git.sesse.net Git - movit/blob - fp16_test.cpp

   1 #include "fp16.h"
   2
   3 #include <math.h>
   4 #include <gtest/gtest.h>
   5
   6 namespace movit {
   7
   8 TEST(FP16Test, Simple) {
   9         EXPECT_EQ(0x0000, fp64_to_fp16(0.0));
  10         EXPECT_DOUBLE_EQ(0.0, fp16_to_fp64(0x0000));
  11
  12         EXPECT_EQ(0x3c00, fp64_to_fp16(1.0));
  13         EXPECT_DOUBLE_EQ(1.0, fp16_to_fp64(0x3c00));
  14
  15         EXPECT_EQ(0x3555, fp64_to_fp16(1.0 / 3.0));
  16         EXPECT_DOUBLE_EQ(0.333251953125, fp16_to_fp64(0x3555));
  17 }
  18
  19 TEST(FP16Test, RoundToNearestEven) {
  20         ASSERT_DOUBLE_EQ(1.0, fp16_to_fp64(0x3c00));
  21
  22         double x0 = fp16_to_fp64(0x3c00);
  23         double x1 = fp16_to_fp64(0x3c01);
  24         double x2 = fp16_to_fp64(0x3c02);
  25         double x3 = fp16_to_fp64(0x3c03);
  26         double x4 = fp16_to_fp64(0x3c04);
  27
  28         EXPECT_EQ(0x3c00, fp64_to_fp16(0.5 * (x0 + x1)));
  29         EXPECT_EQ(0x3c02, fp64_to_fp16(0.5 * (x1 + x2)));
  30         EXPECT_EQ(0x3c02, fp64_to_fp16(0.5 * (x2 + x3)));
  31         EXPECT_EQ(0x3c04, fp64_to_fp16(0.5 * (x3 + x4)));
  32 }
  33
  34 union fp64 {
  35         double f;
  36         unsigned long long ll;
  37 };
  38 union fp32 {
  39         float f;
  40         unsigned int u;
  41 };
  42
  43 TEST(FP16Test, NaN) {
  44         EXPECT_EQ(0xfe00, fp64_to_fp16(0.0 / 0.0));
  45         EXPECT_TRUE(isnan(fp16_to_fp64(0xfe00)));
  46
  47         fp64 borderline_inf;
  48         borderline_inf.ll = 0x7ff0000000000000ull;
  49         fp64 borderline_nan;
  50         borderline_nan.ll = 0x7ff0000000000001ull;
  51
  52         ASSERT_FALSE(isfinite(borderline_inf.f));
  53         ASSERT_FALSE(isnan(borderline_inf.f));
  54
  55         ASSERT_FALSE(isfinite(borderline_nan.f));
  56         ASSERT_TRUE(isnan(borderline_nan.f));
  57
  58         double borderline_inf_roundtrip = fp16_to_fp64(fp64_to_fp16(borderline_inf.f));
  59         double borderline_nan_roundtrip = fp16_to_fp64(fp64_to_fp16(borderline_nan.f));
  60
  61         EXPECT_FALSE(isfinite(borderline_inf_roundtrip));
  62         EXPECT_FALSE(isnan(borderline_inf_roundtrip));
  63
  64         EXPECT_FALSE(isfinite(borderline_nan_roundtrip));
  65         EXPECT_TRUE(isnan(borderline_nan_roundtrip));
  66 }
  67
  68 TEST(FP16Test, Denormals) {
  69         const double smallest_fp16_denormal = 5.9604644775390625e-08;
  70         EXPECT_EQ(0x0001, fp64_to_fp16(smallest_fp16_denormal));
  71         EXPECT_EQ(0x0000, fp64_to_fp16(0.5 * smallest_fp16_denormal));  // Round-to-even.
  72         EXPECT_EQ(0x0001, fp64_to_fp16(0.51 * smallest_fp16_denormal));
  73         EXPECT_EQ(0x0002, fp64_to_fp16(1.5 * smallest_fp16_denormal));
  74
  75         const double smallest_fp16_non_denormal = 6.103515625e-05;
  76         EXPECT_EQ(0x0400, fp64_to_fp16(smallest_fp16_non_denormal));
  77         EXPECT_EQ(0x0400, fp64_to_fp16(smallest_fp16_non_denormal - 0.5 * smallest_fp16_denormal));  // Round-to-even.
  78         EXPECT_EQ(0x03ff, fp64_to_fp16(smallest_fp16_non_denormal - smallest_fp16_denormal));
  79 }
  80
  81 // Randomly test a large number of fp64 -> fp32 conversions, comparing
  82 // against the FPU.
  83 TEST(FP16Test, FP32ReferenceDownconvert) {
  84         srand(12345);
  85
  86         for (int i = 0; i < 1000000; ++i) {
  87                 unsigned r1 = rand();
  88                 unsigned r2 = rand();
  89                 unsigned r3 = rand();
  90                 union fp64 src;
  91                 union fp32 reference, result;
  92
  93                 src.ll = (((unsigned long long)r1) << 33) ^ ((unsigned long long)r2 << 16) ^ r3;
  94                 reference.f = float(src.f);
  95                 result.u = fp64_to_fp32(src.f);
  96
  97                 EXPECT_EQ(isnan(result.f), isnan(reference.f));
  98                 if (!isnan(result.f)) {
  99                         EXPECT_EQ(result.u, reference.u)
 100                             << src.f << " got rounded to " << result.u << " (" << result.f << ")";
 101                 }
 102         }
 103 }
 104
 105 // Randomly test a large number of fp32 -> fp64 conversions, comparing
 106 // against the FPU.
 107 TEST(FP16Test, FP32ReferenceUpconvert) {
 108         srand(12345);
 109
 110         for (int i = 0; i < 1000000; ++i) {
 111                 unsigned r1 = rand();
 112                 unsigned r2 = rand();
 113                 union fp32 src;
 114                 union fp64 reference, result;
 115
 116                 src.u = ((unsigned long long)r1 << 16) ^ r2;
 117                 reference.f = double(src.f);
 118                 result.f = fp32_to_fp64(src.u);
 119
 120                 EXPECT_EQ(isnan(result.f), isnan(reference.f));
 121                 if (!isnan(result.f)) {
 122                         EXPECT_EQ(result.ll, reference.ll)
 123                             << src.f << " got converted to " << result.ll << " (" << result.f << ")";
 124                 }
 125         }
 126 }
 127
 128 }  // namespace movit