makarna/pkg/tensor/simd_dequant_test.go

package tensor

import "testing"

func TestDequantQ8KSimdMatchesScalar(t *testing.T) {
	var b BlockQ8_K
	for i := 0; i < 256; i++ {
		b.QS[i] = int8(i%17 - 8)
	}
	b.D = 0.125

	var outSimd, outScalar [256]float32
	if !dequantQ8KSimd(&b, outSimd[:]) {
		t.Skip("SIMD path not available on this CPU")
	}

	// Force scalar path
	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	DequantizeQ8_K(&b, outScalar[:])
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	for i := range outScalar {
		if diff := absDiff(outSimd[i], outScalar[i]); diff > 1e-6 {
			t.Fatalf("mismatch at %d: simd=%f scalar=%f", i, outSimd[i], outScalar[i])
		}
	}
}

func TestDequantQ4KSimdMatchesScalar(t *testing.T) {
	var b BlockQ4_K
	b.D = 0x3C00 // float16(1.0)
	b.DMin = 0x3800 // float16(0.5)
	for i := range b.Scales {
		b.Scales[i] = uint8((i % 4) + 1)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i % 256)
	}

	var outSimd, outScalar [256]float32
	if !dequantQ4KSimd(&b, outSimd[:]) {
		t.Skip("SIMD path not available on this CPU")
	}

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	DequantizeQ4_K(&b, outScalar[:])
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	for i := range outScalar {
		if diff := absDiff(outSimd[i], outScalar[i]); diff > 1e-5 {
			t.Fatalf("mismatch at %d: simd=%f scalar=%f", i, outSimd[i], outScalar[i])
		}
	}
}

func TestDotQ4KSimdMatchesScalar(t *testing.T) {
	var b BlockQ4_K
	b.D = 0x3C00
	b.DMin = 0x3800
	for i := range b.Scales {
		b.Scales[i] = uint8((i % 4) + 1)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i % 251)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	// SIMD path
	if !hasAVX2 {
		t.Skip("SIMD path not available on this CPU")
	}
	simd := DotQ4_K(&b, x)

	// Force scalar
	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	scalar := DotQ4_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	if diff := absDiff(simd, scalar); diff > 1e-4 {
		t.Fatalf("mismatch: simd=%f scalar=%f", simd, scalar)
	}
}

func TestDotQ5KParamsMatchesScalar(t *testing.T) {
	if !hasAVX2 && !hasAVX512 {
		t.Skip("SIMD path not available on this CPU")
	}
	var b BlockQ5_K
	b.D = 0x3C00
	b.DMin = 0x3800
	for i := range b.Scales {
		b.Scales[i] = uint8(i*7 + 3)
	}
	for i := range b.QH {
		b.QH[i] = uint8(i * 13)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i * 5)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	params := GetQ5KDotParams([]BlockQ5_K{b})

	got := DotQ5_K_Params(&b, &params[0], x)
	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	want := DotQ5_K_Params(&b, &params[0], x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512
	if diff := absDiff(got, want); diff > 1e-4 {
		t.Fatalf("mismatch: got=%f want=%f", got, want)
	}
}

func TestDotQ5KTile8MatchesScalar(t *testing.T) {
	var blocks [8]BlockQ5_K
	for bi := range blocks {
		b := &blocks[bi]
		b.D = 0x3C00
		b.DMin = 0x3800
		for i := range b.Scales {
			b.Scales[i] = uint8((i + bi) * 3)
		}
		for i := range b.QH {
			b.QH[i] = uint8(i*5 + bi)
		}
		for i := range b.QS {
			b.QS[i] = uint8(i*7 + bi)
		}
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%19)-9) * 0.125
	}

	wp := GetQ5KDotParams(blocks[:])
	var sums [8]float32
	DotQ5KTile8(&sums, blocks[:], wp, 0, 1, &x[0], 8)

	for i := 0; i < 8; i++ {
		want := DotQ5_K_Params(&blocks[i], &wp[i], x)
		if diff := absDiff(sums[i], want); diff > 1e-4 {
			t.Fatalf("mismatch at %d: got=%f want=%f", i, sums[i], want)
		}
	}
}

func TestDotQ6KParamsMatchesScalar(t *testing.T) {
	var b BlockQ6_K
	b.D = 0x3C00
	for i := range b.Scales {
		b.Scales[i] = int8((i%11) - 5)
	}
	for i := range b.QL {
		b.QL[i] = uint8(i * 7)
	}
	for i := range b.QH {
		b.QH[i] = uint8(i * 13)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	params := GetQ6KDotParams([]BlockQ6_K{b})
	got := DotQ6_K_Params(&b, &params[0], x)

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	want := DotQ6_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	if diff := absDiff(got, want); diff > 1e-4 {
		t.Fatalf("mismatch: got=%f want=%f", got, want)
	}
}

func TestDotQ6KTile8MatchesScalar(t *testing.T) {
	var blocks [8]BlockQ6_K
	for bi := range blocks {
		b := &blocks[bi]
		b.D = 0x3C00
		for i := range b.Scales {
			b.Scales[i] = int8(((i + bi) % 13) - 6)
		}
		for i := range b.QL {
			b.QL[i] = uint8(i*3 + bi)
		}
		for i := range b.QH {
			b.QH[i] = uint8(i*5 + bi)
		}
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%19)-9) * 0.125
	}

	wp := GetQ6KDotParams(blocks[:])
	var sums [8]float32
	DotQ6KTile8(&sums, blocks[:], wp, 0, 1, &x[0], 8)

	for i := 0; i < 8; i++ {
		want := DotQ6_K_Params(&blocks[i], &wp[i], x)
		if diff := absDiff(sums[i], want); diff > 1e-4 {
			t.Fatalf("mismatch at %d: got=%f want=%f", i, sums[i], want)
		}
	}
}

func TestDotQ8KSimdMatchesScalar(t *testing.T) {
	var b BlockQ8_K
	for i := 0; i < 256; i++ {
		b.QS[i] = int8((i%31)-15)
	}
	b.D = 0.125

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	if !hasAVX2 && !hasAVX512 {
		t.Skip("SIMD path not available on this CPU")
	}
	simd := DotQ8_K(&b, x)

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	scalar := DotQ8_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	if diff := absDiff(simd, scalar); diff > 1e-3 {
		t.Fatalf("mismatch: simd=%f scalar=%f", simd, scalar)
	}
}

func TestDotQ2KMatchesDequantDot(t *testing.T) {
	var b BlockQ2_K
	b.D = 0x3C00
	b.DMin = 0x3800
	for i := range b.Scales {
		b.Scales[i] = uint8((i%16)*16 + i%16)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i * 3)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%19)-9) * 0.125
	}

	got := DotQ2_K(&b, x)

	var deq [256]float32
	DequantizeQ2_K(&b, deq[:])
	var want float32
	for i := 0; i < 256; i++ {
		want += x[i] * deq[i]
	}

	if diff := absDiff(got, want); diff > 1e-4 {
		t.Fatalf("mismatch: got=%f want=%f", got, want)
	}
}

func TestDotQ3KMatchesDequantDot(t *testing.T) {
	var b BlockQ3_K
	b.D = 0x3C00
	for i := range b.Scales {
		b.Scales[i] = uint8(i * 7)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i)
	}
	for i := range b.HMask {
		b.HMask[i] = uint8(i * 5)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%17)-8) * 0.25
	}

	got := DotQ3_K(&b, x)

	var deq [256]float32
	DequantizeQ3_K(&b, deq[:])
	var want float32
	for i := 0; i < 256; i++ {
		want += x[i] * deq[i]
	}

	if diff := absDiff(got, want); diff > 1e-4 {
		t.Fatalf("mismatch: got=%f want=%f", got, want)
	}
}

func TestDotQ2KSimdMatchesScalar(t *testing.T) {
	if !hasAVX2 {
		t.Skip("SIMD path not available on this CPU")
	}
	var b BlockQ2_K
	b.D = 0x3C00
	b.DMin = 0x3800
	for i := range b.Scales {
		b.Scales[i] = uint8((i%16)*16 + i%16)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i * 7)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%29)-14) * 0.125
	}

	simd := DotQ2_K(&b, x)

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	scalar := DotQ2_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	if diff := absDiff(simd, scalar); diff > 1e-4 {
		t.Fatalf("mismatch: simd=%f scalar=%f", simd, scalar)
	}
}

func TestDotQ3KSimdMatchesScalar(t *testing.T) {
	if !hasAVX2 {
		t.Skip("SIMD path not available on this CPU")
	}
	var b BlockQ3_K
	b.D = 0x3C00
	for i := range b.Scales {
		b.Scales[i] = uint8(i*11 + 3)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i * 5)
	}
	for i := range b.HMask {
		b.HMask[i] = uint8(i * 13)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	simd := DotQ3_K(&b, x)

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	scalar := DotQ3_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512
	if diff := absDiff(simd, scalar); diff > 1e-4 {
		t.Fatalf("mismatch: simd=%f scalar=%f", simd, scalar)
	}
}

func TestDotQ2KParamsMatchesScalar(t *testing.T) {
	var b BlockQ2_K
	b.D = 0x3C00
	b.DMin = 0x3800
	for i := range b.Scales {
		b.Scales[i] = uint8(i*7 + 3)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i * 5)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	params := GetQ2KDotParams([]BlockQ2_K{b})
	got := DotQ2_K_Params(&b, &params[0], x)

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	want := DotQ2_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	if diff := absDiff(got, want); diff > 1e-4 {
		t.Fatalf("mismatch: got=%f want=%f", got, want)
	}
}

func TestDotQ2KTile8MatchesScalar(t *testing.T) {
	var blocks [8]BlockQ2_K
	for bi := range blocks {
		b := &blocks[bi]
		b.D = 0x3C00
		b.DMin = 0x3800
		for i := range b.Scales {
			b.Scales[i] = uint8((i + bi) * 3)
		}
		for i := range b.QS {
			b.QS[i] = uint8(i*7 + bi)
		}
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%19)-9) * 0.125
	}

	wp := GetQ2KDotParams(blocks[:])
	var sums [8]float32
	DotQ2KTile8(&sums, blocks[:], wp, 0, 1, &x[0], 8)

	for i := 0; i < 8; i++ {
		want := DotQ2_K_Params(&blocks[i], &wp[i], x)
		if diff := absDiff(sums[i], want); diff > 1e-4 {
			t.Fatalf("mismatch at %d: got=%f want=%f", i, sums[i], want)
		}
	}
}

func TestDotQ8KTile8MatchesScalar(t *testing.T) {
	var blocks [8]BlockQ8_K
	for bi := range blocks {
		b := &blocks[bi]
		for i := 0; i < 256; i++ {
			b.QS[i] = int8((i+bi)%31 - 15)
		}
		b.D = 0.125
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	var sums [8]float32
	DotQ8KTile8(&sums, blocks[:], 0, 1, &x[0], 8)

	for i := 0; i < 8; i++ {
		want := DotQ8_K(&blocks[i], x)
		if diff := absDiff(sums[i], want); diff > 1e-3 {
			t.Fatalf("mismatch at %d: got=%f want=%f", i, sums[i], want)
		}
	}
}

func TestDotQ6KSimdMatchesScalar(t *testing.T) {
	if !hasAVX2 {
		t.Skip("SIMD path not available on this CPU")
	}
	if !q6kSimdReady() {
		t.Skip("Q6K SIMD gate disabled")
	}
	var b BlockQ6_K
	b.D = 0x3C00
	for i := range b.Scales {
		b.Scales[i] = int8((i % 16) - 8)
	}
	for i := range b.QL {
		b.QL[i] = uint8(i)
	}
	for i := range b.QH {
		b.QH[i] = uint8(i * 3)
	}

	x := make([]float32, 256)
	for i := range x {
		x[i] = float32((i%23)-11) * 0.25
	}

	simd := DotQ6_K(&b, x)

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	scalar := DotQ6_K(&b, x)
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	if diff := absDiff(simd, scalar); diff > 1e-3 {
		t.Fatalf("mismatch: simd=%f scalar=%f", simd, scalar)
	}
}

func TestDequantQ2KSimdMatchesScalar(t *testing.T) {
	var b BlockQ2_K
	b.D = 0x3C00 // float16(1.0)
	b.DMin = 0x3800 // float16(0.5)
	for i := range b.Scales {
		b.Scales[i] = uint8((i%16)*16 + i%16) // both scales and mins
	}
	for i := range b.QS {
		b.QS[i] = uint8(i)
	}

	var outSimd, outScalar [256]float32
	if !dequantQ2KSimd(&b, outSimd[:]) {
		t.Skip("SIMD path not available on this CPU")
	}

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	DequantizeQ2_K(&b, outScalar[:])
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	for i := range outScalar {
		if diff := absDiff(outSimd[i], outScalar[i]); diff > 1e-5 {
			t.Fatalf("mismatch at %d: simd=%f scalar=%f", i, outSimd[i], outScalar[i])
		}
	}
}

func TestDequantQ3KSimdMatchesScalar(t *testing.T) {
	var b BlockQ3_K
	b.D = 0x3C00 // float16(1.0)
	for i := range b.Scales {
		b.Scales[i] = uint8(i % 64)
	}
	for i := range b.QS {
		b.QS[i] = uint8(i)
	}
	for i := range b.HMask {
		b.HMask[i] = uint8(i * 7)
	}

	var outSimd, outScalar [256]float32
	if !dequantQ3KSimd(&b, outSimd[:]) {
		t.Skip("SIMD path not available on this CPU")
	}

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	DequantizeQ3_K(&b, outScalar[:])
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	for i := range outScalar {
		if diff := absDiff(outSimd[i], outScalar[i]); diff > 1e-5 {
			t.Fatalf("mismatch at %d: simd=%f scalar=%f", i, outSimd[i], outScalar[i])
		}
	}
}

func TestDequantQ6KSimdMatchesScalar(t *testing.T) {
	var b BlockQ6_K
	b.D = 0x3C00 // float16(1.0)
	for i := range b.Scales {
		b.Scales[i] = int8((i % 16) - 8)
	}
	for i := range b.QL {
		b.QL[i] = uint8(i)
	}
	for i := range b.QH {
		b.QH[i] = uint8(i * 3)
	}

	var outSimd, outScalar [256]float32
	if !dequantQ6KSimd(&b, outSimd[:]) {
		t.Skip("SIMD path not available on this CPU")
	}

	origAVX2, origAVX512 := hasAVX2, hasAVX512
	hasAVX2, hasAVX512 = false, false
	DequantizeQ6_K(&b, outScalar[:])
	hasAVX2, hasAVX512 = origAVX2, origAVX512

	for i := range outScalar {
		if diff := absDiff(outSimd[i], outScalar[i]); diff > 1e-4 {
			t.Fatalf("mismatch at %d: simd=%f scalar=%f", i, outSimd[i], outScalar[i])
		}
	}
}

func absDiff(a, b float32) float32 {
	if a > b {
		return a - b
	}
	return b - a
}