BorlandTalk.com Borland discussion newsgroups   Archives   FAQ   Search   Memberlist   Usergroups   Register   Profile   Log in to check your private messages   Log in  repeated float operatoins.. SSE?          BorlandTalk.com Forum Index -> Delphi Language BASM View previous topic :: View next topic   Author Message Nils Haeck Guest Posted: Wed Mar 21, 2007 4:00 am    Post subject: repeated float operatoins.. SSE? Hi Guys, I need to do the following as fast as possible: procedure Convert(X, Y: pdouble; Count: integer; const A, B: double); var i: integer; begin for i := 0 to Count - 1 do begin Y^ := X^ * A + B; inc(X); inc(Y); end; end; The Count variable is usually something like 100.000 (so X and Y are arrays of 100.000 elements). How can I optimize this using ASM? I'm looking at something with performance for floating point numbers. Preferrably I work with double precision floats, but I could convert the algorithm to work with single precision floats, if it is significantly faster. Thanks in advance, Nils Haeck Back to top Les Guest Posted: Fri Mar 23, 2007 2:38 am    Post subject: Re: repeated float operatoins.. SSE? Quote: I need to do the following as fast as possible: procedure Convert(X, Y: pdouble; Count: integer; const A, B: double); var i: integer; begin for i := 0 to Count - 1 do begin Y^ := X^ * A + B; inc(X); inc(Y); end; end; The Count variable is usually something like 100.000 (so X and Y are arrays of 100.000 elements). Both procedures require SSE2. I have no access to SSE2 CPU right now, so it is completely untested. The loops can be unrolled by 2 or 4 for some extra speed increase (interweave xmm2 with xmm3, xmm4, xmm5). The Count requirements on entry would change if loops were to be unrolled. ConvSingle should theoretically be twice as fast. (Count >= 2) and (Count mod 2 = 0). X and Y must be 16 aligned. If not then use movupd instead of movapd (slower). procedure ConvDouble(X, Y: PDouble; Count: Integer; const A, B: Double); register; asm movsd xmm0, [ebp+20] movsd xmm0, [ebp+12] pshufd xmm0, xmm0, $44 pshufd xmm1, xmm1, $44 and ecx, -2 shl ecx, 3 add eax, ecx add edx, ecx neg ecx @loop: movapd xmm2, [eax+ecx] mulpd xmm2, xmm0 addpd xmm2, xmm1 movapd [edx+ecx], xmm2 add ecx, 16 js @loop end; (Count >= 4) and (Count mod 4 = 0). X and Y must be 16 aligned. If not then use movups instead of movaps (slower). procedure ConvSingle(X, Y: PSingle; Count: Integer; const A, B: Single); register; asm movss xmm0, [ebp+12] movss xmm1, [ebp+8] pshufd xmm0, xmm0, 0 pshufd xmm1, xmm1, 0 and ecx, -4 shl ecx, 2 add eax, ecx add edx, ecx neg ecx @loop: movaps xmm2, [eax+ecx] mulps xmm2, xmm0 addps xmm2, xmm1 movaps [edx+ecx], xmm2 add ecx, 16 js @loop @exit: end; Les. Back to top Nils Haeck Guest Posted: Fri Mar 23, 2007 7:16 pm    Post subject: Re: repeated float operatoins.. SSE? Thank you, Les.. I'll play around with it. Nils Back to top Nils Haeck Guest Posted: Fri Mar 23, 2007 8:12 pm    Post subject: Re: repeated float operatoins.. SSE? I did some testing.. Good news: your code seems to work :) Bad news: there's no difference *at all* in speed, when I use the routine I wrote, or switch to SSE/SSE2. I'm using Delphi7's compiler, and an Intel Pentium4 CPU (2.93Gz). Might it be that somehow the SSE commmands get mapped to "standard" FP processing? Or might it be that Delphi already optimizes my pascal code to use SSE? Nils "Les" <a (AT) b (DOT) c> schreef in bericht news:4602f763$1 (AT) newsgroups (DOT) borland.com... Quote: I need to do the following as fast as possible: procedure Convert(X, Y: pdouble; Count: integer; const A, B: double); var i: integer; begin for i := 0 to Count - 1 do begin Y^ := X^ * A + B; inc(X); inc(Y); end; end; The Count variable is usually something like 100.000 (so X and Y are arrays of 100.000 elements). Both procedures require SSE2. I have no access to SSE2 CPU right now, so it is completely untested. The loops can be unrolled by 2 or 4 for some extra speed increase (interweave xmm2 with xmm3, xmm4, xmm5). The Count requirements on entry would change if loops were to be unrolled. ConvSingle should theoretically be twice as fast. (Count >= 2) and (Count mod 2 = 0). X and Y must be 16 aligned. If not then use movupd instead of movapd (slower). procedure ConvDouble(X, Y: PDouble; Count: Integer; const A, B: Double); register; asm movsd xmm0, [ebp+20] movsd xmm0, [ebp+12] pshufd xmm0, xmm0, $44 pshufd xmm1, xmm1, $44 and ecx, -2 shl ecx, 3 add eax, ecx add edx, ecx neg ecx @loop: movapd xmm2, [eax+ecx] mulpd xmm2, xmm0 addpd xmm2, xmm1 movapd [edx+ecx], xmm2 add ecx, 16 js @loop end; (Count >= 4) and (Count mod 4 = 0). X and Y must be 16 aligned. If not then use movups instead of movaps (slower). procedure ConvSingle(X, Y: PSingle; Count: Integer; const A, B: Single); register; asm movss xmm0, [ebp+12] movss xmm1, [ebp+8] pshufd xmm0, xmm0, 0 pshufd xmm1, xmm1, 0 and ecx, -4 shl ecx, 2 add eax, ecx add edx, ecx neg ecx @loop: movaps xmm2, [eax+ecx] mulps xmm2, xmm0 addps xmm2, xmm1 movaps [edx+ecx], xmm2 add ecx, 16 js @loop @exit: end; Les. Back to top Nils Haeck Guest Posted: Fri Mar 23, 2007 8:22 pm    Post subject: Re: repeated float operatoins.. SSE? Update.. It does seem to be a mem access problem. When I use 300x a list length of 1000000 (1 million), I see no speed difference. When I use 30000 x list length of 10000 (so same number in total), I see significant speed advantage for the SSE methods. Fortunately, this last case is more like "real-life". However, I'd like to know if there's anything I can do to speed up memory access, as this seems to be the limiting factor. Nils Back to top Les Pawelczyk Guest Posted: Fri Mar 23, 2007 11:22 pm    Post subject: Re: repeated float operatoins.. SSE? Quote: It does seem to be a mem access problem. When I use 300x a list length of 1000000 (1 million), I see no speed difference. When I use 30000 x list length of 10000 (so same number in total), I see significant speed advantage for the SSE methods. Fortunately, this last case is more like "real-life". However, I'd like to know if there's anything I can do to speed up memory access, as this seems to be the limiting factor. Try procedure below. (count >= and (count mod 8 = 0). X and Y must be 16 aligned but it will likely be faster if they are 64 aligned. For best performance the offset value in 'prefetchnta' (128) needs to be experimented with on a particular system. Anything between 128 and 512 (in 64 byte increments) may prove to be the best. If the array is really large (much larger than L2 cache) then you can try 'movntpd' instead of 'movapd' when saving back to memory. The second 'prefetchnta' (the one with edx) will not be needed in this case. procedure ConvDouble(X, Y: PDouble; Count: Integer; const A, B: Double); register; asm movsd xmm0, [ebp+20] movsd xmm0, [ebp+12] pshufd xmm0, xmm0, $44 pshufd xmm1, xmm1, $44 and ecx, -8 shl ecx, 3 add eax, ecx add edx, ecx neg ecx @loop: prefetchnta [eax+ecx+128] prefetchnta [edx+ecx+128] movapd xmm2, [eax+ecx] movapd xmm3, [eax+ecx+16] movapd xmm4, [eax+ecx+32] movapd xmm5, [eax+ecx+48] mulpd xmm2, xmm0 mulpd xmm3, xmm0 mulpd xmm4, xmm0 mulpd xmm5, xmm0 addpd xmm2, xmm1 addpd xmm3, xmm1 addpd xmm4, xmm1 addpd xmm5, xmm1 movapd [edx+ecx], xmm2 movapd [edx+ecx+16], xmm3 movapd [edx+ecx+32], xmm4 movapd [edx+ecx+48], xmm5 add ecx, 64 js @loop end; Les. Back to top Nils Haeck Guest Posted: Sat Mar 24, 2007 4:46 am    Post subject: Re: repeated float operatoins.. SSE? Thanks a lot, Les. I'm trying to understand what happens.. I have added some comments here: procedure ConvDoubleSSE2(X, Y: PDouble; Count: Integer; const A, B: Double); //(Count >= 2) and (Count mod 2 = 0). X and Y must be 16 aligned. If not then //use movupd instead of movapd (slower). register; asm movsd xmm0, [ebp+16] // Copy A to xmm0: 0 A movsd xmm1, [ebp+8] // Copy B to xmm1: 0 B pshufd xmm0, xmm0, $44 // xmm0: A A pshufd xmm1, xmm1, $44 // xmm1: B B and ecx, -2 // clear last two bits of ecx (Count now multiple of 4) shl ecx, 3 // Count := Count * 8 add eax, ecx // X pointer + Count add edx, ecx // Y pointer + Count neg ecx // Count := - Count @loop: movapd xmm2, [eax+ecx] // mmx2: X0 X1 (two X doubles) mulpd xmm2, xmm0 // X := X * A (two doubles at a time) addpd xmm2, xmm1 // X := X + B (two doubles at a time) movapd [edx+ecx], xmm2 // Store as Y (two doubles at a time) add ecx, 16 // increment count with 2 * 8 (2 doubles) js @loop // loop as long as count < 0 end; I think I understand most, but I have a question about first two lines. How does it work with e.g. [ebp+16]? I would think +16 would be B and +8 would be A, but this seems not the case. Also, if I would want to do a bit more complex thing, like Z := A * X + B * Y + C Would I still have enough registers for the mem pointers and counters? PS I'll also test the prefetch code. Nils Back to top Nils Haeck Guest Posted: Sat Mar 24, 2007 8:11 am    Post subject: Re: repeated float operatoins.. SSE? Thank you Bob, I'm starting to get the gripes of it :) Quote: -2 has only one clear bit, not 2. ecx is now even; Yes I saw this too late, already realised it. But thanks anyway. So about the stack.. procedure Blabla(...... const A, B: Double); [ebp + 16] points to A [ebp + 8] points to B but in the other example.. procedure Blabla(..... const A, B: Single); I would expect [ebp + 8] points to A [ebp + 4] points to B But obviously when I try it is not correct; I have to use [ebp + 12] points to A [ebp + 8] points to B Is there some "fixed offset" of 8? Then about the other parameters, is it always true that procedure Blabla(x, y, z: integer; ....) eax = x edx = y ecx = z (of course "integer" can also be "pointer" or any other 32bit equivalent) Also, do I have to push/pop any of these to save them? Nils Back to top Bob Gonder Guest Posted: Sat Mar 24, 2007 8:11 am    Post subject: Re: repeated float operatoins.. SSE? Nils Haeck wrote: Quote: and ecx, -2 // clear last two bits of ecx (Count now multiple of 4) -2 has only one clear bit, not 2. ecx is now even; Quote: procedure ConvDoubleSSE2(X, Y: PDouble; Count: Integer; const A, B: Double); movsd xmm0, [ebp+16] // Copy A to xmm0: 0 A movsd xmm1, [ebp+8] // Copy B to xmm1: 0 B I think I understand most, but I have a question about first two lines. How does it work with e.g. [ebp+16]? I would think +16 would be B and +8 would be A, but this seems not the case. Pascal Parameters are pushed onto the stack from left to right. So A is pushed first, then B. The stack grows Down to smaller addresses, so A is above B. Quote: Also, if I would want to do a bit more complex thing, like Z := A * X + B * Y + C Would I still have enough registers for the mem pointers and counters? I don't know SSE, but you are only using 0,1,2, so at least 3 should be available. Back to top Bob Gonder Guest Posted: Sat Mar 24, 2007 7:51 pm    Post subject: Re: repeated float operatoins.. SSE? Nils Haeck wrote: Quote: Is there some "fixed offset" of 8? FunctionEntryPoint: // esp == return address push ebp // esp+4 == return address mov ebp,esp // ebp+4 = return address // ebp+8 == first parameter on stack sub esp, local stack size Quote: Then about the other parameters, is it always true that procedure Blabla(x, y, z: integer; ....) eax = x edx = y ecx = z Only if Pascal is using the fastcall method. And only if the first 3 parameters are 32bit or less. If first param is 64bit integer, could be eax = low dword and edx = high dword, so ecx is second param (if 32bit or less) Quote: Also, do I have to push/pop any of these to save them? Only if your code requires it. Copy: push esi // cannot change these registers push edi mov esi,eax // source mov edi,edx // destination rep movsb // count ecx down to 0 pop edi // restore non-changeable registers pop esi ret Copy2Places: push ebp mov esp,ebp push esi // cannot change these registers push edi push eax // save parameters we will reuse push ecx mov esi,eax mov edi,edx rep movsb mov edi,[ebp+8] // 4th parameter pop ecx // reuse parameters pop esi rep movsb pop edi // restore non-changeable registers pop esi pop ebp ret 4 // remove 4th parameter (4 bytes) Back to top Les Pawelczyk Guest Posted: Mon Mar 26, 2007 7:16 pm    Post subject: Re: repeated float operatoins.. SSE? Quote: Also, if I would want to do a bit more complex thing, like Z := A * X + B * Y + C Would I still have enough registers for the mem pointers and counters? There is eight xmm registers: xmm0 - xmm7. There is sixteen in 64 bit environment. Quote: PS I'll also test the prefetch code. Don't forget that prefetch only needs to be done once every 64 bytes. Les. Back to top Les Pawelczyk Guest Posted: Mon Mar 26, 2007 7:16 pm    Post subject: Re: repeated float operatoins.. SSE? Quote: I would expect [ebp + 8] points to A [ebp + 4] points to B But obviously when I try it is not correct; I have to use [ebp + 12] points to A [ebp + 8] points to B The 'call' instruction used to run your procedure needs 4 bytes to save the return address. Quote: Then about the other parameters, is it always true that procedure Blabla(x, y, z: integer; ....) eax = x edx = y ecx = z Only if you use 'register' calling convention. This is a default calling convention but it doesn't hurt to specify it anyway. Quote: Also, do I have to push/pop any of these to save them? You can freely use eax, edx and ecx without having to restore their original values before return. Every other register must be preserved. Les. Back to top Nils Haeck Guest Posted: Mon Mar 26, 2007 9:56 pm    Post subject: Re: repeated float operatoins.. SSE? Thanks for all the feedback, everyone! It makes it all much more clear now. Nils Back to top Display posts from previous: All Posts1 Day7 Days2 Weeks1 Month3 Months6 Months1 Year Oldest FirstNewest First         BorlandTalk.com Forum Index -> Delphi Language BASM All times are GMT Page 1 of 1   Jump to: Select a forum C++ Builder Newsgroups----------------C++ BuilderC++ Builder (ActiveX)C++ Builder (Announcements)C++ Builder (Command Line Tools)C++ Builder DatabasesC++ Builder Databases (ADO)C++ Builder Databases (Desktop)C++ Builder Databases (InterBase Express)C++ Builder Databases (Multi-Tier)C++ Builder Databases (SQL Servers)C++ Builder (Graphics)C++ Builder (IDE)C++ Builder (Internet)C++ Builder (Internet Socket)C++ Builder (Internet Web)C++ Builder (Jobs)C++ Builder (Language)C++ Builder (Language C++)C++ Builder (Compatibility with MFC)C++ Builder (Native API)C++ Builder (Non-Technical)C++ Builder (Students)C++ Builder (Third Party Tools)C++ Builder (Upgrade)C++ Builder (VCL)C++ Builder (VCL Components Usage)C++ Builder (VCL Components Development)C++ Builder (VCL Usage)C++ Builder (WinAPI)C++ Builder Installation Delphi Newsgroups----------------Delphi (General)Delphi ActiveX Controls UsageDelphi ActiveX Controls DevelopmentDelphi As400Delphi DatabasesDelphi Databases (ADO)Delphi Databases (dbExpress)Delphi Databases (Desktop)Delphi Databases (InterBase Express)Delphi Databases (Multi-Tier)Delphi Databases (SQL Servers)Delphi DeploymentDelphi GraphicsDelphi IDEDelphi InternationalizationDelphi InternetDelphi Internet Isapi-WebBrokerDelphi Internet WebSnapDelphi Internet WinsockDelphi JobsDelphi Language BASMDelphi Language ObjectPascalDelphi Model Driven Architecture (General)Delphi Model Driven Architecture (Third Party)Delphi MultimediaDelphi Native APIDelphi Non-TechnicalDelphi ObjectPascalDelphi OLE AutomationDelphi OO designDelphi OpenToolsAPIDelphi Reporting-ChartingDelphi RTLDelphi StudentsDelphi TeamSourceDelphi Thirdparty ToolsDelphi Thirdparty Tools (General)Delphi Thirdparty Tools (IntraWeb)Delphi Thirdparty Tools (RaveReports)Delphi UpgradeDelphi VCLDelphi VCL ComponentsDelphi VCL Components UsingDelphi VCL Components WritingDelphi WebServices SOAPDelphi WinAPIDelphi XMLDelphi Installationcomp.lang.pascal.delphi.components.misccomp.lang.pascal.delphi.components.usagecomp.lang.pascal.delphi.components.writingcomp.lang.pascal.delphi.databasescomp.lang.pascal.delphi.misccomp.sources.delphi JBuilder Newsgroups----------------JBuilderJBuilder AnnouncementsJBuilder Applet IssuesJBuilder As400JBuilder CompilerJBuilder CORBA/RMIJBuilder DatabasesJBuilder DB SwingJBuilder DebuggerJBuilder DeploymentJBuilder DocumentationJBuilder EnterpriseJBuilder HandheldJBuilder IDEJBuilder JavaJBuilder JavaAPIJBuilder Java LanguageJBuilder Java SwingJBuilder Java BeansJBuilder Java Beans UsageJBuilder Java Beans DevelopmentJBuilder JBCLJBuilder JDataStoreJBuilder JobsJBuilder LinuxJBuilder MacintoshJBuilder & MobileSetJBuilder MultiLingual AppsJBuilder Non-TechnicalJBuilder OpenToolsJBuilder Servlets (JSP)JBuilder StudentsJBuilder Team DevelopmentJBuilder Thirdparty ToolsJBuilder UpgradeJBuilder Installation InterBase Newsgroups----------------InterBaseInterBase Bug DiscussionInterBase GeneralInterBase IBConsoleInterBase OpenSourceInterBase ReplicationInterBase SQLInterBase ThirdpartyInterBase Installation Miscellaneous Borland Newsgroups----------------AppCenterAppServerEnteraVisiBroker  You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum Powered by phpBB © 2001, 2006 phpBB Group SEO toolkit © 2004-2006 webmedic.