feat: initial release

Assisted-by: GLM 5.3 Flash

signal/conv_bench_test.go

@@ -0,0 +1,48 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: MIT
+
+package signal
+
+import (
+	"fmt"
+	"testing"
+)
+
+// Channel-blocked kernel benchmarks at inference-like sizes. The
+// channel count is the axis the kernels divide work along, so each
+// size climbs a cache level: 4 channels stay in L1, 32 cross L2 and
+// 128 make the whole-volume walk pay for memory bandwidth.
+
+// BenchmarkConv3D measures a (1, C, 64, 64, 64) input through C×C
+// kernels of 3×3×3, stride 1, padding 1.
+func BenchmarkConv3D(b *testing.B) {
+	for _, ch := range []int{4, 32, 128} {
+		b.Run(fmt.Sprintf("%d channels", ch), func(b *testing.B) {
+			input := benchSig(b, 1, ch*64*64*64, 1, ch, 64, 64, 64)
+			kernel := benchSig(b, 2, ch*ch*27, ch, ch, 3, 3, 3)
+			b.ReportAllocs()
+			for b.Loop() {
+				if _, err := Conv3D(input, kernel, nil, 1, [3]int{1, 1, 1}, [3]int{1, 1, 1}); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
+// BenchmarkConvTranspose2D measures a (1, C, 128, 128) input through
+// C×C kernels of 3×3 at the upsampling stride of 2, padding 1.
+func BenchmarkConvTranspose2D(b *testing.B) {
+	for _, ch := range []int{4, 32, 128} {
+		b.Run(fmt.Sprintf("%d channels", ch), func(b *testing.B) {
+			input := benchSig(b, 3, ch*128*128, 1, ch, 128, 128)
+			kernel := benchSig(b, 4, ch*ch*9, ch, ch, 3, 3)
+			b.ReportAllocs()
+			for b.Loop() {
+				if _, err := ConvTranspose2D(input, kernel, nil, 2, 1); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}