//go:build cuda

package matmul

import (
	"testing"

	"makarna/pkg/backend/cpu"
	"makarna/pkg/tensor"
)

func TestLinearCudaMatchesCPU(t *testing.T) {
	// Small matrix to compare CPU vs CUDA paths.
	M, K, N := 4, 8, 3
	aCPU := cpu.NewTensor(tensor.Shape{M, K}, nil)
	wCPU := cpu.NewTensor(tensor.Shape{N, K}, nil)
	outCPU := cpu.NewTensor(tensor.Shape{M, N}, nil)
	outCUDA := cpu.NewTensor(tensor.Shape{M, N}, nil)

	fillSeq(aCPU.DataFloat32())
	fillSeq(wCPU.DataFloat32())

	if err := linearCPU(aCPU, wCPU, outCPU); err != nil {
		t.Fatalf("cpu linear failed: %v", err)
	}
	if err := Linear(aCPU, wCPU, outCUDA); err != nil {
		t.Fatalf("cuda linear failed: %v", err)
	}

	for i, v := range outCPU.DataFloat32() {
		got := outCUDA.DataFloat32()[i]
		if diff := abs32(v - got); diff > 1e-4 {
			t.Fatalf("mismatch at %d: cpu=%f cuda=%f", i, v, got)
		}
	}
}

func fillSeq(dst []float32) {
	for i := range dst {
		dst[i] = float32(i%7 + 1)
	}
}

func abs32(v float32) float32 {
	if v < 0 {
		return -v
	}
	return v
}