fix: use truncation instead of rounding in ToBF16 (#1244)

Author: zhangzhenghao

common/floats/floats.go

@@ -232,8 +232,7 @@ func ToBF16(values []float32) []uint16 {
 	encoded := make([]uint16, len(values))
 	for i, value := range values {
 		bits := math.Float32bits(value)
-		roundingBias := uint32(0x7FFF + ((bits >> 16) & 1))
-		encoded[i] = uint16((bits + roundingBias) >> 16)
+		encoded[i] = uint16(bits >> 16)
 	}
 	return encoded
 }

common/floats/floats_test.go

@@ -42,17 +42,17 @@ func TestZero(t *testing.T) {
 
 func TestToBF16(t *testing.T) {
 	assert.Nil(t, ToBF16(nil))
-	assert.Equal(t, []uint16{0x0000, 0x3f80, 0xc020, 0x3f8d}, ToBF16([]float32{0, 1, -2.5, 1.1}))
+	assert.Equal(t, []uint16{0x0000, 0x3f80, 0xc020, 0x3f8c}, ToBF16([]float32{0, 1, -2.5, 1.1}))
 }
 
 func TestFromBF16(t *testing.T) {
 	assert.Nil(t, FromBF16(nil))
-	decoded := FromBF16([]uint16{0x0000, 0x3f80, 0xc020, 0x3f8d})
+	decoded := FromBF16([]uint16{0x0000, 0x3f80, 0xc020, 0x3f8c})
 	assert.Len(t, decoded, 4)
 	assert.Equal(t, uint32(0x00000000), math.Float32bits(decoded[0]))
 	assert.Equal(t, uint32(0x3f800000), math.Float32bits(decoded[1]))
 	assert.Equal(t, uint32(0xc0200000), math.Float32bits(decoded[2]))
-	assert.Equal(t, uint32(0x3f8d0000), math.Float32bits(decoded[3]))
+	assert.Equal(t, uint32(0x3f8c0000), math.Float32bits(decoded[3]))
 }
 
 func TestAdd(t *testing.T) {

model/ctr/data_test.go

@@ -174,7 +174,7 @@ func TestDataset_Split(t *testing.T) {
 		dataSet.Index.CountUsers() + dataSet.Index.CountItems() + dataSet.Index.CountUserLabels() + 8,
 		0,
 	}, features)
-	assert.InDeltaSlice(t, []float32{2, 2.1, 2.2}, floats.FromBF16(embeddings[0]), 0.01)
+	assert.InDeltaSlice(t, []float32{2, 2.09375, 2.1875}, floats.FromBF16(embeddings[0]), 0.001)
 	assert.Equal(t, []float32{1, 1, 1, 1, 1, 1, 1, 0.5}, values)
 	assert.Equal(t, float32(-1), target)