From 38ed6dd995f9c7c0e6e00d5aeff56964978ec774 Mon Sep 17 00:00:00 2001
From: margarita0303 <lashina.2000@gmail.com>
Date: Tue, 2 Aug 2022 20:59:11 +0300
Subject: [PATCH] added benchmarks with different sizes, added check for
 accuracy

---
 .../kscience/kmath/benchmarks/SVDBenchmark.kt | 117 +++++++++++++++++-
 1 file changed, 112 insertions(+), 5 deletions(-)

diff --git a/benchmarks/src/jvmMain/kotlin/space/kscience/kmath/benchmarks/SVDBenchmark.kt b/benchmarks/src/jvmMain/kotlin/space/kscience/kmath/benchmarks/SVDBenchmark.kt
index 6879960d7..ee581f778 100644
--- a/benchmarks/src/jvmMain/kotlin/space/kscience/kmath/benchmarks/SVDBenchmark.kt
+++ b/benchmarks/src/jvmMain/kotlin/space/kscience/kmath/benchmarks/SVDBenchmark.kt
@@ -8,27 +8,134 @@ import kotlinx.benchmark.Benchmark
 import kotlinx.benchmark.Blackhole
 import kotlinx.benchmark.Scope
 import kotlinx.benchmark.State
+import org.ejml.UtilEjml.assertTrue
+import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.diagonalEmbedding
+import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.dot
+import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.eq
 import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.svdGolubKahan
+import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.transpose
 import space.kscience.kmath.tensors.core.DoubleTensorAlgebra
 import space.kscience.kmath.tensors.core.DoubleTensorAlgebra.Companion.svdPowerMethod
 
 @State(Scope.Benchmark)
 class SVDBenchmark {
     companion object {
-        val tensor = DoubleTensorAlgebra.randomNormal(intArrayOf(10, 10, 10), 0)
+        val tensorSmall = DoubleTensorAlgebra.randomNormal(intArrayOf(5, 5), 0)
+        val tensorMedium = DoubleTensorAlgebra.randomNormal(intArrayOf(10, 10), 0)
+        val tensorLarge = DoubleTensorAlgebra.randomNormal(intArrayOf(50, 50), 0)
+        val tensorVeryLarge = DoubleTensorAlgebra.randomNormal(intArrayOf(100, 100), 0)
+        val epsilon = 1e-9
     }
 
     @Benchmark
-    fun svdPowerMethod(blackhole: Blackhole) {
+    fun svdPowerMethodSmall(blackhole: Blackhole) {
+        val svd = tensorSmall.svdPowerMethod()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorSmall, epsilon))
         blackhole.consume(
-            tensor.svdPowerMethod()
+            tensorSmall.svdPowerMethod()
         )
     }
 
     @Benchmark
-    fun svdGolubKahan(blackhole: Blackhole) {
+    fun svdPowerMethodMedium(blackhole: Blackhole) {
+        val svd = tensorMedium.svdPowerMethod()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorMedium, epsilon))
         blackhole.consume(
-            tensor.svdGolubKahan()
+            tensorMedium.svdPowerMethod()
+        )
+    }
+
+    @Benchmark
+    fun svdPowerMethodLarge(blackhole: Blackhole) {
+        val svd = tensorLarge.svdPowerMethod()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorLarge, epsilon))
+        blackhole.consume(
+            tensorLarge.svdPowerMethod()
+        )
+    }
+
+    @Benchmark
+    fun svdPowerMethodVeryLarge(blackhole: Blackhole) {
+        val svd = tensorVeryLarge.svdPowerMethod()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorVeryLarge, epsilon))
+        blackhole.consume(
+            tensorVeryLarge.svdPowerMethod()
+        )
+    }
+
+    @Benchmark
+    fun svdGolubKahanSmall(blackhole: Blackhole) {
+        val svd = tensorSmall.svdGolubKahan()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorSmall, epsilon))
+        blackhole.consume(
+            tensorSmall.svdGolubKahan()
+        )
+    }
+
+    @Benchmark
+    fun svdGolubKahanMedium(blackhole: Blackhole) {
+        val svd = tensorMedium.svdGolubKahan()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorMedium, epsilon))
+        blackhole.consume(
+            tensorMedium.svdGolubKahan()
+        )
+    }
+
+    @Benchmark
+    fun svdGolubKahanLarge(blackhole: Blackhole) {
+        val svd = tensorLarge.svdGolubKahan()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorLarge, epsilon))
+        blackhole.consume(
+            tensorLarge.svdGolubKahan()
+        )
+    }
+
+    @Benchmark
+    fun svdGolubKahanVeryLarge(blackhole: Blackhole) {
+        val svd = tensorVeryLarge.svdGolubKahan()
+        val tensorSVD = svd.first
+            .dot(
+                diagonalEmbedding(svd.second)
+                    .dot(svd.third.transpose())
+            )
+        assertTrue(tensorSVD.eq(tensorVeryLarge, epsilon))
+        blackhole.consume(
+            tensorVeryLarge.svdGolubKahan()
         )
     }
 }
\ No newline at end of file