9 0f 29 45 30 movaps XMMWORD PTR t3$[rbp], xmm0 ; 19 : ; 20 : target->colum[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0)); 000ed 0f 28 45 00 movaps xmm0, XMMWORD PTR t0$[rbp] 000f1 0f c6 45 20 88 shufps xmm0, XMMWORD PTR t2$[rbp], 136 ; 00000088H 000f6 b8 10 00 00 00 mov eax, 16 000fb 48 6b c0 00 imul rax, rax, 0 000ff 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 00106 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 21 : target->colum[1] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 1, 3, 1)); 0010a 0f 28 45 00 movaps xmm0, XMMWORD PTR t0$[rbp] 0010e 0f c6 45 20 dd shufps xmm0, XMMWORD PTR t2$[rbp], 221 ; 000000ddH 00113 b8 10 00 00 00 mov eax, 16 00118 48 6b c0 01 imul rax, rax, 1 0011c 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 00123 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 22 : target->colum[2] = _mm_shuffle_ps(t1, t3, _MM_SHUFFLE(2, 0, 2, 0)); 00127 0f 28 45 10 movaps xmm0, XMMWORD PTR t1$[rbp] 0012b 0f c6 45 30 88 shufps xmm0, XMMWORD PTR t3$[rbp], 136 ; 00000088H 00130 b8 10 00 00 00 mov eax, 16 00135 48 6b c0 02 imul rax, rax, 2 00139 48 8b 8d 98 00 00 00 mov rcx, QWORD PTR target$[rbp] 00140 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 23 : } 00144 48 8d a5 80 00 00 00 lea rsp, QWORD PTR [rbp+128] 0014b 5d pop rbp 0014c c3 ret 0 transpose3to3 ENDP _TEXT ENDS ; Function compile flags: /Odtp /ZI ; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\源代码\StgShqrpGraphic\mat3.c ; COMDAT transpose3to2 _TEXT SEGMENT t0$ = 0 t2$ = 16 source$ = 112 target$ = 120 transpose3to2 PROC ; COMDAT ; 5 : { $LN3: 00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx 00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx 0000a 55 push rbp 0000b 48 81 ec 80 00 00 00 sub rsp, 128 ; 00000080H 00012 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32] 00017 48 8d 0d 00 00 00 00 lea rcx, OFFSET FLAT:__68875EC1_mat3@c 0001e e8 00 00 00 00 call __CheckForDebuggerJustMyCode ; 6 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->colum[0]), ALIGN(source->colum[1]), _MM_SHUFFLE(1, 0, 1, 0)); 00023 b8 10 00 00 00 mov eax, 16 00028 48 6b c0 01 imul rax, rax, 1 0002c 48 8b 4d 70 mov rcx, QWORD PTR source$[rbp] 00030 48 03 c8 add rcx, rax 00033 48 8b c1 mov rax, rcx 00036 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00039 b8 10 00 00 00 mov eax, 16 0003e 48 6b c0 00 imul rax, rax, 0 00042 48 8b 4d 70 mov rcx, QWORD PTR source$[rbp] 00046 48 03 c8 add rcx, rax 00049 48 8b c1 mov rax, rcx 0004c 0f 10 08 movups xmm1, XMMWORD PTR [rax] 0004f 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H 00053 0f 28 c1 movaps xmm0, xmm1 00056 0f 29 45 00 movaps XMMWORD PTR t0$[rbp], xmm0 ; 7 : __m128 t2 = _mm_shuffle_ps(ALIGN(source->colum[2]), *defaultVecPtr, _MM_SHUFFLE(1, 0, 1, 0)); 0005a b8 10 00 00 00 mov eax, 16 0005f 48 6b c0 02 imul rax, rax, 2 00063 48 8b 4d 70 mov rcx, QWORD PTR source$[rbp] 00067 48 03 c8 add rcx, rax 0006a 48 8b c1 mov rax, rcx 0006d 0f 10 00 movups xmm0, XMMWORD PTR [rax] 00070 48 8b 05 00 00 00 00 mov rax, QWORD PTR defaultVecPtr 00077 0f c6 00 44 shufps xmm0, XMMWORD PTR [rax], 68 ; 00000044H 0007b 0f 29 45 10 movaps XMMWORD PTR t2$[rbp], xmm0 ; 8 : ; 9 : target->colum[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0)); 0007f 0f 28 45 00 movaps xmm0, XMMWORD PTR t0$[rbp] 00083 0f c6 45 10 88 shufps xmm0, XMMWORD PTR t2$[rbp], 136 ; 00000088H 00088 b8 10 00 00 00 mov eax, 16 0008d 48 6b c0 00 imul rax, rax, 0 00091 48 8b 4d 78 mov rcx, QWORD PTR target$[rbp] 00095 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 10 : target->colum[1] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 1, 3, 1)); 00099 0f 28 45 00 movaps xmm0, XMMWORD PTR t0$[rbp] 0009d 0f c6 45 10 dd shufps xmm0, XMMWORD PTR t2$[rbp], 221 ; 000000ddH 000a2 b8 10 00 00 00 mov eax, 16 000a7 48 6b c0 01 imul rax, rax, 1 000ab 48 8b 4d 78 mov rcx, QWORD PTR target$[rbp] 000af 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0 ; 11 : } 000b3 48 8d 65 60 lea rsp, QWORD PTR [rbp+96] 000b7 5d pop rbp 000b8 c3 ret 0 transpose3to2 ENDP _TEXT ENDS END