tensor/signal/conv_transpose_test.go

// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
// SPDX-License-Identifier: MIT

package signal

import (
	"math"
	"testing"
)

// convTranspose2DReference computes the transpose convolution straight
// from its scatter definition, each output element summing the input
// positions that map onto it in the (input channel, kernel row,
// kernel column) order, the bias added last. It is the test's own
// arithmetic, not a second kernel: the loops mirror the definition
// the documentation states.
func convTranspose2DReference(in []float64, k []float64, bias []float64,
	n, cIn, hIn, wIn, cOut, kH, kW, stride, padding int,
) []float64 {
	hOut := (hIn-1)*stride - 2*padding + kH
	wOut := (wIn-1)*stride - 2*padding + kW
	out := make([]float64, n*cOut*hOut*wOut)
	for b := range n {
		for oc := range cOut {
			for oh := range hOut {
				for ow := range wOut {
					sum := 0.0
					for ic := range cIn {
						for kh := range kH {
							ih := oh + padding - kh
							if ih < 0 || ih%stride != 0 {
								continue
							}
							ih /= stride
							if ih >= hIn {
								continue
							}
							for kw := range kW {
								iw := ow + padding - kw
								if iw < 0 || iw%stride != 0 {
									continue
								}
								iw /= stride
								if iw >= wIn {
									continue
								}
								sum += in[((b*cIn+ic)*hIn+ih)*wIn+iw] * k[((ic*cOut+oc)*kH+kh)*kW+kw]
							}
						}
					}
					if bias != nil {
						sum += bias[oc]
					}
					out[((b*cOut+oc)*hOut+oh)*wOut+ow] = sum
				}
			}
		}
	}
	return out
}

// TestConvTranspose2DChannelBlocking pins the channel-blocked kernel
// against the scatter definition on shapes the single-channel test
// cannot reach: several input and output channels (so the kernel's
// channel stride matters), a stride above one, a bias, and an output
// wider than one channel block, which leaves a partial block at the
// end.
func TestConvTranspose2DChannelBlocking(t *testing.T) {
	seed := 1
	rand := func() float64 {
		seed = (7*seed + 3) % 97
		return float64(seed)/97 - 0.5
	}
	fill := func(n int) []float64 {
		v := make([]float64, n)
		for i := range v {
			v[i] = rand()
		}
		return v
	}

	t.Run("several channels, stride two, bias", func(t *testing.T) {
		const n, cIn, cOut, k = 2, 2, 3, 2
		const hIn, wIn, stride, padding = 3, 3, 2, 1
		in := fill(n * cIn * hIn * wIn)
		kk := fill(cIn * cOut * k * k)
		bias := fill(cOut)
		got, err := ConvTranspose2D(mustFromFloats(t, in, n, cIn, hIn, wIn), mustFromFloats(t, kk, cIn, cOut, k, k), mustFromFloats(t, bias, cOut), stride, padding)
		if err != nil {
			t.Fatalf("ConvTranspose2D: %v", err)
		}
		want := convTranspose2DReference(in, kk, bias, n, cIn, hIn, wIn, cOut, k, k, stride, padding)
		outF := got.RawFloats()
		for i := range want {
			if math.Float64bits(outF[i]) != math.Float64bits(want[i]) {
				t.Fatalf("element %d = %v, want %v", i, outF[i], want[i])
			}
		}
	})
	t.Run("partial channel block", func(t *testing.T) {
		// An output row of 1025 caps one block at three channels, so
		// four output channels run as a full block and a partial one.
		const n, cIn, cOut, k = 1, 1, 4, 3
		const hIn, wIn, stride, padding = 1025, 1025, 1, 1
		in := fill(n * cIn * hIn * wIn)
		kk := fill(cIn * cOut * k * k)
		bias := fill(cOut)
		got, err := ConvTranspose2D(mustFromFloats(t, in, n, cIn, hIn, wIn), mustFromFloats(t, kk, cIn, cOut, k, k), mustFromFloats(t, bias, cOut), stride, padding)
		if err != nil {
			t.Fatalf("ConvTranspose2D: %v", err)
		}
		want := convTranspose2DReference(in, kk, bias, n, cIn, hIn, wIn, cOut, k, k, stride, padding)
		outF := got.RawFloats()
		for i := range want {
			if math.Float64bits(outF[i]) != math.Float64bits(want[i]) {
				t.Fatalf("element %d = %v, want %v", i, outF[i], want[i])
			}
		}
	})
}