| // Copyright 2019 The SwiftShader Authors. All Rights Reserved. |
| // |
| // Licensed under the Apache License, Version 2.0 (the "License"); |
| // you may not use this file except in compliance with the License. |
| // You may obtain a copy of the License at |
| // |
| // http://www.apache.org/licenses/LICENSE-2.0 |
| // |
| // Unless required by applicable law or agreed to in writing, software |
| // distributed under the License is distributed on an "AS IS" BASIS, |
| // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| // See the License for the specific language governing permissions and |
| // limitations under the License. |
| |
| #include "System/CPUID.hpp" |
| #include "System/Half.hpp" |
| #include "System/Math.hpp" |
| |
| #include <gmock/gmock.h> |
| #include <gtest/gtest.h> |
| |
| #include <cstdlib> |
| #include <cmath> |
| |
| using std::isnan; |
| using std::isinf; |
| using std::signbit; |
| |
| using namespace sw; |
| |
| // Implementation of frexp() which satisfies C++ <cmath> requirements. |
| float fast_frexp(float val, int *exp) |
| { |
| int isNotZero = (val != 0.0f) ? 0xFFFFFFFF : 0x00000000; |
| int v = bit_cast<int>(val); |
| int isInfOrNaN = (v & 0x7F800000) == 0x7F800000 ? 0xFFFFFFFF : 0x00000000; |
| |
| // When val is a subnormal value we can't directly use its mantissa to construct the significand in |
| // the range [0.5, 1.0). We need to multiply it by a factor that makes it normalized. For large |
| // values the factor must avoid overflow to inifity. |
| int factor = ((127 + 23) << 23) - (v & 0x3F800000); |
| int nval = bit_cast<int>(val * bit_cast<float>(factor)); |
| |
| // Extract the exponent of the normalized value and subtract the exponent of the normalizing factor. |
| int exponent = ((((nval & 0x7F800000) - factor) >> 23) + 1) & isNotZero; |
| |
| // Substitute the exponent of 0.5f (if not zero) to obtain the significand. |
| float significand = bit_cast<float>((nval & 0x807FFFFF) | (0x3F000000 & isNotZero) | (0x7F800000 & isInfOrNaN)); |
| |
| *exp = exponent; |
| return significand; |
| } |
| |
| TEST(MathTest, Frexp) |
| { |
| for(bool flush : { false, true }) |
| { |
| CPUID::setDenormalsAreZero(flush); |
| CPUID::setFlushToZero(flush); |
| |
| std::vector<float> a = { |
| 2.3f, |
| 0.1f, |
| 0.7f, |
| 1.7f, |
| 0.0f, |
| -2.3f, |
| -0.1f, |
| -0.7f, |
| -1.7f, |
| -0.0f, |
| 100000000.0f, |
| -100000000.0f, |
| 0.000000001f, |
| -0.000000001f, |
| FLT_MIN, |
| -FLT_MIN, |
| FLT_MAX, |
| -FLT_MAX, |
| FLT_TRUE_MIN, |
| -FLT_TRUE_MIN, |
| INFINITY, |
| -INFINITY, |
| NAN, |
| bit_cast<float>(0x007FFFFF), // Largest subnormal |
| bit_cast<float>(0x807FFFFF), |
| bit_cast<float>(0x00000001), // Smallest subnormal |
| bit_cast<float>(0x80000001), |
| }; |
| |
| for(float f : a) |
| { |
| int exp = -1000; |
| float sig = fast_frexp(f, &exp); |
| |
| if(f == 0.0f) // Could be subnormal if `flush` is true |
| { |
| // We don't rely on std::frexp here to produce a reference result because it may |
| // return non-zero significands and exponents for subnormal arguments., while our |
| // implementation is meant to respect denormals-are-zero / flush-to-zero. |
| |
| ASSERT_EQ(sig, 0.0f) << "Argument: " << std::hexfloat << f; |
| ASSERT_TRUE(signbit(sig) == signbit(f)) << "Argument: " << std::hexfloat << f; |
| ASSERT_EQ(exp, 0) << "Argument: " << std::hexfloat << f; |
| } |
| else |
| { |
| int ref_exp = -1000; |
| float ref_sig = std::frexp(f, &ref_exp); |
| |
| if(!isnan(f)) |
| { |
| ASSERT_EQ(sig, ref_sig) << "Argument: " << std::hexfloat << f; |
| } |
| else |
| { |
| ASSERT_TRUE(isnan(sig)) << "Significand: " << std::hexfloat << sig; |
| } |
| |
| if(!isinf(f) && !isnan(f)) // If the argument is NaN or Inf the exponent is unspecified. |
| { |
| ASSERT_EQ(exp, ref_exp) << "Argument: " << std::hexfloat << f; |
| } |
| } |
| } |
| } |
| } |
| |
| // Returns the whole-number ULP error of `a` relative to `x`. |
| // Use the doouble-precision version below. This just illustrates the principle. |
| [[deprecated]] float ULP_32(float x, float a) |
| { |
| // Flip the last mantissa bit to compute the 'unit in the last place' error. |
| float x1 = bit_cast<float>(bit_cast<uint32_t>(x) ^ 0x00000001); |
| float ulp = abs(x1 - x); |
| |
| return abs(a - x) / ulp; |
| } |
| |
| double ULP_32(double x, double a) |
| { |
| // binary64 has 52 mantissa bits, while binary32 has 23, so the ULP for the latter is 29 bits shifted. |
| double x1 = bit_cast<double>(bit_cast<uint64_t>(x) ^ 0x0000000020000000ull); |
| double ulp = abs(x1 - x); |
| |
| return abs(a - x) / ulp; |
| } |
| |
| float ULP_16(float x, float a) |
| { |
| // binary32 has 23 mantissa bits, while binary16 has 10, so the ULP for the latter is 13 bits shifted. |
| double x1 = bit_cast<float>(bit_cast<uint32_t>(x) ^ 0x00002000); |
| float ulp = abs(x1 - x); |
| |
| return abs(a - x) / ulp; |
| } |
| |
| // lolremez --float -d 2 -r "0:2^23" "(log2(x/2^23+1)-x/2^23)/x" "1/x" |
| // ULP-16: 0.797363281, abs: 0.0991751999 |
| float f(float x) |
| { |
| float u = 2.8017103e-22f; |
| u = u * x + -8.373131e-15f; |
| return u * x + 5.0615534e-8f; |
| } |
| |
| float Log2Relaxed(float x) |
| { |
| // Reinterpretation as an integer provides a piecewise linear |
| // approximation of log2(). Scale to the radix and subtract exponent bias. |
| int im = bit_cast<int>(x); |
| float y = (float)im * (1.0f / (1 << 23)) - 127.0f; |
| |
| // Handle log2(inf) = inf. |
| if(im == 0x7F800000) y = INFINITY; |
| |
| float m = (float)(im & 0x007FFFFF); // Unnormalized mantissa of x. |
| |
| // Add a polynomial approximation of log2(m+1)-m to the result's mantissa. |
| return f(m) * m + y; |
| } |
| |
| TEST(MathTest, Log2RelaxedExhaustive) |
| { |
| CPUID::setDenormalsAreZero(true); |
| CPUID::setFlushToZero(true); |
| |
| float worst_margin = 0; |
| float worst_ulp = 0; |
| float worst_x = 0; |
| float worst_val = 0; |
| float worst_ref = 0; |
| |
| float worst_abs = 0; |
| |
| for(float x = 0.0f; x <= INFINITY; x = inc(x)) |
| { |
| float val = Log2Relaxed(x); |
| |
| double ref = log2((double)x); |
| |
| if(ref == (int)ref) |
| { |
| ASSERT_EQ(val, ref); |
| } |
| else if(x >= 0.5f && x <= 2.0f) |
| { |
| const float tolerance = pow(2.0f, -7.0f); // Absolute |
| |
| float margin = abs(val - ref) / tolerance; |
| |
| if(margin > worst_abs) |
| { |
| worst_abs = margin; |
| } |
| } |
| else |
| { |
| const float tolerance = 3; // ULP |
| |
| float ulp = (float)ULP_16(ref, (double)val); |
| float margin = ulp / tolerance; |
| |
| if(margin > worst_margin) |
| { |
| worst_margin = margin; |
| worst_ulp = ulp; |
| worst_x = x; |
| worst_val = val; |
| worst_ref = ref; |
| } |
| } |
| } |
| |
| ASSERT_TRUE(worst_margin < 1.0f) << " worst_x " << worst_x << " worst_val " << worst_val << " worst_ref " << worst_ref << " worst_ulp " << worst_ulp; |
| ASSERT_TRUE(worst_abs <= 1.0f) << " worst_x " << worst_x << " worst_val " << worst_val << " worst_ref " << worst_ref << " worst_ulp " << worst_ulp; |
| |
| CPUID::setDenormalsAreZero(false); |
| CPUID::setFlushToZero(false); |
| } |
| |
| // lolremez --float -d 2 -r "0:1" "(2^x-x-1)/x" "1/x" |
| // ULP-16: 0.130859017 |
| float Pr(float x) |
| { |
| float u = 7.8145574e-2f; |
| u = u * x + 2.2617357e-1f; |
| return u * x + -3.0444314e-1f; |
| } |
| |
| float Exp2Relaxed(float x) |
| { |
| x = min(x, 128.0f); |
| x = max(x, bit_cast<float>(int(0xC2FDFFFF))); // -126.999992 |
| |
| // 2^f - f - 1 as P(f) * f |
| // This is a correction term to be added to 1+x to obtain 2^x. |
| float f = x - floor(x); |
| float y = Pr(f) * f + x; |
| |
| // bit_cast<float>(int(x * 2^23)) is a piecewise linear approximation of 2^(x-127). |
| // See "Fast Exponential Computation on SIMD Architectures" by Malossi et al. |
| return bit_cast<float>(int((1 << 23) * y + (127 << 23))); |
| } |
| |
| TEST(MathTest, Exp2RelaxedExhaustive) |
| { |
| CPUID::setDenormalsAreZero(true); |
| CPUID::setFlushToZero(true); |
| |
| float worst_margin = 0; |
| float worst_ulp = 0; |
| float worst_x = 0; |
| float worst_val = 0; |
| float worst_ref = 0; |
| |
| for(float x = -10; x <= 10; x = inc(x)) |
| { |
| float val = Exp2Relaxed(x); |
| |
| double ref = exp2((double)x); |
| |
| if(x == (int)x) |
| { |
| ASSERT_EQ(val, ref); |
| } |
| |
| const float tolerance = (1 + 2 * abs(x)); |
| float ulp = ULP_16((float)ref, val); |
| float margin = ulp / tolerance; |
| |
| if(margin > worst_margin) |
| { |
| worst_margin = margin; |
| worst_ulp = ulp; |
| worst_x = x; |
| worst_val = val; |
| worst_ref = ref; |
| } |
| } |
| |
| ASSERT_TRUE(worst_margin <= 1.0f) << " worst_x " << worst_x << " worst_val " << worst_val << " worst_ref " << worst_ref << " worst_ulp " << worst_ulp; |
| |
| CPUID::setDenormalsAreZero(false); |
| CPUID::setFlushToZero(false); |
| } |
| |
| // lolremez --float -d 7 -r "0:1" "(log2(x+1)-x)/x" "1/x" |
| // ULP-32: 1.69571960, abs: 0.360798746 |
| float Pl(float x) |
| { |
| float u = -9.3091638e-3f; |
| u = u * x + 5.2059003e-2f; |
| u = u * x + -1.3752135e-1f; |
| u = u * x + 2.4186478e-1f; |
| u = u * x + -3.4730109e-1f; |
| u = u * x + 4.786837e-1f; |
| u = u * x + -7.2116581e-1f; |
| return u * x + 4.4268988e-1f; |
| } |
| |
| float Log2(float x) |
| { |
| // Reinterpretation as an integer provides a piecewise linear |
| // approximation of log2(). Scale to the radix and subtract exponent bias. |
| int im = bit_cast<int>(x); |
| float y = (float)(im - (127 << 23)) * (1.0f / (1 << 23)); |
| |
| // Handle log2(inf) = inf. |
| if(im == 0x7F800000) y = INFINITY; |
| |
| float m = (float)(im & 0x007FFFFF) * (1.0f / (1 << 23)); // Normalized mantissa of x. |
| |
| // Add a polynomial approximation of log2(m+1)-m to the result's mantissa. |
| return Pl(m) * m + y; |
| } |
| |
| TEST(MathTest, Log2Exhaustive) |
| { |
| CPUID::setDenormalsAreZero(true); |
| CPUID::setFlushToZero(true); |
| |
| float worst_margin = 0; |
| float worst_ulp = 0; |
| float worst_x = 0; |
| float worst_val = 0; |
| float worst_ref = 0; |
| |
| float worst_abs = 0; |
| |
| for(float x = 0.0f; x <= INFINITY; x = inc(x)) |
| { |
| float val = Log2(x); |
| |
| double ref = log2((double)x); |
| |
| if(ref == (int)ref) |
| { |
| ASSERT_EQ(val, ref); |
| } |
| else if(x >= 0.5f && x <= 2.0f) |
| { |
| const float tolerance = pow(2.0f, -21.0f); // Absolute |
| |
| float margin = abs(val - ref) / tolerance; |
| |
| if(margin > worst_abs) |
| { |
| worst_abs = margin; |
| } |
| } |
| else |
| { |
| const float tolerance = 3; // ULP |
| |
| float ulp = (float)ULP_32(ref, (double)val); |
| float margin = ulp / tolerance; |
| |
| if(margin > worst_margin) |
| { |
| worst_margin = margin; |
| worst_ulp = ulp; |
| worst_x = x; |
| worst_val = val; |
| worst_ref = ref; |
| } |
| } |
| } |
| |
| ASSERT_TRUE(worst_margin < 1.0f) << " worst_x " << worst_x << " worst_val " << worst_val << " worst_ref " << worst_ref << " worst_ulp " << worst_ulp; |
| ASSERT_TRUE(worst_abs <= 1.0f) << " worst_x " << worst_x << " worst_val " << worst_val << " worst_ref " << worst_ref << " worst_ulp " << worst_ulp; |
| |
| CPUID::setDenormalsAreZero(false); |
| CPUID::setFlushToZero(false); |
| } |
| |
| // lolremez --float -d 4 -r "0:1" "(2^x-x-1)/x" "1/x" |
| // ULP_32: 2.14694786, Vulkan margin: 0.686957061 |
| float P(float x) |
| { |
| float u = 1.8852974e-3f; |
| u = u * x + 8.9733787e-3f; |
| u = u * x + 5.5835927e-2f; |
| u = u * x + 2.4015281e-1f; |
| return u * x + -3.0684753e-1f; |
| } |
| |
| float Exp2(float x) |
| { |
| x = min(x, 128.0f); |
| x = max(x, bit_cast<float>(0xC2FDFFFF)); // -126.999992 |
| |
| // 2^f - f - 1 as P(f) * f |
| // This is a correction term to be added to 1+x to obtain 2^x. |
| float f = x - floor(x); |
| float y = P(f) * f + x; |
| |
| // bit_cast<float>(int(x * 2^23)) is a piecewise linear approximation of 2^(x-127). |
| // See "Fast Exponential Computation on SIMD Architectures" by Malossi et al. |
| return bit_cast<float>(int(y * (1 << 23)) + (127 << 23)); |
| } |
| |
| TEST(MathTest, Exp2Exhaustive) |
| { |
| CPUID::setDenormalsAreZero(true); |
| CPUID::setFlushToZero(true); |
| |
| float worst_margin = 0; |
| float worst_ulp = 0; |
| float worst_x = 0; |
| float worst_val = 0; |
| float worst_ref = 0; |
| |
| for(float x = -10; x <= 10; x = inc(x)) |
| { |
| float val = Exp2(x); |
| |
| double ref = exp2((double)x); |
| |
| if(x == (int)x) |
| { |
| ASSERT_EQ(val, ref); |
| } |
| |
| const float tolerance = (3 + 2 * abs(x)); |
| float ulp = (float)ULP_32(ref, (double)val); |
| float margin = ulp / tolerance; |
| |
| if(margin > worst_margin) |
| { |
| worst_margin = margin; |
| worst_ulp = ulp; |
| worst_x = x; |
| worst_val = val; |
| worst_ref = ref; |
| } |
| } |
| |
| ASSERT_TRUE(worst_margin <= 1.0f) << " worst_x " << worst_x << " worst_val " << worst_val << " worst_ref " << worst_ref << " worst_ulp " << worst_ulp; |
| |
| CPUID::setDenormalsAreZero(false); |
| CPUID::setFlushToZero(false); |
| } |
| |
| // Polynomial approximation of order 5 for sin(x * 2 * pi) in the range [-1/4, 1/4] |
| static float sin5(float x) |
| { |
| // A * x^5 + B * x^3 + C * x |
| // Exact at x = 0, 1/12, 1/6, 1/4, and their negatives, which correspond to x * 2 * pi = 0, pi/6, pi/3, pi/2 |
| const float A = (36288 - 20736 * sqrt(3)) / 5; |
| const float B = 288 * sqrt(3) - 540; |
| const float C = (47 - 9 * sqrt(3)) / 5; |
| |
| float x2 = x * x; |
| |
| return ((A * x2 + B) * x2 + C) * x; |
| } |
| |
| TEST(MathTest, SinExhaustive) |
| { |
| const float tolerance = powf(2.0f, -12.0f); // Vulkan requires absolute error <= 2^−11 inside the range [−pi, pi] |
| const float pi = 3.1415926535f; |
| |
| for(float x = -pi; x <= pi; x = inc(x)) |
| { |
| // Range reduction and mirroring |
| float x_2 = 0.25f - x * (0.5f / pi); |
| float z = 0.25f - fabs(x_2 - round(x_2)); |
| |
| float val = sin5(z); |
| |
| ASSERT_NEAR(val, sinf(x), tolerance); |
| } |
| } |
| |
| TEST(MathTest, CosExhaustive) |
| { |
| const float tolerance = powf(2.0f, -12.0f); // Vulkan requires absolute error <= 2^−11 inside the range [−pi, pi] |
| const float pi = 3.1415926535f; |
| |
| for(float x = -pi; x <= pi; x = inc(x)) |
| { |
| // Phase shift, range reduction, and mirroring |
| float x_2 = x * (0.5f / pi); |
| float z = 0.25f - fabs(x_2 - round(x_2)); |
| |
| float val = sin5(z); |
| |
| ASSERT_NEAR(val, cosf(x), tolerance); |
| } |
| } |
| |
| TEST(MathTest, UnsignedFloat11_10) |
| { |
| // Test the largest value which causes underflow to 0, and the smallest value |
| // which produces a denormalized result. |
| |
| EXPECT_EQ(R11G11B10F::float32ToFloat11(bit_cast<float>(0x3500007F)), 0x0000); |
| EXPECT_EQ(R11G11B10F::float32ToFloat11(bit_cast<float>(0x35000080)), 0x0001); |
| |
| EXPECT_EQ(R11G11B10F::float32ToFloat10(bit_cast<float>(0x3580003F)), 0x0000); |
| EXPECT_EQ(R11G11B10F::float32ToFloat10(bit_cast<float>(0x35800040)), 0x0001); |
| } |
| |
| // Clamps to the [0, hi] range. NaN input produces 0, hi must be non-NaN. |
| float clamp0hi(float x, float hi) |
| { |
| // If x=NaN, x > 0 will compare false and we return 0. |
| if(!(x > 0)) |
| { |
| return 0; |
| } |
| |
| // x is non-NaN at this point, so std::min() is safe for non-NaN hi. |
| return std::min(x, hi); |
| } |
| |
| unsigned int RGB9E5_reference(float r, float g, float b) |
| { |
| // Vulkan 1.1.117 section 15.2.1 RGB to Shared Exponent Conversion |
| |
| // B is the exponent bias (15) |
| constexpr int g_sharedexp_bias = 15; |
| |
| // N is the number of mantissa bits per component (9) |
| constexpr int g_sharedexp_mantissabits = 9; |
| |
| // Emax is the maximum allowed biased exponent value (31) |
| constexpr int g_sharedexp_maxexponent = 31; |
| |
| constexpr float g_sharedexp_max = |
| ((static_cast<float>(1 << g_sharedexp_mantissabits) - 1) / |
| static_cast<float>(1 << g_sharedexp_mantissabits)) * |
| static_cast<float>(1 << (g_sharedexp_maxexponent - g_sharedexp_bias)); |
| |
| const float red_c = clamp0hi(r, g_sharedexp_max); |
| const float green_c = clamp0hi(g, g_sharedexp_max); |
| const float blue_c = clamp0hi(b, g_sharedexp_max); |
| |
| const float max_c = fmax(fmax(red_c, green_c), blue_c); |
| const float exp_p = fmax(-g_sharedexp_bias - 1, floor(log2(max_c))) + 1 + g_sharedexp_bias; |
| const int max_s = static_cast<int>(floor((max_c / exp2(exp_p - g_sharedexp_bias - g_sharedexp_mantissabits)) + 0.5f)); |
| const int exp_s = static_cast<int>((max_s < exp2(g_sharedexp_mantissabits)) ? exp_p : exp_p + 1); |
| |
| unsigned int R = static_cast<unsigned int>(floor((red_c / exp2(exp_s - g_sharedexp_bias - g_sharedexp_mantissabits)) + 0.5f)); |
| unsigned int G = static_cast<unsigned int>(floor((green_c / exp2(exp_s - g_sharedexp_bias - g_sharedexp_mantissabits)) + 0.5f)); |
| unsigned int B = static_cast<unsigned int>(floor((blue_c / exp2(exp_s - g_sharedexp_bias - g_sharedexp_mantissabits)) + 0.5f)); |
| unsigned int E = exp_s; |
| |
| return (E << 27) | (B << 18) | (G << 9) | R; |
| } |
| |
| TEST(MathTest, SharedExponentSparse) |
| { |
| for(uint64_t i = 0; i < 0x0000000100000000; i += 0x400) |
| { |
| float f = bit_cast<float>(i); |
| |
| unsigned int ref = RGB9E5_reference(f, 0.0f, 0.0f); |
| unsigned int val = RGB9E5(f, 0.0f, 0.0f); |
| |
| EXPECT_EQ(ref, val); |
| } |
| } |
| |
| TEST(MathTest, SharedExponentRandom) |
| { |
| srand(0); |
| |
| unsigned int x = 0; |
| unsigned int y = 0; |
| unsigned int z = 0; |
| |
| for(int i = 0; i < 10000000; i++) |
| { |
| float r = bit_cast<float>(x); |
| float g = bit_cast<float>(y); |
| float b = bit_cast<float>(z); |
| |
| unsigned int ref = RGB9E5_reference(r, g, b); |
| unsigned int val = RGB9E5(r, g, b); |
| |
| EXPECT_EQ(ref, val); |
| |
| x += rand(); |
| y += rand(); |
| z += rand(); |
| } |
| } |
| |
| TEST(MathTest, SharedExponentExhaustive) |
| { |
| for(uint64_t i = 0; i < 0x0000000100000000; i += 1) |
| { |
| float f = bit_cast<float>(i); |
| |
| unsigned int ref = RGB9E5_reference(f, 0.0f, 0.0f); |
| unsigned int val = RGB9E5(f, 0.0f, 0.0f); |
| |
| EXPECT_EQ(ref, val); |
| } |
| } |