Modernisation/Performance boost POC #39

TimothyMakkison · 2022-12-21T23:29:26Z

Partially resolving #38 , #37 #35 . Modernised and implemented interfaces to be more like Guid, optimized code using intrinsics and bit hacks.

Implemented IComparable.
Implemented ISpanParsable and ISpanParsable<Ulid>.
Optimised get Time by 30%.
Vectorised and optimised public Ulid(Guid) for a 80% and 20% speed up.
Added vector equality checks and modern unsafe code for equality checks for a 20-30% speedup.
Modernised GetHashCode to use modern unsafe code.
Vectorised ToGuid and added bit hack optimization for a 80-90% and 20% speed up.
Added .Net 7.0
Benchmarks used can be found here.

I've updated Ulid, PerfBenchmarks and Ulid,Tests to support .Net 7.0, this is to add ISpanParsable and modern Vector methods. All changes should be back compatible with older versions of dotnet - modern code is added with the #if pre processor directives.

A lot of intrinsics code appears to be repeated for .Net 3+ and for .Net 7, this is deliberate as the cross platform Vector128.Shuffle was only added in .Net 7. This way Ssse compatible devices will get the benefit of vectorisation in most modern versions of .Net.

Because of JIT related issues (I think) if .Net 3 vector code doesn't constant vectors outside of method bodies for performance reasons. Likewise .Net 7 vector code should only use constant vectors placed inside its method body, again for performance reasons. I'm not sure why this happens (might be similar to this or this), all I know is that when I use a shuffle indices vector the wrong way the performance takes a hit. I'd be curious to see if other people experience the same performance improvements I've had.

Note that my benchmarks may not be accurate, the times are in the nano seconds and vary quite a bit. For some benchmarks I've added a loop to get more reliable numbers, but due to the magic of software these numbers are probably not representative. I've labeled these results accordingly. Some of that changes may not be significant and It's possible that I've added performance regressions for other versions of .Net or devices.

Finally please note that I'm still unexperienced with unsafe code and vectorisation. It's likely that I've made a mistake or overlooked key points. ie memory alignment, pointer safety, architecture compatibility, Unsafe.As vs Unsafe.As(Unsafe.AsRef), accidental mutations.

Implemented

IComparable

Implemented IComparable, returns 1 if null, raises an exception if not an Ulid, otherwise the default CompareTo method is used. I also added a comparison unit test.

ISpanFormattable & ISpanParsable

Implemented both interfaces for their respective versions with additional unit tests. I'm not familiar with these interfaces (can't find any docs/articles) and treated the methods as wrappers for TryWriteStringify/ToString and Parse/TryParse, ignoring IFormatProvider and ReadOnlySpan<char> format.
Looking at the Guid internals they ignored IFormatProvider but would change the serialization method depending on values inside ReadOnlySpan<char> format. I assumed that it isn't mandatory for implementors to use these parameters. Perhaps you could consider custom formatting options here.

Changed

Time

Updated the Time property to use Unsafe code to reverse the first 6 bytes. I kept the original method for big endian devices.
Benchmarking shows a 30% speedup with times equal to or faster than vectorisation so I opted to only use byte shuffling.

Method	Runtime	Mean	Error	StdDev	Gen0	Allocated
UlidUpdated_	.NET Core 3.1	18.22 ns	0.289 ns	0.075 ns	-	-
UlidVector_	.NET Core 3.1	17.64 ns	0.088 ns	0.023 ns	-	-
UlidOriginal_	.NET Core 3.1	23.48 ns	0.132 ns	0.034 ns	-	-
NUlid_	.NET Core 3.1	34.48 ns	0.525 ns	0.081 ns	0.0408	64 B
UlidUpdated_	.NET 5.0	17.35 ns	0.342 ns	0.089 ns	-	-
UlidVector_	.NET 5.0	17.33 ns	0.079 ns	0.012 ns	-	-
UlidOriginal_	.NET 5.0	23.51 ns	0.094 ns	0.024 ns	-	-
NUlid_	.NET 5.0	32.94 ns	0.310 ns	0.048 ns	0.0408	64 B
UlidUpdated_	.NET 7.0	12.94 ns	0.179 ns	0.028 ns	-	-
UlidVector_	.NET 7.0	13.24 ns	0.163 ns	0.042 ns	-	-
UlidOriginal_	.NET 7.0	18.71 ns	0.189 ns	0.029 ns	-	-
NUlid_	.NET 7.0	29.44 ns	0.179 ns	0.046 ns	0.0408	64 B

Looped benchmark (100 iterations)

Method	Runtime	Mean	Error	StdDev	Gen0	Allocated
UlidUpdated_	.NET Core 3.1	1,970.66 ns	12.498 ns	1.934 ns	-	-
UlidVector_	.NET Core 3.1	1,910.51 ns	20.386 ns	5.294 ns	-	-
UlidOriginal_	.NET Core 3.1	2,531.84 ns	86.987 ns	13.461 ns	-	-
NUlid_	.NET Core 3.1	33.57 ns	0.466 ns	0.072 ns	0.0408	64 B
UlidUpdated_	.NET 5.0	1,872.80 ns	11.559 ns	3.002 ns	-	-
UlidVector_	.NET 5.0	1,908.42 ns	11.305 ns	1.749 ns	-	-
UlidOriginal_	.NET 5.0	2,499.66 ns	11.249 ns	2.921 ns	-	-
NUlid_	.NET 5.0	32.65 ns	0.585 ns	0.091 ns	0.0408	64 B
UlidUpdated_	.NET 7.0	1,478.76 ns	13.499 ns	3.506 ns	-	-
UlidVector_	.NET 7.0	1,476.74 ns	10.306 ns	1.595 ns	-	-
UlidOriginal_	.NET 7.0	2,105.35 ns	12.327 ns	3.201 ns	-	-
NUlid_	.NET 7.0	29.10 ns	0.501 ns	0.130 ns	0.0408	64 B

Constructor public Ulid Guid

Added/updated 3 methods for public Ulid(Guid guid). For .Net 7+ or .Net 3+ I used shuffle intrinsics to shuffle the structure of the guid and cast it into a Ulid. For little endian intrinsic incompatible devices I used Unsafe and BinaryPrimitives.ReverseEndianness to alter the Guid structure to Ulid.

Method	Runtime	Mean	Error	StdDev	Gen0	Allocated
UlidUpdatedBitwise_	.NET Core 3.1	9.794 ns	0.1290 ns	0.0200 ns	-	-
UlidUpdatedVec_	.NET Core 3.1	2.747 ns	0.0699 ns	0.0182 ns	-	-
UlidOriginal_	.NET Core 3.1	12.401 ns	0.0551 ns	0.0143 ns	-	-
NUlid_	.NET Core 3.1	19.591 ns	0.2255 ns	0.0586 ns	0.0255	40 B
UlidUpdatedBitwise_	.NET 5.0	9.548 ns	0.4433 ns	0.0686 ns	-	-
UlidUpdatedVec_	.NET 5.0	2.348 ns	0.0683 ns	0.0106 ns	-	-
UlidOriginal_	.NET 5.0	12.083 ns	0.0419 ns	0.0109 ns	-	-
NUlid_	.NET 5.0	18.995 ns	0.3381 ns	0.0878 ns	0.0255	40 B
UlidUpdatedBitwise_	.NET 7.0	9.015 ns	0.1410 ns	0.0366 ns	-	-
UlidUpdatedVec_	.NET 7.0	2.094 ns	0.1952 ns	0.0507 ns	-	-
UlidOriginal_	.NET 7.0	11.003 ns	0.3082 ns	0.0801 ns	-	-
NUlid_	.NET 7.0	17.700 ns	0.2000 ns	0.0519 ns	0.0255	40 B

Equality

Added a vectorized equality check using either .Net 7+ Vector128 extensions, a manual Ssse3 comparison for .Net 3.0-.Net 6.0 or a refactored unsafe equality check.

Method	Runtime	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	.NET Core 3.1	2.0128 ns	0.0352 ns	0.0091 ns	1.00	0.00	-	-	NA
UlidUnsafe_	.NET Core 3.1	1.8601 ns	0.0629 ns	0.0097 ns	0.93	0.01	-	-	NA
UlidVector_	.NET Core 3.1	2.0095 ns	0.1870 ns	0.0486 ns	1.00	0.03	-	-	NA
UlidOriginal_	.NET Core 3.1	2.5127 ns	0.0408 ns	0.0106 ns	1.25	0.01	-	-	NA
NUlid_	.NET Core 3.1	36.7950 ns	2.3330 ns	0.6059 ns	18.28	0.36	0.0510	80 B	NA

Guid_	.NET 5.0	1.9346 ns	0.0226 ns	0.0035 ns	1.00	0.00	-	-	NA
UlidUnsafe_	.NET 5.0	2.0288 ns	0.2671 ns	0.0694 ns	1.04	0.03	-	-	NA
UlidVector_	.NET 5.0	1.6987 ns	0.0409 ns	0.0063 ns	0.88	0.00	-	-	NA
UlidOriginal_	.NET 5.0	2.3973 ns	0.3028 ns	0.0786 ns	1.23	0.05	-	-	NA
NUlid_	.NET 5.0	37.4773 ns	4.3714 ns	1.1352 ns	19.23	0.60	0.0510	80 B	NA

Guid_	.NET 7.0	1.6386 ns	0.1061 ns	0.0276 ns	1.00	0.00	-	-	NA
UlidUnsafe_	.NET 7.0	0.4664 ns	0.0366 ns	0.0057 ns	0.29	0.00	-	-	NA
UlidVector_	.NET 7.0	0.3541 ns	0.0080 ns	0.0021 ns	0.22	0.00	-	-	NA
UlidOriginal_	.NET 7.0	1.1777 ns	0.6814 ns	0.1769 ns	0.72	0.12	-	-	NA
NUlid_	.NET 7.0	36.0873 ns	19.4051 ns	5.0394 ns	22.04	3.26	0.0510	80 B	NA

Looped benchmark (100 iterations)

Method	Runtime	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	.NET Core 3.1	2.034 ns	0.0409 ns	0.0106 ns	1.00	0.00	-	-	NA
UlidUnsafe_	.NET Core 3.1	242.943 ns	18.8697 ns	4.9004 ns	119.46	2.47	-	-	NA
UlidVector_	.NET Core 3.1	206.211 ns	1.3328 ns	0.3461 ns	101.40	0.41	-	-	NA
UlidOriginal_	.NET Core 3.1	337.801 ns	4.4441 ns	1.1541 ns	166.10	0.55	-	-	NA
NUlid_	.NET Core 3.1	34.217 ns	0.9053 ns	0.1401 ns	16.83	0.11	0.0510	80 B	NA

Guid_	.NET 5.0	1.951 ns	0.1234 ns	0.0191 ns	1.00	0.00	-	-	NA
UlidUnsafe_	.NET 5.0	297.125 ns	98.1249 ns	25.4827 ns	152.33	16.48	-	-	NA
UlidVector_	.NET 5.0	206.018 ns	2.8097 ns	0.4348 ns	105.62	0.83	-	-	NA
UlidOriginal_	.NET 5.0	305.997 ns	1.9817 ns	0.5146 ns	156.87	1.70	-	-	NA
NUlid_	.NET 5.0	34.712 ns	0.6706 ns	0.1741 ns	17.81	0.18	0.0510	80 B	NA

Guid_	.NET 7.0	1.636 ns	0.0610 ns	0.0094 ns	1.00	0.00	-	-	NA
UlidUnsafe_	.NET 7.0	62.873 ns	24.1450 ns	6.2704 ns	38.12	4.36	-	-	NA
UlidVector_	.NET 7.0	42.068 ns	0.6935 ns	0.1073 ns	25.71	0.18	-	-	NA
UlidOriginal_	.NET 7.0	144.580 ns	14.0945 ns	3.6603 ns	87.89	2.37	-	-	NA
NUlid_	.NET 7.0	32.020 ns	0.7219 ns	0.1875 ns	19.59	0.22	0.0510	80 B	NA

GetHashCode

Refactored the previous unsafe scoped implementation to use modern Unsafe code.

Method	Runtime	Mean	Error	StdDev	Median	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	.NET Core 3.1	1.4569 ns	1.5604 ns	0.4052 ns	1.3952 ns	1.00	0.00	-	-	NA
UlidUpdate_	.NET Core 3.1	0.4813 ns	1.0117 ns	0.2627 ns	0.3474 ns	0.37	0.24	-	-	NA
UlidOriginal_	.NET Core 3.1	1.0379 ns	0.5577 ns	0.0863 ns	0.9981 ns	0.77	0.17	-	-	NA
NUlid_	.NET Core 3.1	22.5968 ns	0.6576 ns	0.1708 ns	22.6603 ns	16.47	4.40	0.0255	40 B	NA

Guid_	.NET 7.0	0.0059 ns	0.0178 ns	0.0046 ns	0.0041 ns	1.00	0.00	-	-	NA
UlidUpdate_	.NET 7.0	0.0055 ns	0.0053 ns	0.0008 ns	0.0053 ns	3.73	5.33	-	-	NA
UlidOriginal_	.NET 7.0	0.0679 ns	0.0232 ns	0.0036 ns	0.0693 ns	46.73	65.15	-	-	NA
NUlid_	.NET 7.0	20.6823 ns	0.5543 ns	0.1440 ns	20.6432 ns	11,398.66	17,269.11	0.0255	40 B	NA

ToString

Used string.Create this lets TryWriteStringify directly mutate the string during constructions, before it is made immutable.

ToGuid

Added/updated 3 methods for public Ulid(Guid guid). For .Net 7+ or .Net 3+ I used shuffle intrinsics to shuffle the structure of the Ulid and cast it into a Guid. For little endian intrinsic incompatible devices I used Unsafe and BinaryPrimitives.ReverseEndianness to alter the Ulid structure to Guid.

Method	Runtime	Mean	Error	StdDev	Gen0	Allocated
UlidUpdatedBitwise_	.NET Core 3.1	7.5734 ns	0.0181 ns	0.0028 ns	-	-
UlidUpdatedVec_	.NET Core 3.1	1.2751 ns	0.0302 ns	0.0047 ns	-	-
UlidOriginal_	.NET Core 3.1	9.8598 ns	0.0609 ns	0.0158 ns	-	-
NUlid_	.NET Core 3.1	17.2219 ns	0.3480 ns	0.0539 ns	0.0255	40 B
UlidUpdatedBitwise_	.NET 5.0	7.1705 ns	0.0137 ns	0.0021 ns	-	-
UlidUpdatedVec_	.NET 5.0	1.0241 ns	0.0296 ns	0.0077 ns	-	-
UlidOriginal_	.NET 5.0	9.7610 ns	0.0179 ns	0.0028 ns	-	-
NUlid_	.NET 5.0	17.2024 ns	0.2036 ns	0.0529 ns	0.0255	40 B
UlidUpdatedBitwise_	.NET 7.0	6.6813 ns	0.0260 ns	0.0067 ns	-	-
UlidUpdatedVec_	.NET 7.0	0.6562 ns	0.0389 ns	0.0060 ns	-	-
UlidOriginal_	.NET 7.0	8.6386 ns	0.0742 ns	0.0115 ns	-	-
NUlid_	.NET 7.0	18.0501 ns	0.2538 ns	0.0393 ns	0.0255	40 B

Method	EnvironmentVariables	Runtime	Mean	Error	StdDev	Median	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	COMPlus_EnableSSE2=0	.NET 6.0	44.831 ns	0.8923 ns	0.8346 ns	44.760 ns	1.00	0.00	-	-	NA
Ulid_	COMPlus_EnableSSE2=0	.NET 6.0	6.925 ns	0.0585 ns	0.0547 ns	6.914 ns	0.15	0.00	-	-	NA
NUlid_	COMPlus_EnableSSE2=0	.NET 6.0	19.132 ns	0.1224 ns	0.1145 ns	19.136 ns	0.43	0.01	0.0255	40 B	NA

Guid_	COMPlus_EnableSSE2=0	.NET 7.0	52.820 ns	0.3402 ns	0.2841 ns	52.692 ns	1.00	0.00	-	-	NA
Ulid_	COMPlus_EnableSSE2=0	.NET 7.0	6.583 ns	0.0428 ns	0.0379 ns	6.585 ns	0.12	0.00	-	-	NA
NUlid_	COMPlus_EnableSSE2=0	.NET 7.0	20.839 ns	0.4598 ns	0.7683 ns	20.684 ns	0.39	0.01	0.0255	40 B	NA

Guid_	Empty	.NET 6.0	46.470 ns	0.9681 ns	0.9056 ns	46.405 ns	1.00	0.00	-	-	NA
Ulid_	Empty	.NET 6.0	1.766 ns	0.0562 ns	0.0469 ns	1.765 ns	0.04	0.00	-	-	NA
NUlid_	Empty	.NET 6.0	19.523 ns	0.2070 ns	0.1835 ns	19.518 ns	0.42	0.01	0.0255	40 B	NA

Guid_	Empty	.NET 7.0	53.839 ns	0.9657 ns	0.8064 ns	53.778 ns	1.00	0.00	-	-	NA
Ulid_	Empty	.NET 7.0	1.513 ns	0.0778 ns	0.1039 ns	1.496 ns	0.03	0.00	-	-	NA
NUlid_	Empty	.NET 7.0	18.989 ns	0.4115 ns	0.8769 ns	18.490 ns	0.37	0.02	0.0255	40 B	NA

Testing

To support vectorised/non vectorised code a way of testing with and without hardware intrinsics enabled should be added. I'm not familiar with github actions but perhaps the following could be added to build-debug.yml.

- run: dotnet test -c Debug  --environment "COMPlus_EnableSSE2=0"
- run: dotnet test -c Debug  --environment "COMPlus_EnableHWIntrinsic=0;COMPlus_EnableSSE2=0"

Future changes/ Thoughts

Modernise the code to use C# 11 - use file scoped namespaces, simplify slices.
Enable nullable across the project and update relevant methods.
Consider adding comparison operators <, <=, >, >=.
Add 'NotNullWhen' attributes to methods such as 'TryParse'.

public string ToBase64 uses ArrayPool, stackalloc could instead be used here. The SkipLocalsInit could be used to remove the overhead of clearing the span.

Would the logic used in the original get Time and internal Ulid(long timestampMilliseconds... work on big endian machines?

Little endian
TimestampMilliseconds -> Ulid first 6 bytes
|A|B|C|D|E|F|0|0|              -> |F|E|D|C|B|A|

Big endian
TimestampMilliseconds -> Ulid first 6 bytes
|0|0|F|E|D|C|B|A|              -> |F|E|D|C|0|0|

I have a working vectorized implementation of CrockfordBase32 that is compatible with the Ulid/Nulid implementation (Accepts inputs of 0-122, treats lowercase as uppercase, invalid characters become 31). It needs more work, decoding is 30% faster than the array version but encoding is still slower.

Experimental change to `internal Ulid(long,..)`

Internal Ulid long

Experimented with using Unsafe and BinaryPrimitives.ReverseEndianness to speed up internal constructors. Benchmarks showed little improvement.

Benchmark

Method	Runtime	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	.NET Core 3.1	89.71 ns	4.371 ns	0.676 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET Core 3.1	138.93 ns	56.850 ns	14.764 ns	1.58	0.18	-	-	NA
Ulid_	.NET Core 3.1	135.14 ns	1.101 ns	0.286 ns	1.51	0.01	-	-	NA
UlidOrginal_	.NET Core 3.1	130.13 ns	2.616 ns	0.679 ns	1.45	0.01	-	-	NA
NUlid_	.NET Core 3.1	272.83 ns	3.776 ns	0.584 ns	3.04	0.02	0.0663	104 B	NA

Guid_	.NET 5.0	87.75 ns	0.983 ns	0.255 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET 5.0	120.98 ns	0.612 ns	0.095 ns	1.38	0.00	-	-	NA
Ulid_	.NET 5.0	128.48 ns	10.173 ns	1.574 ns	1.47	0.02	-	-	NA
UlidOrginal_	.NET 5.0	122.75 ns	1.357 ns	0.352 ns	1.40	0.01	-	-	NA
NUlid_	.NET 5.0	264.45 ns	12.622 ns	3.278 ns	3.01	0.04	0.0663	104 B	NA

Guid_	.NET 7.0	86.09 ns	1.055 ns	0.163 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET 7.0	52.90 ns	0.735 ns	0.191 ns	0.61	0.00	-	-	NA
Ulid_	.NET 7.0	52.97 ns	0.182 ns	0.047 ns	0.62	0.00	-	-	NA
UlidOrginal_	.NET 7.0	54.26 ns	0.209 ns	0.032 ns	0.63	0.00	-	-	NA
NUlid_	.NET 7.0	188.91 ns	0.492 ns	0.076 ns	2.19	0.00	0.0663	104 B	NA

Looped benchmark (100 iterations)

Method	Runtime	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	.NET Core 3.1	88.18 ns	1.145 ns	0.297 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET Core 3.1	13,310.85 ns	2,060.225 ns	318.822 ns	150.88	3.98	-	-	NA
UlidOrginal_	.NET Core 3.1	13,086.56 ns	109.310 ns	28.387 ns	148.42	0.70	-	-	NA
NUlid_	.NET Core 3.1	276.40 ns	6.419 ns	1.667 ns	3.13	0.02	0.0663	104 B	NA

Guid_	.NET 5.0	88.72 ns	3.598 ns	0.934 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET 5.0	12,813.03 ns	430.020 ns	111.675 ns	144.42	1.14	-	-	NA
UlidOrginal_	.NET 5.0	12,600.82 ns	131.549 ns	20.357 ns	141.81	1.68	-	-	NA
NUlid_	.NET 5.0	264.17 ns	1.700 ns	0.263 ns	2.97	0.04	0.0663	104 B	NA

Guid_	.NET 7.0	85.85 ns	1.576 ns	0.244 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET 7.0	5,504.00 ns	15.282 ns	3.969 ns	64.13	0.15	-	-	NA
UlidOrginal_	.NET 7.0	5,644.55 ns	68.978 ns	10.674 ns	65.75	0.20	-	-	NA
NUlid_	.NET 7.0	191.18 ns	14.316 ns	2.215 ns	2.23	0.02	0.0663	104 B	NA

Method	Runtime	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
Guid_	.NET Core 3.1	89.83 ns	8.693 ns	2.257 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET Core 3.1	13,079.05 ns	229.835 ns	59.687 ns	145.66	3.54	-	-	NA
UlidOrginal_	.NET Core 3.1	13,201.82 ns	38.433 ns	9.981 ns	147.03	3.70	-	-	NA
NUlid_	.NET Core 3.1	273.02 ns	3.563 ns	0.925 ns	3.04	0.08	0.0663	104 B	NA

Guid_	.NET 5.0	88.77 ns	3.623 ns	0.941 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET 5.0	12,436.28 ns	65.738 ns	17.072 ns	140.11	1.63	-	-	NA
UlidOrginal_	.NET 5.0	12,628.54 ns	224.992 ns	58.430 ns	142.27	1.86	-	-	NA
NUlid_	.NET 5.0	264.49 ns	1.955 ns	0.508 ns	2.98	0.03	0.0663	104 B	NA

Guid_	.NET 7.0	85.78 ns	1.533 ns	0.237 ns	1.00	0.00	-	-	NA
UlidUpdated_	.NET 7.0	5,489.08 ns	36.487 ns	9.475 ns	64.00	0.14	-	-	NA
UlidOrginal_	.NET 7.0	5,744.08 ns	113.194 ns	17.517 ns	66.97	0.36	-	-	NA
NUlid_	.NET 7.0	191.33 ns	37.974 ns	9.862 ns	2.25	0.11	0.0663	104 B	NA

gfoidl

There's quite some code duplication.
Art the officially unsupported targets (pre .NET 6) worth to vectorized them?
Can the both Ulid.cs-implementations be unified? Maybe with the help of common helper-methods if not doable otherwise.

gfoidl · 2022-12-29T12:33:57Z

src/Ulid.Unity/Assets/Scripts/Ulid/Ulid.cs

@@ -155,21 +183,58 @@ internal Ulid(ReadOnlySpan<char> base32)
            randomness8 = (byte)((CharToBase32[base32[22]] << 7) | (CharToBase32[base32[23]] << 2) | (CharToBase32[base32[24]] >> 3));
        }

+#if NETCOREAPP3_0_OR_GREATER
+        private static readonly Vector128<byte> guidVecShuffle = Vector128.Create((byte)3, 2, 1, 0, 5, 4, 7, 6, 8, 9, 10, 11, 12, 13, 14, 15);


For newer targets (.NET 6.0 onwards) this constant should be given "inline" for allowing the JIT to emit better code. As at the moment only .NET 6+ is officially supported I wouldn't bother with .NET Core 3.1 optimizations anymore. So the below #if def can be dropped. For the Shuffle you could create a private method that handles this depending on the target.

TimothyMakkison · 2023-01-01T20:38:13Z

Art the officially unsupported targets (pre .NET 6) worth to vectorized them? As at the moment only .NET 6+ is officially supported I wouldn't bother with .NET Core 3.1 optimizations anymore.
There's quite some code duplication.

Good point, it probably isn't worth supporting them, I'll remove them unless the owner want to keep them. Should remove some bloat. I'll see if I can create one or two helper functions.

Could I ask you a question? From what I gathered from testing .Net 3, 3.1, 5 should have constants declared outside the method and .Net 7 should have constants delcared inline. Finally .Net 6 uniquely runs the same with constants declared outside or inline. Is this correct?

gfoidl · 2023-01-02T09:41:26Z

.NET 6 onwards should use the vector constants inline. The JIT has then the ability to optimize the usage. It streamlines the C# code also.
Before .NET 6 it was better to have the vector constants declared outside.
See dotnet/runtime#54827 (and the linked issue) for more info.

michelbieleveld · 2023-03-17T13:29:58Z

Hi, seems a lot of great progress has been made. Any chance this might get merged and a new nuget version produced ?

hadashiA · 2023-08-18T08:24:17Z

Sorry for the late reply and thanks for the great PR.
We have confirmed the changes and would be happy to merge them.
We would also like to adopt your suggestion about adding a test environment.
Thanks a lot :)

hadashiA · 2023-08-19T02:40:34Z

There seems to be some discussion left unanswered.

Modernise the code to use C# 11 - use file scoped namespaces, simplify slices.
Enable nullable across the project and update relevant methods.

Currently, we want the same code to work in Unity 2019.3 so we may not change it until we change the supported version.

Consider adding comparison operators <, <=, >, >=.
Add 'NotNullWhen' attributes to methods such as 'TryParse'.
public string ToBase64 uses ArrayPool, stackalloc could instead be used here. The SkipLocalsInit could be used to remove the overhead of clearing the span.

This is probably a reasonable change.
We would like to consider this.

Would the logic used in the original get Time and internal Ulid(long timestampMilliseconds... work on big endian machines?

Hmm, it looks like you are right. I would like to check with the original author.

Experimental change to internal Ulid(long,..)

Although we do not want to add too much complexity to the value generation for the sake of speed, BinaryPrimitives.ReverseEndianness may be reasonable.

Thanks for the suggestions. further pr is welcome.

TimothyMakkison added 17 commits December 18, 2022 18:25

Add .Net 7

1bec1b1

Add .Net 7+ support

21d3b0d

Edit get Time to use bitwise shuffling

41220a5

Correct fisrtByte typo to firstByte

4b03c36

Add Unsafe GetHashCode

22721c7

Edit ToGuid to use either intrinsics or a binary shuffle

d49b6aa

Add vectorized equality

5a11727

Add byte shuffle Ulid constructor

c76fd05

Vectorize and add byte shuffle to the Guid to Ulid constructor

1a51a77

Add GetHashCode test

779b196

Implement IComparable and comparison tests

3666413

Upgrade to C# 11

06b448a

Implement ISpanFormattable and ISpanParsable

16e85b4

Add ISpanFormattable and ISpanParsable tests

5ae3b68

Add mirror changess

f9db016

Add string.Create

9cb48a7

Remove experimental constructor and hoist constant vector

82ff62f

TimothyMakkison mentioned this pull request Dec 28, 2022

Possible improvements from Cysharp/Ulid RobThree/NUlid#23

Open

gfoidl reviewed Dec 29, 2022

View reviewed changes

Added shuffle method and targeted .Net 6+ only

3698799

hadashiA merged commit 4449a82 into Cysharp:master Aug 18, 2023

hadashiA mentioned this pull request Aug 18, 2023

Code modernisation #38

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Modernisation/Performance boost POC #39

Modernisation/Performance boost POC #39

TimothyMakkison commented Dec 21, 2022 •

edited

gfoidl left a comment

gfoidl Dec 29, 2022

TimothyMakkison commented Jan 1, 2023

gfoidl commented Jan 2, 2023

michelbieleveld commented Mar 17, 2023

hadashiA commented Aug 18, 2023

hadashiA commented Aug 19, 2023

Modernisation/Performance boost POC #39

Modernisation/Performance boost POC #39

Conversation

TimothyMakkison commented Dec 21, 2022 • edited

Implemented

IComparable

ISpanFormattable & ISpanParsable

Changed

Time

Looped benchmark (100 iterations)

Constructor public Ulid Guid

Equality

Looped benchmark (100 iterations)

GetHashCode

ToString

ToGuid

Testing

Future changes/ Thoughts

Internal Ulid long

Benchmark

Looped benchmark (100 iterations)

gfoidl left a comment

Choose a reason for hiding this comment

gfoidl Dec 29, 2022

Choose a reason for hiding this comment

TimothyMakkison commented Jan 1, 2023

gfoidl commented Jan 2, 2023

michelbieleveld commented Mar 17, 2023

hadashiA commented Aug 18, 2023

hadashiA commented Aug 19, 2023

TimothyMakkison commented Dec 21, 2022 •

edited