X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=fp16_test.cpp;h=e0920e9ce60bf499bb636ac4c7ab8eac9b2b1b5c;hp=d5d7fc04c22396a48fc7bbf73b239629f6fd39e6;hb=6f1efa8348a90a393187c12d70fd10d81bbd2c99;hpb=82464681eca54bf11be3ca53eed20841ff6eb204

diff --git a/fp16_test.cpp b/fp16_test.cpp
index d5d7fc0..e0920e9 100644
--- a/fp16_test.cpp
+++ b/fp16_test.cpp
@@ -4,21 +4,48 @@
 #include <gtest/gtest.h>
 
 namespace movit {
+namespace {
+
+fp16_int_t make_fp16(unsigned short x)
+{
+	fp16_int_t ret;
+	ret.val = x;
+	return ret;
+}
+
+fp32_int_t make_fp32(unsigned int x)
+{
+	fp32_int_t ret;
+	ret.val = x;
+	return ret;
+}
+
+}  // namespace
 
 TEST(FP16Test, Simple) {
-	EXPECT_EQ(0x0000, fp64_to_fp16(0.0));
-	EXPECT_DOUBLE_EQ(0.0, fp16_to_fp64(0x0000));
+	EXPECT_EQ(0x0000, fp64_to_fp16(0.0).val);
+	EXPECT_DOUBLE_EQ(0.0, fp16_to_fp64(make_fp16(0x0000)));
 
-	EXPECT_EQ(0x3c00, fp64_to_fp16(1.0));
-	EXPECT_DOUBLE_EQ(1.0, fp16_to_fp64(0x3c00));
+	EXPECT_EQ(0x3c00, fp64_to_fp16(1.0).val);
+	EXPECT_DOUBLE_EQ(1.0, fp16_to_fp64(make_fp16(0x3c00)));
 
-	EXPECT_EQ(0x3555, fp64_to_fp16(1.0 / 3.0));
-	EXPECT_DOUBLE_EQ(0.333251953125, fp16_to_fp64(0x3555));
+	EXPECT_EQ(0x3555, fp64_to_fp16(1.0 / 3.0).val);
+	EXPECT_DOUBLE_EQ(0.333251953125, fp16_to_fp64(make_fp16(0x3555)));
 }
 
-TEST(FP16Test, NaN) {
-	EXPECT_EQ(0xfe00, fp64_to_fp16(0.0 / 0.0));
-	EXPECT_TRUE(isnan(fp16_to_fp64(0xfe00)));
+TEST(FP16Test, RoundToNearestEven) {
+	ASSERT_DOUBLE_EQ(1.0, fp16_to_fp64(make_fp16(0x3c00)));
+
+	double x0 = fp16_to_fp64(make_fp16(0x3c00));
+	double x1 = fp16_to_fp64(make_fp16(0x3c01));
+	double x2 = fp16_to_fp64(make_fp16(0x3c02));
+	double x3 = fp16_to_fp64(make_fp16(0x3c03));
+	double x4 = fp16_to_fp64(make_fp16(0x3c04));
+
+	EXPECT_EQ(0x3c00, fp64_to_fp16(0.5 * (x0 + x1)).val);
+	EXPECT_EQ(0x3c02, fp64_to_fp16(0.5 * (x1 + x2)).val);
+	EXPECT_EQ(0x3c02, fp64_to_fp16(0.5 * (x2 + x3)).val);
+	EXPECT_EQ(0x3c04, fp64_to_fp16(0.5 * (x3 + x4)).val);
 }
 
 union fp64 {
@@ -30,6 +57,45 @@ union fp32 {
 	unsigned int u;
 };
 
+TEST(FP16Test, NaN) {
+	// Ignore the sign bit.
+	EXPECT_EQ(0x7e00, fp64_to_fp16(0.0 / 0.0).val & 0x7fff);
+	EXPECT_TRUE(isnan(fp16_to_fp64(make_fp16(0xfe00))));
+
+	fp64 borderline_inf;
+	borderline_inf.ll = 0x7ff0000000000000ull;
+	fp64 borderline_nan;
+	borderline_nan.ll = 0x7ff0000000000001ull;
+
+	ASSERT_FALSE(isfinite(borderline_inf.f));
+	ASSERT_FALSE(isnan(borderline_inf.f));
+
+	ASSERT_FALSE(isfinite(borderline_nan.f));
+	ASSERT_TRUE(isnan(borderline_nan.f));
+
+	double borderline_inf_roundtrip = fp16_to_fp64(fp64_to_fp16(borderline_inf.f));
+	double borderline_nan_roundtrip = fp16_to_fp64(fp64_to_fp16(borderline_nan.f));
+
+	EXPECT_FALSE(isfinite(borderline_inf_roundtrip));
+	EXPECT_FALSE(isnan(borderline_inf_roundtrip));
+
+	EXPECT_FALSE(isfinite(borderline_nan_roundtrip));
+	EXPECT_TRUE(isnan(borderline_nan_roundtrip));
+}
+
+TEST(FP16Test, Denormals) {
+	const double smallest_fp16_denormal = 5.9604644775390625e-08;
+	EXPECT_EQ(0x0001, fp64_to_fp16(smallest_fp16_denormal).val);
+	EXPECT_EQ(0x0000, fp64_to_fp16(0.5 * smallest_fp16_denormal).val);  // Round-to-even.
+	EXPECT_EQ(0x0001, fp64_to_fp16(0.51 * smallest_fp16_denormal).val);
+	EXPECT_EQ(0x0002, fp64_to_fp16(1.5 * smallest_fp16_denormal).val);
+
+	const double smallest_fp16_non_denormal = 6.103515625e-05;
+	EXPECT_EQ(0x0400, fp64_to_fp16(smallest_fp16_non_denormal).val);
+	EXPECT_EQ(0x0400, fp64_to_fp16(smallest_fp16_non_denormal - 0.5 * smallest_fp16_denormal).val);  // Round-to-even.
+	EXPECT_EQ(0x03ff, fp64_to_fp16(smallest_fp16_non_denormal - smallest_fp16_denormal).val);
+}
+
 // Randomly test a large number of fp64 -> fp32 conversions, comparing
 // against the FPU.
 TEST(FP16Test, FP32ReferenceDownconvert) {
@@ -44,7 +110,7 @@ TEST(FP16Test, FP32ReferenceDownconvert) {
 
 		src.ll = (((unsigned long long)r1) << 33) ^ ((unsigned long long)r2 << 16) ^ r3;
 		reference.f = float(src.f);
-		result.u = fp64_to_fp32(src.f);
+		result.u = fp64_to_fp32(src.f).val;
 
 		EXPECT_EQ(isnan(result.f), isnan(reference.f));
 		if (!isnan(result.f)) {
@@ -67,7 +133,7 @@ TEST(FP16Test, FP32ReferenceUpconvert) {
 
 		src.u = ((unsigned long long)r1 << 16) ^ r2;
 		reference.f = double(src.f);
-		result.f = fp32_to_fp64(src.u);
+		result.f = fp32_to_fp64(make_fp32(src.u));
 
 		EXPECT_EQ(isnan(result.f), isnan(reference.f));
 		if (!isnan(result.f)) {