Hi Roger, Thanks for your reply and thoughts! Comments interspersed below:
On 8/28/20 10:54 AM, Roger Riggs wrote:
Hi Corey, A few comments on core-libs side... The naming convention for methods that end in '0' is usually to indicate they are the bottom-most method or a native method. So I think you can/should rename the methods to make the most sense as to their function.
Ok, I will fix that.
Comparing with the way that the Base64 encoder was intrinsified, the method that is intrinsified should have a method body that does the same function, so it is interchangable. That likely will just shift the "fast path" code into the decodeBlock method. Keeping the symmetry between encoder and decoder will make it easier to maintain the code.
Good point. I'll investigate what this looks like in terms of the actual code, and will report back (perhaps in a new webrev).
Given intrinsic only handles 2 of the three cases, and the java code handles all three, I would add an extra arg to decodeBlock to reflect the isMime caseand have the intrinsic take an early exit until it was implemented.
I did consider doing that, but didn't for two reasons:* Implementing isMIME using vector hardware would be very difficult due to the need to ignore non-base64 characters. This requires eliminating those characters from the vector, then reading and shifting more in, repeatedly until there are no non-base64 characters left. This isn't a trivial/fast thing to do, at least on Power arch. None of the published base64 encode/decode functions for vector processors address the MIME case. In fact they don't address isURL=true either, but fortunately that is a relatively easy addition.
* If isMIME=true is not implemented by the intrinsic, it will cost unnecessary overhead for that case, because of the need to martial the parameters, call the intrinsic, and then do an early return. I benchmarked this approach before, and saw an approx 5% drop in performance when isMIME = true. So that's why we decided to leave the isMIME=true case as a later optimization. Because of the extra complexity of the algorithm, it probably shouldn't share the same intrinsic anyway; only the isMIME=true case should take the performance hit.
It is unfortunate that taking advantage of vectorization has to be hand coded.If/when the Vector API is ready (JEP 338 https://openjdk.java.net/jeps/338) the java code should be replaced to use the Vector API and then it would work for a new hardware without specific coding for each platform. "Just" implement the Vector API. There's a lot more bang for the buck going for that approach.
The kind of vector processing used in this intrinsic operates mostly on bytes within one vector, not between two vectors (for example in matrix-multiply algorithms), otherwise known as SWAR (https://en.wikipedia.org/wiki/SWAR). Because of that, it's very sensitive to which exact instructions are available in the vector processor. There isn't much standardization of SWAR instructions between different arches, so I think it would be hard to get a generic SWAR API that gives good performance across several arches. From briefly looking at the link you provided, it doesn't appear to address SWAR operations, so it doesn't seem to me that waiting for the vector API would be worth the wait, and in fact may not provide any method at all to boost performance of base64 decode/encode.
Regards, - CoreyP.S. I work only two days a week, so the updates will be slower compared to other developers.
Thanks, Roger On 8/24/20 9:21 PM, Corey Ashford wrote:Here's a revised webrev which includes a JMH benchmark for the decode operation.http://cr.openjdk.java.net/~mhorie/8248188/webrev.03/The added benchmark tries to be "fair" in that it doesn't prefer a large buffer size, which would favor the intrinsic. It pseudo-randomly (but reproducibly) chooses a buffer size between 8 and 20k+8 bytes, and fills it with random data to encode and decode. As part of the TearDown of an invocation, it also checks the decoded output data for correctness.Example runs on the Power9-based machine I use for development shows a 3X average improvement across these random buffer sizes. Here's an excerpt of the output when run with -XX:-UseBASE64Intrinsics :Iteration 1: 70795.623 ops/s Iteration 2: 71070.607 ops/s Iteration 3: 70867.544 ops/s Iteration 4: 71107.992 ops/s Iteration 5: 71048.281 ops/s And here's the output with the intrinsic enabled: Iteration 1: 208794.022 ops/s Iteration 2: 208630.904 ops/s Iteration 3: 208238.822 ops/s Iteration 4: 208714.967 ops/s Iteration 5: 209060.894 ops/s Taking the best of the two runs: 209060/71048 = 2.94From other experiments where the benchmark uses a fixed-size, larger buffer, the performance ratio rises to about 4.0.Power10 should have a slightly higher ratio due to several factors, but I have not yet benchmarked on Power10.Other arches ought to be able to do at least this well, if not better, because of wider vector registers (> 128 bits) being available. Only a Power9/10 implementation is included in this webrev, however.Regards, - Corey On 8/19/20 11:20 AM, Roger Riggs wrote:Hi Corey,For changes obviously performance motivated, it is conventional to run a JMH perf test to demonstatethe improvement and prove it is worthwhile to add code complexity.I don't see any existing Base64 JMH tests but they would be in the repo below or near:test/micro/org/openjdk/bench/java/util/ Please contribute a JMH test and results to show the difference. Regards, Roger On 8/19/20 2:10 PM, Corey Ashford wrote:Michihiro Horie posted up a new iteration of this webrev for me. This time the webrev includes a complete implementation of the intrinsic for Power9 and Power10.You can find it here: http://cr.openjdk.java.net/~mhorie/8248188/webrev.02/Changes in webrev.02 vs. webrev.01:* The method header for the intrinsic in the Base64 code has been rewritten using the Javadoc style. The clarity of the comments has been improved and some verbosity has been removed. There are no additional functional changes to Base64.java.* The code needed to martial and check the intrinsic parameters has been added, using the base64 encodeBlock intrinsic as a guideline.* A complete intrinsic implementation for Power9 and Power10 is included.* Adds some Power9 and Power10 assembler instructions needed by the intrinsic which hadn't been defined before.The intrinsic implementation in this patch accelerates the decoding of large blocks of base64 data by a factor of about 3.5X on Power9.I'm attaching two Java test cases I am using for testing and benchmarking. The TestBase64_VB encodes and decodes randomly-sized buffers of random data and checks that original data matches the encoded-then-decoded data. TestBase64Errors encodes a 48K block of random bytes, then corrupts each byte of the encoded data, one at a time, checking to see if the decoder catches the illegal byte.Any comments/suggestions would be appreciated. Thanks, - Corey On 7/27/20 6:49 PM, Corey Ashford wrote:Michihiro Horie uploaded a new revision of the Base64 decodeBlock intrinsic API for me:http://cr.openjdk.java.net/~mhorie/8248188/webrev.01/ It has the following changes with respect to the original one posted:* In the event of encountering a non-base64 character, instead of having a separate error code of -1, the intrinsic can now just return either 0, or the number of data bytes produced up to the point where the illegal base64 character was encountered. This reduces the number of special cases, and also provides a way to speed up the process of finding the bad character by the slower, pure-Java algorithm.* The isMIME boolean is removed from the API for two reasons:- The current API is not sufficient to handle the isMIME case, because there isn't a strict relationship between the number of input bytes and the number of output bytes, because there can be an arbitrary number of non-base64 characters in the source. - If an intrinsic only implements the (isMIME == false) case as ours does, it will always return 0 bytes processed, which will slightly slow down the normal path of processing an (isMIME == true) instantiation. - We considered adding a separate hotspot candidate for the (isMIME == true) case, but since we don't have an intrinsic implementation to test that, we decided to leave it as a future optimization.Comments and suggestions are welcome. Thanks for your consideration. - Corey On 6/23/20 6:23 PM, Michihiro Horie wrote:Hi Corey, Following is the issue I created. https://bugs.openjdk.java.net/browse/JDK-8248188 I will upload a webrev when you're ready as we talked in private. Best regards, MichihiroInactive hide details for "Corey Ashford" ---2020/06/24 09:40:10---Currently in java.util.Base64, there is a HotSpotIntrinsicCa"Corey Ashford" ---2020/06/24 09:40:10---Currently in java.util.Base64, there is a HotSpotIntrinsicCandidate and API for encodeBlock, but noFrom: "Corey Ashford" <cjash...@linux.ibm.com>To: "hotspot-compiler-...@openjdk.java.net" <hotspot-compiler-...@openjdk.java.net>, "ppc-aix-port-...@openjdk.java.net" <ppc-aix-port-...@openjdk.java.net> Cc: Michihiro Horie/Japan/IBM@IBMJP, Kazunori Ogata/Japan/IBM@IBMJP, jos...@br.ibm.comDate: 2020/06/24 09:40Subject: RFR(S): [PATCH] Add HotSpotIntrinsicCandidate and API for Base64 decoding------------------------------------------------------------------------Currently in java.util.Base64, there is a HotSpotIntrinsicCandidate andAPI for encodeBlock, but none for decoding. This means that only encoding gets acceleration from the underlying CPU's vector hardware. I'd like to propose adding a new intrinsic for decodeBlock. The considerations I have for this new intrinsic's API: * Don't make any assumptions about the underlying capability of thehardware. For example, do not impose any specific block size granularity.* Don't assume the underlying intrinsic can handle isMIME or isURLmodes, but also let them decide if they will process the data regardlessof the settings of the two booleans. * Any remaining data that is not processed by the intrinsic will beprocessed by the pure Java implementation. This allows the intrinsic toprocess whatever block sizes it's good at without the complexity of handling the end fragments.* If any illegal character is discovered in the decoding process, theintrinsic will simply return -1, instead of requiring it to throw a proper exception from the context of the intrinsic. In the event ofgetting a -1 returned from the intrinsic, the Java Base64 library code simply calls the pure Java implementation to have it find the error andproperly throw an exception. This is a performance trade-off in the case of an error (which I expect to be very rare). * One thought I have for a further optimization (not implemented inthe current patch), is that when the intrinsic decides not to process ablock because of some combination of isURL and isMIME settings itdoesn't handle, it could return extra bits in the return code, encodedas a negative number. For example: Illegal_Base64_char = 0b001; isMIME_unsupported = 0b010; isURL_unsupported = 0b100;These can be OR'd together as needed and then negated (flip the sign).The Base64 library code could then cache these flags, so it will know not to call the intrinsic again when another decodeBlock is requested but with an unsupported mode. This will save the performance hit of calling the intrinsic when it is guaranteed to fail. I've tested the attached patch with an actual intrinsic coded up forPower9/Power10, but those runtime intrinsics and arch-specific patchesaren't attached today. I want to get some consensus on the library-level intrinsic API first.Also attached is a simple test case to test that the new intrinsic APIdoesn't break anything. I'm open to any comments about this. Thanks for your consideration, - Corey Corey Ashford IBM Systems, Linux Technology Center, OpenJDK team cjashfor at us dot ibm dot com[attachment "decodeBlock_api-20200623.patch" deleted by Michihiro Horie/Japan/IBM] [attachment "TestBase64.java" deleted by Michihiro Horie/Japan/IBM]
