00:00:00 --- log: started retro/12.11.28
00:25:02 --- join: oPless (~oPless@lart.doosh.net) joined #retro
03:30:11 --- join: tangentstorm (~michal@108-218-151-22.lightspeed.rcsntx.sbcglobal.net) joined #retro
08:24:29 --- join: karswell (~coat@93-97-29-243.zone5.bethere.co.uk) joined #retro
10:55:32 --- join: Mat2 (~claude@91-65-144-133-dynip.superkabel.de) joined #retro
10:55:39 <Mat2> hello !
10:57:26 <Mat2> crc: thanks for the link, that is the C version I searched for. For my surprise it is not open-source so I'm out of luck :(
11:26:55 <crc> sorry :(
11:29:54 <Mat2> I use know the Lua vm as benchmark reference
11:30:29 <Mat2> (and gforth-fast)
11:36:31 * Mat2 find a way to efficient SIMD processing despite Intel SSE4
11:56:28 * Mat2 found a way to efficient SIMD processing on Intel cpu's
11:59:38 <tangentstorm> How's the new technique working out, Mat2?
12:11:59 <Mat2> works fine, specially on Intel Atom64 because SSE instructions can be paired with two integer ALU's
12:13:17 <Mat2> that doubles performance near the theoretical maximum (from ~670 to ~1600 mips)
12:14:23 <tangentstorm> :)
12:15:38 <Mat2> the drawback is: My fan turns on
12:18:57 <Mat2> somehow I have the feeling that these Intel Atom cpu's were not designed for such applications...
12:19:25 <tangentstorm> probably not!
12:19:36 <tangentstorm> can you throttle it?
12:19:58 <tangentstorm> they're probably also not used to people actually using the cpu to its full potential
12:20:22 <Mat2> I don't think so, anyhow: It works
12:21:20 <Mat2> SSE sucks by the way
12:21:27 <tangentstorm> can you use the mmx registers alongside sse?
12:22:10 <Mat2> yes, because there are mapped to the floating point stack
12:22:45 <Mat2> and I can use all 64 bit integer register of course
12:23:27 <Mat2> (the second 8 ones are used as call stack for example)
12:26:45 <Mat2> I use this manual: The microarchitecture of Intel, AMD and VIA CPUs: An optimization guide for assembly programmers and compiler makers
12:26:50 <Mat2> The microarchitecture of Intel, AMD and VIA CPUs: An optimization guide for assembly programmers and compiler makers
12:27:06 <Mat2> sorry: http://www.agner.org/optimize/
12:28:20 <Mat2> If you are interested in effective assembly programming I would recommend it
12:30:22 <tangentstorm> it's definitely interesting to me, i just don't have time to follow through on everything i'm interested in :)
12:31:01 <Mat2> understandable :)
12:33:16 <Mat2> I upload my promised sources this night for sure by the way, sorry for the delay but but I wanted to get the maximum performance out of these cpu
12:34:18 <tangentstorm> no problem. i've been working on a grammar / parsing engine
12:34:58 <tangentstorm> https://github.com/sabren/b4/blob/master/pre/test_pre.pas
12:36:05 <tangentstorm> all tests pass, finally, but all it does is match... next up is actually build a syntax tree
12:37:02 <Mat2> looks good
12:39:24 <tangentstorm> I've also been digging through SWAG - an archive of old public domain pascal code from the DOS days.
12:39:24 <tangentstorm> i think i might set up a website that categorizes it... some of it is completely archaic but much is still relevant
12:39:42 <tangentstorm> http://kd5col.info/swag/
12:40:43 <tangentstorm> also... i know you're not a huge fan of LLVM, but ... http://code.google.com/p/llvm-pascal/
12:42:39 <Mat2> it is always good to have a compiler as alternative
12:44:24 <Mat2> the problem with LLVM are the limitations of there IL code which results in inefficient code for some applications
12:44:47 <Mat2> like efficient interpretation
12:54:48 <Mat2> the VGA related sources on these website are very useful for driver development (offer nice tricks for a basic framebuffer)
12:56:35 <tangentstorm> i'm curious how much of this stuff still works :)
12:57:54 <tangentstorm> i found a small calculator example on there last night that converts expressions to RPN
13:03:30 <Mat2> oh, algorithms don't get older, only better :D
13:04:57 <tangentstorm> :) yeah. i'm reading "brinch hansen on pascal compilers" ... from like 1985 or so. it's really good
13:13:02 <Mat2> I have here the book from Wirth about the oberon compiler, nice read
13:14:53 <tangentstorm> Yeah, he's put all his old books online now in PDF. They're really good.
13:15:19 <tangentstorm> in fact, i was going to use his RISC machine for my stuff until i found ngaro.
13:15:51 <Mat2> now you using a virtual MISC machine :)
13:16:45 <tangentstorm> yep :)
13:20:50 <Mat2> I'll recommend the site of Jeff Fox, you can find there a lot about optimizing code for MISC style cpu's
13:21:41 <Mat2> http://www.ultratechnology.com/
13:22:14 <Mat2> most is colour forth related
13:22:43 <tangentstorm> oh yeah, i've seen this... i completely ignored the "chips" section, but he's got all kinds of fpga stuff on there. nice.
13:23:56 <Mat2> he was very influential 
15:08:21 --- quit: saper (Read error: Operation timed out)
15:09:13 --- join: saper (saper@wikipedia/saper) joined #retro
15:51:31 <Mat2> tangentstorm: ok, the interpreter stub is uploaded, I only need to complete the primitive routines. That will be my work tomorrow
15:53:07 <tangentstorm> :)
15:53:10 * tangentstorm takes a look
15:56:26 <tangentstorm> that's actually pretty readable for assembly code
15:57:06 <tangentstorm> what do Li and Ld mean in the opcodes?
15:57:17 <tangentstorm> int / double ?
15:58:33 <Mat2> Load Immediate and Load Address
15:59:11 <Mat2> LI -> LIT, LD -> LOAD
16:00:17 <Mat2> BEA = Branch if Equal Always <adr>
16:08:15 <Mat2> it was an interesting exercise to schedule the instructions so both in-order paths work in parallel
16:11:24 <tangentstorm> well, i said readable "for assembly code" but I don't really understand assembly code :) i'm not sure what you're talking about
16:12:54 <tangentstorm> I think Dsp, Rsp, and Qsp are stacks?
16:13:28 <Mat2> the Atom cpu have two separate pipelines so two instructions can be executed parallel if there are no ressource conflicts
16:13:31 <tangentstorm> CND doesn't seem to be set to anything but false
16:13:59 <tangentstorm> oh, so you arranged the instructions to avoid conflicts
16:14:35 <tangentstorm> i guess DSP is the stack pointer for the D stack
16:14:47 <tangentstorm> so that would be Data, Return and Q-Stack
16:15:17 <Mat2> oh, that variable is not needed anymore because I now use the native flag register
16:15:19 <Mat2> yes
16:15:36 <Mat2> the Q stack is for closures
16:16:19 <tangentstorm> you're only using xmm0 right?
16:16:29 <Mat2> right
16:17:57 <Mat2> offsets into the call table are computed with the SIMD unit
16:20:37 <Mat2> and this unit is fully piplined (a hidden, third pipeline). The official documentation conceal this detail
16:21:40 <Mat2> so with the right sheduling SIMD processing cost no extra clock cycles
16:21:57 <tangentstorm> your instructions are 4-bit, right?
16:22:02 <Mat2> yes
16:22:45 <tangentstorm> so you could store 32 of them in each of the other sse registers
16:23:23 <tangentstorm> there are instructions to rotate those registers, too
16:23:36 <tangentstorm> so each one could basically be a loop
16:24:35 <tangentstorm> so you could have 7 threads running concurrently, with the complete code all stored directly in the registers
16:24:51 <tangentstorm> if the code fits in 32 instructions i mean
16:24:57 <Mat2> yes
16:26:33 <Mat2> exactly 15 because there exist 16 xmm register
16:27:03 <Mat2> that's the plan for multiprocessing
16:27:26 <tangentstorm> or if yo could fit it in 16 instructions, then you could have a 32-bit TOS and NOS for each one ( just use half of each register )
16:28:48 <Mat2> for caching these virtual registers I will use the mmx register set
16:29:39 <Mat2> so working with register halves is not really needed
16:30:44 <Mat2> my guess is the fp unit is also pipelined (no extra cycles if carefully used)
16:31:32 <tangentstorm> neat stuff
16:32:14 <Mat2> I need some sleep, see you
16:32:21 <Mat2> ciao
16:32:34 --- quit: Mat2 (Quit: Verlassend)
19:28:17 --- quit: tangentstorm (Quit: trying out new tmux.conf)
19:48:21 --- join: tangentstorm (~michal@108-218-151-22.lightspeed.rcsntx.sbcglobal.net) joined #retro
20:08:19 --- quit: tangentstorm (Quit: WeeChat 0.3.2)
20:18:06 --- join: tangentstorm (~michal@108.218.151.22) joined #retro
20:20:46 --- quit: tangentstorm (Client Quit)
20:24:57 --- join: tangentstorm (~michal@108-218-151-22.lightspeed.rcsntx.sbcglobal.net) joined #retro
23:56:49 <tangentstorm> docl: I saw a giant book that was all about programming MUDS today
23:59:51 <tangentstorm> pretty sure it was this : http://www.amazon.com/Game-Programming-Premier-Press-Development/dp/1592000908
23:59:59 --- log: ended retro/12.11.28