Benchmark some CPU vs GPU 7-card poker hand evaluators on iOS and macOS.
All unit tests use XCTest. Tests are split into macOS and iOS testing targets, but are built from the same Swift test files.
macOS unit tests are performed with the standalone Showdown_macOSTests target.
iOS unit tests on an iOS device require an app testing target, so a minimal empty Showdown_iOS app is included as the testing target for the Showdown_iOSTests target. Metal tests will fail in the iOS simulator.
Each evaluator is encapsulted in a Swift framework subproject. Framework targets build for both iOS and macOS architectures.
SpecialK is a C++ evaluator that uses a brute force precalculated perfect hash with 136 kB of lookup tables as described here.
This SKPokerEval framework is Swift wrapper around a C implementation of the SpecialK v1.9 evaluator. Evaluation is performed on a single CPU thread.
import SKPokerEval
let hands = randomSKHands(count: 10_000)
// With Swift arrays
let scores = SKPokerEval.score(hands: hands)
// With pointers
let scoreBuffer = UnsafeMutableBufferPointer<UInt16>
.allocate(capacity: hands.count)
score(handPointer: UnsafePointer<UInt64>(hands),
scorePointer: scoreBuffer.baseAddress!,
count: skHands.count)A Swift convenience class for a Metal implementation of the SpecialK v1.9 evaluator. An initialized Evaluator object loads the Metal pipeline and copies the lookup tables to Metal buffers. Evaluation is performed in parallel at the execution width of the default Metal GPU.
import SKMetalPokerEval
let hands = randomSKHands(count: 10_000)
let evaluator = Evaluator()!
// With Swift arrays
let scores = evaluator.score(hands: hands)
// With Metal buffers
let handBuffer = evaluator.makeHandBuffer(hands: hands)!
let scoreBuffer = evaluator.makeScoreBuffer(count: hands.count)!
let commandBuffer = evaluator.dispatchScoreCommand(
handBuffer: handBuffer,
scoreBuffer: scoreBuffer,
count: skHands.count
)!
commandBuffer.waitUntilCompleted()A Swift wrapper around a bit twiddling (lookup tableless) evaluator implemented in C. Evaluation is performed on a single CPU thread.
import BTPokerEval
let hands = randomBTHands(count: 10_000)
// With Swift arrays
let scores = SKPokerEval.score(hands: hands)
// With pointers
let scoreBuffer = UnsafeMutableBufferPointer<UInt16>
.allocate(capacity: hands.count)
score(handPointer: UnsafePointer<UInt64>(hands),
scorePointer: scoreBuffer.baseAddress!,
count: skHands.count)A Swift convenience class for a Metal implementation of BTPokerEval. An initialized Evaluator object loads the Metal pipeline. Evaluation is performed in parallel at the execution width of the default Metal GPU.
import BTMetalPokerEval
let hands = randomBTHands(count: 10_000)
let evaluator = Evaluator()!
// With Swift arrays
let scores = evaluator.score(hands: hands)
// With Metal buffers
let handBuffer = evaluator.makeHandBuffer(hands: hands)!
let scoreBuffer = evaluator.makeScoreBuffer(count: hands.count)!
let commandBuffer = evaluator.dispatchScoreCommand(
handBuffer: handBuffer,
scoreBuffer: scoreBuffer,
count: skHands.count
)!
commandBuffer.waitUntilCompleted()Contributions are welcome, including improvements to the benchmarking code, individual evaluator performance enhancements, and addition of other evaluators.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent