Initial commit: E4B-MarkBase model integration with passing tests

- E4B-MarkBase model (42 layers, 4.4GB) loaded successfully - All Phase 1-6 tests passed (model loading, forward pass, vision/audio towers, token generation, performance) - All stress tests passed (5/5 in 127.6s) - Concurrent inference - Memory stress (67.5 tok/s, 0 NaN) - Continuous generation - Batch processing - Long-running stability - Swift Metal inference engine with multimodal support
2026-06-23 18:12:35 +08:00
commit ac75faa0cc
301 changed files with 63426 additions and 0 deletions
@@ -0,0 +1,50 @@
+import XCTest
+@testable import MarkBase
+
+final class AudioGPUTest: XCTestCase {
+    func testGPUvsCPU() throws {
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let extractor = AudioFeatureExtractor(sampleRate: 16000, nMels: 128, nFft: 400, hopLength: 160)
+
+        var audioData = [Float](repeating: 0, count: 16000)
+        for i in 0..<audioData.count {
+            audioData[i] = sin(2.0 * Float.pi * 440.0 * Float(i) / 16000.0) * 0.5
+        }
+
+        let startCPU = CFAbsoluteTimeGetCurrent()
+        let cpuResult = extractor.extractMelSpectrogram(from: audioData)
+        let cpuTime = CFAbsoluteTimeGetCurrent() - startCPU
+
+        let startGPU = CFAbsoluteTimeGetCurrent()
+        let gpuResult = try extractor.extractMelSpectrogramGPU(engine: engine, audioData: audioData)
+        let gpuTime = CFAbsoluteTimeGetCurrent() - startGPU
+
+        print("CPU: \(cpuResult.count)×\(cpuResult[0].count) in \(String(format: "%.1f", cpuTime * 1000))ms")
+        print("GPU: \(gpuResult.count)×\(gpuResult[0].count) in \(String(format: "%.1f", gpuTime * 1000))ms")
+        print("Speedup: \(String(format: "%.1f", cpuTime / gpuTime))×")
+
+        // Compare in linear scale (log10 amplifies noise-floor differences)
+        var maxLinearDiff: Float = 0
+        var maxRelativeDiff: Float = 0
+        var worstF = 0, worstM = 0
+        for f in 0..<cpuResult.count {
+            for m in 0..<cpuResult[0].count {
+                let cLin = pow(10, cpuResult[f][m])
+                let gLin = pow(10, gpuResult[f][m])
+                let lDiff = abs(cLin - gLin)
+                maxLinearDiff = max(maxLinearDiff, lDiff)
+                if cLin > 1e-8 {
+                    let relDiff = lDiff / cLin
+                    if relDiff > maxRelativeDiff {
+                        maxRelativeDiff = relDiff
+                        worstF = f; worstM = m
+                    }
+                }
+            }
+        }
+        print("Max linear diff: \(maxLinearDiff)")
+        print("Max relative diff: \(maxRelativeDiff) at frame=\(worstF) mel=\(worstM)")
+        print("  CPU(lin)=\(pow(10, cpuResult[worstF][worstM])) GPU(lin)=\(pow(10, gpuResult[worstF][worstM]))")
+        XCTAssertLessThan(maxLinearDiff, 1e-3, "GPU mel linear values should match CPU")
+    }
+}