Chris,
All I'm trying to say is Learn From My Fail. After dealing with such code for a while (in C --- yuck), I realized it was much better to use the compiler intrinsics. Once I had the new code running, I deleted numerous implementations of the "let's swap bytes around" business. IIRC it was a net lines-of-code loss, and less code is great.
I'm not disputing the performance gains. But consider how much faster and simpler still the compiler intrinsic approach could be.
The Smalltalk code does 8 at:, which are bound checked. There are also 7 additions, which have to be checked against overflow. Then there are some comparisons, some ifFalse:, more bitShift: (overflow check), more additions (overflow check), and so on. Creating large integers is going to be costly. Even with optimized bounds / overflow checks, surely that's going to expand to tens of assembly instructions, if not a couple hundred.
I'm not saying this is an inefficient Smalltalk way of doing things, however I'd point out it's reimplementing what's available in hardware.
In constrast, if that code was made into a primitive, a relevant compiler intrinsic would be available to help. The VM would grab a 64 bit integer with a MOVBE instruction, perform one test for overflow, and if all is well then return tagging the integer with a single LEA instruction. If the result must be a large integer, the VM might as well create it. Surely a decent compiler can express all that in very little code, because it's effectively sweeping all the complexity under MOVBE.
For the sake of illustration, and assuming only one tag bit for simplicity, the assembly would be something like:
; calculate the pointer to dereference in rax, then... movbe rax, [rax] test rax, rax js overflowToLargeInteger lea rax, [rax+rax+1] ret
; ok, it didn't fit, so... overflowToLargeInteger: call largePositiveIntegerFromRAX ; returning in RAX ret
I'd imagine the integer arithmetic cannot possibly fit in that space.
Andres.
On 8/31/15 20:21 , Chris Cunningham wrote:
Hi Andres,
ByteArray currently doesn't have a primitive that handles any part of getting bytes from the ByteArray and forming them into an integer. If it did have one, I would be happy to alter the code around that.
The long drawn out method is 4x faster for small (SmallInteger) results, and 25% faster for LargeInteger results (those that excercise all 8 bytes). This because it does at most 2 LargeInteger bitShifts, and as little as no LargeInteger bitShifts. The 'macro' version does a minimum of 1 LargeInteger bitShifts, and up to 3 of them.
For BigEndian platforms, speed may be important; in any case, it is nice.
You are probably aware, but the current Squeak has does not have #unsignedLong64At:bigEndian: in the image at all - that diff was from my first attempt.
-cbc
On Mon, Aug 31, 2015 at 7:39 PM, Andres Valloud <avalloud@smalltalk.comcastbiz.net mailto:avalloud@smalltalk.comcastbiz.net> wrote:
Interesting about the fading relevancy of big endian platforms. Just in case the point was lost, I meant the macro-style approach in contrast with this (from Squeak-dev): =============== Diff against Collections-cbc.650 =============== Item was changed: ----- Method: ByteArray>>unsignedLong64At:bigEndian: (in category 'platform independent access') ----- unsignedLong64At: index bigEndian: aBool + "Avoid as much largeInteger as we can" + | b0 b2 b3 b5 b6 w n2 n3 | + + aBool ifFalse: [ + w := self at: index. + b6 := self at: index+1. + b5 := self at: index+2. + n2 := self at: index+3. + b3 := self at: index+4. + b2 := self at: index+5. + n3 := self at: index+6. + b0 := self at: index+7. + ] ifTrue: [ + b0 := self at: index. + n3 := self at: index+1. + b2 := self at: index+2. + b3 := self at: index+3. + n2 := self at: index+4. + b5 := self at: index+5. + b6 := self at: index+6. + w := self at: index+7. + ]. + + "Minimize LargeInteger arithmetic" + b6 = 0 ifFalse:[w := (b6 bitShift: 8) + w]. + b5 = 0 ifFalse:[w := (b5 bitShift: 16) + w]. + + b3 = 0 ifFalse:[n2 := (b3 bitShift: 8) + n2]. + b2 = 0 ifFalse:[n2 := (b2 bitShift: 16) + n2]. + n2 == 0 ifFalse: [w := (n2 bitShift: 24) + w]. + + b0 = 0 ifFalse:[n3 := (b0 bitShift: 8) + n3]. + n3 == 0 ifFalse: [w := (n3 bitShift: 48) + w]. + + ^w! - | n1 n2 | - aBool - ifTrue: [ - n2 := self unsignedLongAt: index bigEndian: true. - n1 := self unsignedLongAt: index+4 bigEndian: true. - ] - ifFalse: [ - n1 := self unsignedLongAt: index bigEndian: false. - n2 := self unsignedLongAt: index+4 bigEndian: false. - ]. - ^(n2 bitShift: 32) + n1! I'd rather have that pushed down enough so that the compiler intrinsic becomes visible. And at that point, all that code is reduced to a single instruction. Andres. On 8/31/15 19:12 , Eliot Miranda wrote: Hi Andres, On Aug 31, 2015, at 5:52 PM, Andres Valloud <avalloud@smalltalk.comcastbiz.net <mailto:avalloud@smalltalk.comcastbiz.net>> wrote: FWIW... IMO it's better to enable access to the relevant compiler intrinsic with platform specific macros, rather than implementing instructions such as Intel's BSWAP or MOVBE by hand. In HPS, isolating endianness concerns from the large integer arithmetic primitives with such macros enabled 25-40% faster performance on big endian platforms. Just as importantly, the intrinsic approach takes significantly less code to implement. Makes sense, and the performance increases are impressive. The only issue I have is that the Cog JIT (which would have the easiest time generating those intrinsics) currently runs only in little-endianness platforms and I seriously doubt it will run in a big endianness platform in the next five years. PowerPC is the only possibility I see. Yes, ARM is biendian but all the popular applications I know of are little endian. VW's a different beast; significant big endian legacy. But what you say about isolating makes perfect sense. Thanks On 8/31/15 10:25 , Eliot Miranda wrote: Hi Chrises, my vote would be to write these as 12 numbered primitives, (2,4 & 8 bytes) * (at: & at:put:) * (big & little endian) because they can be performance critical and implementing them like this means the maximum efficiency in both 32-bit and 64-bit Spur, plus the possibility of the JIT implementing the primitives. On Sun, Aug 30, 2015 at 10:01 PM, Chris Cunningham <cunningham.cb@gmail.com <mailto:cunningham.cb@gmail.com> <mailto:cunningham.cb@gmail.com <mailto:cunningham.cb@gmail.com>>> wrote: Hi Chris, I'm all for having the fastest that in the image that works. If you could make your version handle endianess, then I'm all for including it (at least in the 3 variants that are faster). My first use for this (interface for KAFKA) apparently requires bigEndianess, so I really want that supported. It might be best to keep my naming, though - it follows the name pattern that is already in the class. Or will yours also support 128? -cbc On Sun, Aug 30, 2015 at 2:38 PM, Chris Muller <asqueaker@gmail.com <mailto:asqueaker@gmail.com> <mailto:asqueaker@gmail.com <mailto:asqueaker@gmail.com>>> wrote: Hi Chris, I think these methods belong in the image with the fastest implementation we can do. I implemented 64-bit unsigned access for Ma Serializer back in 2005. I modeled my implementation after Andreas' original approach which tries to avoid LI arithmetic. I was curious whether your implementations would be faster, because if they are then it could benefit Magma. After loading "Ma Serializer" 1.5 (or head) into a trunk image, I used the following script to take comparison measurements: | smallN largeN maBa cbBa | smallN := ((2 raisedTo: 13) to: (2 raisedTo: 14)) atRandom. largeN := ((2 raisedTo: 63) to: (2 raisedTo: 64)) atRandom. maBa := ByteArray new: 8. cbBa := ByteArray new: 8. maBa maUint: 64 at: 0 put: largeN. cbBa unsignedLong64At: 1 put: largeN bigEndian: false. self assert: (cbBa maUnsigned64At: 1) = (maBa unsignedLong64At: 1 bigEndian: false). { 'cbc smallN write' -> [ cbBa unsignedLong64At: 1 put: smallN bigEndian: false] bench. 'ma smallN write' -> [cbBa maUint: 64 at: 0 put: smallN ] bench. 'cbc smallN access' -> [ cbBa unsignedLong64At: 1 bigEndian: false. ] bench. 'ma smallN access' -> [ cbBa maUnsigned64At: 1] bench. 'cbc largeN write' -> [ cbBa unsignedLong64At: 1 put: largeN bigEndian: false] bench. 'ma largeN write' -> [cbBa maUint: 64 at: 0 put: largeN ] bench. 'cbc largeN access' -> [ cbBa unsignedLong64At: 1 bigEndian: false ] bench. 'ma largeN access' -> [ cbBa maUnsigned64At: 1] bench. } Here are the results: 'cbc smallN write'->'3,110,000 per second. 322 nanoseconds per run.' . 'ma smallN write'->'4,770,000 per second. 210 nanoseconds per run.' . 'cbc smallN access'->'4,300,000 per second. 233 nanoseconds per run.' . 'ma smallN access'->'16,400,000 per second. 60.9 nanoseconds per run.' . 'cbc largeN write'->'907,000 per second. 1.1 microseconds per run.' . 'ma largeN write'->'6,620,000 per second. 151 nanoseconds per run.' . 'cbc largeN access'->'1,900,000 per second. 527 nanoseconds per run.' . 'ma largeN access'->'1,020,000 per second. 982 nanoseconds per run.' It looks like your 64-bit access is 86% faster for accessing the high-end of the 64-bit range, but slower in the other 3 metrics. Noticeably, it was only 14% as fast for writing the high-end of the 64-bit range, and similarly as much slower for small-number access.. On Fri, Aug 28, 2015 at 6:01 PM, Chris Cunningham <cunningham.cb@gmail.com <mailto:cunningham.cb@gmail.com> <mailto:cunningham.cb@gmail.com <mailto:cunningham.cb@gmail.com>>> wrote: > Hi. > > I've committed a change to the inbox with changes to allow getting/putting > 64bit values to ByteArrays (similar to 32 and 16 bit accessors). Could this > be added to trunk? > > Also, first time I used the selective commit function - very nice! the > changes I didn't want committed didn't, in fact, get commited. Just the > desirable bits! > > -cbc > > > -- _,,,^..^,,,_ best, Eliot .