Marat Dukhan | f44f022 | 2020-12-14 11:53:27 -0800 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <cmath> |
| 8 | #include <cstddef> |
| 9 | #include <cstdint> |
| 10 | #include <cstdlib> |
| 11 | #include <iomanip> |
| 12 | #include <ios> |
| 13 | #include <vector> |
| 14 | |
| 15 | #include <gtest/gtest.h> |
| 16 | |
| 17 | #include <fp16.h> |
| 18 | |
| 19 | #include <xnnpack/AlignedAllocator.h> |
| 20 | #include <xnnpack/common.h> |
| 21 | #include <xnnpack/isa-checks.h> |
| 22 | #include <xnnpack/math-stubs.h> |
| 23 | |
| 24 | |
| 25 | constexpr int kBlockSize = 1024; |
| 26 | |
| 27 | |
| 28 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 29 | TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, negative_zero) { |
| 30 | TEST_REQUIRES_ARM_NEON; |
| 31 | |
| 32 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 33 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 34 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 35 | xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 36 | const float reference_output = 0.0f; |
| 37 | ASSERT_EQ(reference_output, outputs[0]) |
| 38 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 39 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 40 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 41 | } |
| 42 | |
| 43 | TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, negative_saturation) { |
| 44 | TEST_REQUIRES_ARM_NEON; |
| 45 | |
| 46 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 47 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 48 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 49 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 50 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 51 | } |
| 52 | xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 53 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 54 | const float reference_output = -1.0f; |
| 55 | ASSERT_EQ(reference_output, outputs[i]) |
| 56 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 57 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 58 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 59 | } |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, positive_nan) { |
| 64 | TEST_REQUIRES_ARM_NEON; |
| 65 | |
| 66 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 67 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 68 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 69 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 70 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 71 | } |
| 72 | xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 73 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 74 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 75 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 76 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 77 | } |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | TEST(EXPM1MINUS__NEON_RR2_LUT16_P3, negative_nan) { |
| 82 | TEST_REQUIRES_ARM_NEON; |
| 83 | |
| 84 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 85 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 86 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 87 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 88 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 89 | } |
| 90 | xnn_math_f32_expm1minus__neon_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 91 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 92 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 93 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 94 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 95 | } |
| 96 | } |
| 97 | } |
| 98 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 99 | |
| 100 | |
| 101 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 102 | TEST(EXPM1MINUS__NEON_RR2_P6, negative_zero) { |
| 103 | TEST_REQUIRES_ARM_NEON; |
| 104 | |
| 105 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 106 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 107 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 108 | xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 109 | const float reference_output = 0.0f; |
| 110 | ASSERT_EQ(reference_output, outputs[0]) |
| 111 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 112 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 113 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 114 | } |
| 115 | |
| 116 | TEST(EXPM1MINUS__NEON_RR2_P6, negative_saturation) { |
| 117 | TEST_REQUIRES_ARM_NEON; |
| 118 | |
| 119 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 120 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 121 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 122 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 123 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 124 | } |
| 125 | xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 126 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 127 | const float reference_output = -1.0f; |
| 128 | ASSERT_EQ(reference_output, outputs[i]) |
| 129 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 130 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 131 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 132 | } |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | TEST(EXPM1MINUS__NEON_RR2_P6, positive_nan) { |
| 137 | TEST_REQUIRES_ARM_NEON; |
| 138 | |
| 139 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 140 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 141 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 142 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 143 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 144 | } |
| 145 | xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 146 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 147 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 148 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 149 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 150 | } |
| 151 | } |
| 152 | } |
| 153 | |
| 154 | TEST(EXPM1MINUS__NEON_RR2_P6, negative_nan) { |
| 155 | TEST_REQUIRES_ARM_NEON; |
| 156 | |
| 157 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 158 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 159 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 160 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 161 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 162 | } |
| 163 | xnn_math_f32_expm1minus__neon_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 164 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 165 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 166 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 167 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 168 | } |
| 169 | } |
| 170 | } |
| 171 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 172 | |
| 173 | |
| 174 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 175 | TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, negative_zero) { |
| 176 | TEST_REQUIRES_ARM_NEON_FMA; |
| 177 | |
| 178 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 179 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 180 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 181 | xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 182 | const float reference_output = 0.0f; |
| 183 | ASSERT_EQ(reference_output, outputs[0]) |
| 184 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 185 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 186 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 187 | } |
| 188 | |
| 189 | TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, negative_saturation) { |
| 190 | TEST_REQUIRES_ARM_NEON_FMA; |
| 191 | |
| 192 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 193 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 194 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 195 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 196 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 197 | } |
| 198 | xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 199 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 200 | const float reference_output = -1.0f; |
| 201 | ASSERT_EQ(reference_output, outputs[i]) |
| 202 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 203 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 204 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 205 | } |
| 206 | } |
| 207 | } |
| 208 | |
| 209 | TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, positive_nan) { |
| 210 | TEST_REQUIRES_ARM_NEON_FMA; |
| 211 | |
| 212 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 213 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 214 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 215 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 216 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 217 | } |
| 218 | xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 219 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 220 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 221 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 222 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 223 | } |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | TEST(EXPM1MINUS__NEONFMA_RR1_LUT16_P3, negative_nan) { |
| 228 | TEST_REQUIRES_ARM_NEON_FMA; |
| 229 | |
| 230 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 231 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 232 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 233 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 234 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 235 | } |
| 236 | xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 237 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 238 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 239 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 240 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 241 | } |
| 242 | } |
| 243 | } |
| 244 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 245 | |
| 246 | |
| 247 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 248 | TEST(EXPM1MINUS__NEONFMA_RR1_P6, negative_zero) { |
| 249 | TEST_REQUIRES_ARM_NEON_FMA; |
| 250 | |
| 251 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 252 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 253 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 254 | xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 255 | const float reference_output = 0.0f; |
| 256 | ASSERT_EQ(reference_output, outputs[0]) |
| 257 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 258 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 259 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 260 | } |
| 261 | |
| 262 | TEST(EXPM1MINUS__NEONFMA_RR1_P6, negative_saturation) { |
| 263 | TEST_REQUIRES_ARM_NEON_FMA; |
| 264 | |
| 265 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 266 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 267 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 268 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 269 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 270 | } |
| 271 | xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 272 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 273 | const float reference_output = -1.0f; |
| 274 | ASSERT_EQ(reference_output, outputs[i]) |
| 275 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 276 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 277 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 278 | } |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | TEST(EXPM1MINUS__NEONFMA_RR1_P6, positive_nan) { |
| 283 | TEST_REQUIRES_ARM_NEON_FMA; |
| 284 | |
| 285 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 286 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 287 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 288 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 289 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 290 | } |
| 291 | xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 292 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 293 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 294 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 295 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 296 | } |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | TEST(EXPM1MINUS__NEONFMA_RR1_P6, negative_nan) { |
| 301 | TEST_REQUIRES_ARM_NEON_FMA; |
| 302 | |
| 303 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 304 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 305 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 306 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 307 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 308 | } |
| 309 | xnn_math_f32_expm1minus__neonfma_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 310 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 311 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 312 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 313 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 314 | } |
| 315 | } |
| 316 | } |
| 317 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 318 | |
| 319 | |
| 320 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 321 | TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, negative_zero) { |
| 322 | TEST_REQUIRES_X86_AVX512F; |
| 323 | |
| 324 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 325 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 326 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 327 | xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 328 | const float reference_output = 0.0f; |
| 329 | ASSERT_EQ(reference_output, outputs[0]) |
| 330 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 331 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 332 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 333 | } |
| 334 | |
| 335 | TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, negative_saturation) { |
| 336 | TEST_REQUIRES_X86_AVX512F; |
| 337 | |
| 338 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 339 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 340 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 341 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 342 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 343 | } |
| 344 | xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 345 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 346 | const float reference_output = -1.0f; |
| 347 | ASSERT_EQ(reference_output, outputs[i]) |
| 348 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 349 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 350 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 351 | } |
| 352 | } |
| 353 | } |
| 354 | |
| 355 | TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, positive_nan) { |
| 356 | TEST_REQUIRES_X86_AVX512F; |
| 357 | |
| 358 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 359 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 360 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 361 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 362 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 363 | } |
| 364 | xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 365 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 366 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 367 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 368 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | TEST(EXPM1MINUS__AVX512F_RR1_LUT16_P3_PERM, negative_nan) { |
| 374 | TEST_REQUIRES_X86_AVX512F; |
| 375 | |
| 376 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 377 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 378 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 379 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 380 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 381 | } |
| 382 | xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 383 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 384 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 385 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 386 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 387 | } |
| 388 | } |
| 389 | } |
| 390 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 391 | |
| 392 | |
| 393 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 394 | TEST(EXPM1MINUS__AVX512F_RR1_P6, negative_zero) { |
| 395 | TEST_REQUIRES_X86_AVX512F; |
| 396 | |
| 397 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 398 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 399 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 400 | xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 401 | const float reference_output = 0.0f; |
| 402 | ASSERT_EQ(reference_output, outputs[0]) |
| 403 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 404 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 405 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 406 | } |
| 407 | |
| 408 | TEST(EXPM1MINUS__AVX512F_RR1_P6, negative_saturation) { |
| 409 | TEST_REQUIRES_X86_AVX512F; |
| 410 | |
| 411 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 412 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 413 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 414 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 415 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 416 | } |
| 417 | xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 418 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 419 | const float reference_output = -1.0f; |
| 420 | ASSERT_EQ(reference_output, outputs[i]) |
| 421 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 422 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 423 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 424 | } |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | TEST(EXPM1MINUS__AVX512F_RR1_P6, positive_nan) { |
| 429 | TEST_REQUIRES_X86_AVX512F; |
| 430 | |
| 431 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 432 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 433 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 434 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 435 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 436 | } |
| 437 | xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 438 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 439 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 440 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 441 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 442 | } |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | TEST(EXPM1MINUS__AVX512F_RR1_P6, negative_nan) { |
| 447 | TEST_REQUIRES_X86_AVX512F; |
| 448 | |
| 449 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 450 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 451 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 452 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 453 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 454 | } |
| 455 | xnn_math_f32_expm1minus__avx512f_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 456 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 457 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 458 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 459 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 460 | } |
| 461 | } |
| 462 | } |
| 463 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 464 | |
| 465 | |
| 466 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 467 | TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, negative_zero) { |
| 468 | TEST_REQUIRES_X86_AVX2; |
| 469 | |
| 470 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 471 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 472 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 473 | xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 474 | const float reference_output = 0.0f; |
| 475 | ASSERT_EQ(reference_output, outputs[0]) |
| 476 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 477 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 478 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 479 | } |
| 480 | |
| 481 | TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, negative_saturation) { |
| 482 | TEST_REQUIRES_X86_AVX2; |
| 483 | |
| 484 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 485 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 486 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 487 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 488 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 489 | } |
| 490 | xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 491 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 492 | const float reference_output = -1.0f; |
| 493 | ASSERT_EQ(reference_output, outputs[i]) |
| 494 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 495 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 496 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 497 | } |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, positive_nan) { |
| 502 | TEST_REQUIRES_X86_AVX2; |
| 503 | |
| 504 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 505 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 506 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 507 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 508 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 509 | } |
| 510 | xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 511 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 512 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 513 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 514 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 515 | } |
| 516 | } |
| 517 | } |
| 518 | |
| 519 | TEST(EXPM1MINUS__AVX2_RR1_LUT4_P4_PERM, negative_nan) { |
| 520 | TEST_REQUIRES_X86_AVX2; |
| 521 | |
| 522 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 523 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 524 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 525 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 526 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 527 | } |
| 528 | xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 529 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 530 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 531 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 532 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 533 | } |
| 534 | } |
| 535 | } |
| 536 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 537 | |
| 538 | |
| 539 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 540 | TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, negative_zero) { |
| 541 | TEST_REQUIRES_X86_AVX2; |
| 542 | |
| 543 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 544 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 545 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 546 | xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 547 | const float reference_output = 0.0f; |
| 548 | ASSERT_EQ(reference_output, outputs[0]) |
| 549 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 550 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 551 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 552 | } |
| 553 | |
| 554 | TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, negative_saturation) { |
| 555 | TEST_REQUIRES_X86_AVX2; |
| 556 | |
| 557 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 558 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 559 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 560 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 561 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 562 | } |
| 563 | xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 564 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 565 | const float reference_output = -1.0f; |
| 566 | ASSERT_EQ(reference_output, outputs[i]) |
| 567 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 568 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 569 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 570 | } |
| 571 | } |
| 572 | } |
| 573 | |
| 574 | TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, positive_nan) { |
| 575 | TEST_REQUIRES_X86_AVX2; |
| 576 | |
| 577 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 578 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 579 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 580 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 581 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 582 | } |
| 583 | xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 584 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 585 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 586 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 587 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 588 | } |
| 589 | } |
| 590 | } |
| 591 | |
| 592 | TEST(EXPM1MINUS__AVX2_RR1_LUT8_P4_PERM, negative_nan) { |
| 593 | TEST_REQUIRES_X86_AVX2; |
| 594 | |
| 595 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 596 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 597 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 598 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 599 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 600 | } |
| 601 | xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 602 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 603 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 604 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 605 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 606 | } |
| 607 | } |
| 608 | } |
| 609 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 610 | |
| 611 | |
| 612 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 613 | TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, negative_zero) { |
| 614 | TEST_REQUIRES_X86_AVX2; |
| 615 | |
| 616 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 617 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 618 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 619 | xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 620 | const float reference_output = 0.0f; |
| 621 | ASSERT_EQ(reference_output, outputs[0]) |
| 622 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 623 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 624 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 625 | } |
| 626 | |
| 627 | TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, negative_saturation) { |
| 628 | TEST_REQUIRES_X86_AVX2; |
| 629 | |
| 630 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 631 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 632 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 633 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 634 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 635 | } |
| 636 | xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 637 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 638 | const float reference_output = -1.0f; |
| 639 | ASSERT_EQ(reference_output, outputs[i]) |
| 640 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 641 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 642 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 643 | } |
| 644 | } |
| 645 | } |
| 646 | |
| 647 | TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, positive_nan) { |
| 648 | TEST_REQUIRES_X86_AVX2; |
| 649 | |
| 650 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 651 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 652 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 653 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 654 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 655 | } |
| 656 | xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 657 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 658 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 659 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 660 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 661 | } |
| 662 | } |
| 663 | } |
| 664 | |
| 665 | TEST(EXPM1MINUS__AVX2_RR1_LUT16_P3_GATHER, negative_nan) { |
| 666 | TEST_REQUIRES_X86_AVX2; |
| 667 | |
| 668 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 669 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 670 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 671 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 672 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 673 | } |
| 674 | xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 675 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 676 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 677 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 678 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 679 | } |
| 680 | } |
| 681 | } |
| 682 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 683 | |
| 684 | |
| 685 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 686 | TEST(EXPM1MINUS__AVX2_RR1_P6, negative_zero) { |
| 687 | TEST_REQUIRES_X86_AVX2; |
| 688 | |
| 689 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 690 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 691 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 692 | xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 693 | const float reference_output = 0.0f; |
| 694 | ASSERT_EQ(reference_output, outputs[0]) |
| 695 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 696 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 697 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 698 | } |
| 699 | |
| 700 | TEST(EXPM1MINUS__AVX2_RR1_P6, negative_saturation) { |
| 701 | TEST_REQUIRES_X86_AVX2; |
| 702 | |
| 703 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 704 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 705 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 706 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 707 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 708 | } |
| 709 | xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 710 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 711 | const float reference_output = -1.0f; |
| 712 | ASSERT_EQ(reference_output, outputs[i]) |
| 713 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 714 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 715 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 716 | } |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | TEST(EXPM1MINUS__AVX2_RR1_P6, positive_nan) { |
| 721 | TEST_REQUIRES_X86_AVX2; |
| 722 | |
| 723 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 724 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 725 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 726 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 727 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 728 | } |
| 729 | xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 730 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 731 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 732 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 733 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 734 | } |
| 735 | } |
| 736 | } |
| 737 | |
| 738 | TEST(EXPM1MINUS__AVX2_RR1_P6, negative_nan) { |
| 739 | TEST_REQUIRES_X86_AVX2; |
| 740 | |
| 741 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 742 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 743 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 744 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 745 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 746 | } |
| 747 | xnn_math_f32_expm1minus__avx2_rr1_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 748 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 749 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 750 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 751 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 752 | } |
| 753 | } |
| 754 | } |
| 755 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 756 | |
| 757 | |
| 758 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 759 | TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, negative_zero) { |
| 760 | TEST_REQUIRES_X86_AVX; |
| 761 | |
| 762 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 763 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 764 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 765 | xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 766 | const float reference_output = 0.0f; |
| 767 | ASSERT_EQ(reference_output, outputs[0]) |
| 768 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 769 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 770 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 771 | } |
| 772 | |
| 773 | TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, negative_saturation) { |
| 774 | TEST_REQUIRES_X86_AVX; |
| 775 | |
| 776 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 777 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 778 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 779 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 780 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 781 | } |
| 782 | xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 783 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 784 | const float reference_output = -1.0f; |
| 785 | ASSERT_EQ(reference_output, outputs[i]) |
| 786 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 787 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 788 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 789 | } |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, positive_nan) { |
| 794 | TEST_REQUIRES_X86_AVX; |
| 795 | |
| 796 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 797 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 798 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 799 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 800 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 801 | } |
| 802 | xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 803 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 804 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 805 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 806 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 807 | } |
| 808 | } |
| 809 | } |
| 810 | |
| 811 | TEST(EXPM1MINUS__AVX_RR2_LUT4_P4_PERM, negative_nan) { |
| 812 | TEST_REQUIRES_X86_AVX; |
| 813 | |
| 814 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 815 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 816 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 817 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 818 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 819 | } |
| 820 | xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 821 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 822 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 823 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 824 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 825 | } |
| 826 | } |
| 827 | } |
| 828 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 829 | |
| 830 | |
| 831 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 832 | TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, negative_zero) { |
| 833 | TEST_REQUIRES_X86_AVX; |
| 834 | |
| 835 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 836 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 837 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 838 | xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 839 | const float reference_output = 0.0f; |
| 840 | ASSERT_EQ(reference_output, outputs[0]) |
| 841 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 842 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 843 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 844 | } |
| 845 | |
| 846 | TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, negative_saturation) { |
| 847 | TEST_REQUIRES_X86_AVX; |
| 848 | |
| 849 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 850 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 851 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 852 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 853 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 854 | } |
| 855 | xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 856 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 857 | const float reference_output = -1.0f; |
| 858 | ASSERT_EQ(reference_output, outputs[i]) |
| 859 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 860 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 861 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 862 | } |
| 863 | } |
| 864 | } |
| 865 | |
| 866 | TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, positive_nan) { |
| 867 | TEST_REQUIRES_X86_AVX; |
| 868 | |
| 869 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 870 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 871 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 872 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 873 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 874 | } |
| 875 | xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 876 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 877 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 878 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 879 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 880 | } |
| 881 | } |
| 882 | } |
| 883 | |
| 884 | TEST(EXPM1MINUS__AVX_RR2_LUT16_P3, negative_nan) { |
| 885 | TEST_REQUIRES_X86_AVX; |
| 886 | |
| 887 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 888 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 889 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 890 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 891 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 892 | } |
| 893 | xnn_math_f32_expm1minus__avx_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 894 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 895 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 896 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 897 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 898 | } |
| 899 | } |
| 900 | } |
| 901 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 902 | |
| 903 | |
| 904 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 905 | TEST(EXPM1MINUS__AVX_RR2_P6, negative_zero) { |
| 906 | TEST_REQUIRES_X86_AVX; |
| 907 | |
| 908 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 909 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 910 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 911 | xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 912 | const float reference_output = 0.0f; |
| 913 | ASSERT_EQ(reference_output, outputs[0]) |
| 914 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 915 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 916 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 917 | } |
| 918 | |
| 919 | TEST(EXPM1MINUS__AVX_RR2_P6, negative_saturation) { |
| 920 | TEST_REQUIRES_X86_AVX; |
| 921 | |
| 922 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 923 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 924 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 925 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 926 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 927 | } |
| 928 | xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 929 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 930 | const float reference_output = -1.0f; |
| 931 | ASSERT_EQ(reference_output, outputs[i]) |
| 932 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 933 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 934 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 935 | } |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | TEST(EXPM1MINUS__AVX_RR2_P6, positive_nan) { |
| 940 | TEST_REQUIRES_X86_AVX; |
| 941 | |
| 942 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 943 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 944 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 945 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 946 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 947 | } |
| 948 | xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 949 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 950 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 951 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 952 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 953 | } |
| 954 | } |
| 955 | } |
| 956 | |
| 957 | TEST(EXPM1MINUS__AVX_RR2_P6, negative_nan) { |
| 958 | TEST_REQUIRES_X86_AVX; |
| 959 | |
| 960 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 961 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 962 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 963 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 964 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 965 | } |
| 966 | xnn_math_f32_expm1minus__avx_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 967 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 968 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 969 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 970 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 971 | } |
| 972 | } |
| 973 | } |
| 974 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 975 | |
| 976 | |
| 977 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 978 | TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, negative_zero) { |
| 979 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 980 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 981 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 982 | xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 983 | const float reference_output = 0.0f; |
| 984 | ASSERT_EQ(reference_output, outputs[0]) |
| 985 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 986 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 987 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 988 | } |
| 989 | |
| 990 | TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, negative_saturation) { |
| 991 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 992 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 993 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 994 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 995 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 996 | } |
| 997 | xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 998 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 999 | const float reference_output = -1.0f; |
| 1000 | ASSERT_EQ(reference_output, outputs[i]) |
| 1001 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1002 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1003 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1004 | } |
| 1005 | } |
| 1006 | } |
| 1007 | |
| 1008 | TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, positive_nan) { |
| 1009 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1010 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1011 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1012 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1013 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1014 | } |
| 1015 | xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1016 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1017 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1018 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1019 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1020 | } |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | TEST(EXPM1MINUS__SSE2_RR2_LUT16_P3, negative_nan) { |
| 1025 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1026 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1027 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1028 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1029 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1030 | } |
| 1031 | xnn_math_f32_expm1minus__sse2_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1032 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1033 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1034 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1035 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1036 | } |
| 1037 | } |
| 1038 | } |
| 1039 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1040 | |
| 1041 | |
| 1042 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1043 | TEST(EXPM1MINUS__SSE2_RR2_P6, negative_zero) { |
| 1044 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1045 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1046 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1047 | xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1048 | const float reference_output = 0.0f; |
| 1049 | ASSERT_EQ(reference_output, outputs[0]) |
| 1050 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1051 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1052 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1053 | } |
| 1054 | |
| 1055 | TEST(EXPM1MINUS__SSE2_RR2_P6, negative_saturation) { |
| 1056 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1057 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1058 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1059 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1060 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1061 | } |
| 1062 | xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1063 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1064 | const float reference_output = -1.0f; |
| 1065 | ASSERT_EQ(reference_output, outputs[i]) |
| 1066 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1067 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1068 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1069 | } |
| 1070 | } |
| 1071 | } |
| 1072 | |
| 1073 | TEST(EXPM1MINUS__SSE2_RR2_P6, positive_nan) { |
| 1074 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1075 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1076 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1077 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1078 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1079 | } |
| 1080 | xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1081 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1082 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1083 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1084 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1085 | } |
| 1086 | } |
| 1087 | } |
| 1088 | |
| 1089 | TEST(EXPM1MINUS__SSE2_RR2_P6, negative_nan) { |
| 1090 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1091 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1092 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1093 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1094 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1095 | } |
| 1096 | xnn_math_f32_expm1minus__sse2_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1097 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1098 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1099 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1100 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1101 | } |
| 1102 | } |
| 1103 | } |
| 1104 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1105 | |
| 1106 | |
| 1107 | #if XNN_ARCH_WASMSIMD |
| 1108 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, negative_zero) { |
| 1109 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1110 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1111 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1112 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1113 | const float reference_output = 0.0f; |
| 1114 | ASSERT_EQ(reference_output, outputs[0]) |
| 1115 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1116 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1117 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1118 | } |
| 1119 | |
| 1120 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, negative_saturation) { |
| 1121 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1122 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1123 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1124 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1125 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1126 | } |
| 1127 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1128 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1129 | const float reference_output = -1.0f; |
| 1130 | ASSERT_EQ(reference_output, outputs[i]) |
| 1131 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1132 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1133 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1134 | } |
| 1135 | } |
| 1136 | } |
| 1137 | |
| 1138 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, positive_nan) { |
| 1139 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1140 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1141 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1142 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1143 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1144 | } |
| 1145 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1146 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1147 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1148 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1149 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1150 | } |
| 1151 | } |
| 1152 | } |
| 1153 | |
| 1154 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_ANDNOT, negative_nan) { |
| 1155 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1156 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1157 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1158 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1159 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1160 | } |
| 1161 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1162 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1163 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1164 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1165 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1166 | } |
| 1167 | } |
| 1168 | } |
| 1169 | #endif // XNN_ARCH_WASMSIMD |
| 1170 | |
| 1171 | |
| 1172 | #if XNN_ARCH_WASMSIMD |
| 1173 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, negative_zero) { |
| 1174 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1175 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1176 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1177 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1178 | const float reference_output = 0.0f; |
| 1179 | ASSERT_EQ(reference_output, outputs[0]) |
| 1180 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1181 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1182 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1183 | } |
| 1184 | |
| 1185 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, negative_saturation) { |
| 1186 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1187 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1188 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1189 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1190 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1191 | } |
| 1192 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1193 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1194 | const float reference_output = -1.0f; |
| 1195 | ASSERT_EQ(reference_output, outputs[i]) |
| 1196 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1197 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1198 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1199 | } |
| 1200 | } |
| 1201 | } |
| 1202 | |
| 1203 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, positive_nan) { |
| 1204 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1205 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1206 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1207 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1208 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1209 | } |
| 1210 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1211 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1212 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1213 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1214 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1215 | } |
| 1216 | } |
| 1217 | } |
| 1218 | |
| 1219 | TEST(EXPM1MINUS__WASMSIMD_RR2_LUT16_P3_MAX, negative_nan) { |
| 1220 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1221 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1222 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1223 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1224 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1225 | } |
| 1226 | xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1227 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1228 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1229 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1230 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1231 | } |
| 1232 | } |
| 1233 | } |
| 1234 | #endif // XNN_ARCH_WASMSIMD |
| 1235 | |
| 1236 | |
| 1237 | #if XNN_ARCH_WASMSIMD |
| 1238 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, negative_zero) { |
| 1239 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1240 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1241 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1242 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1243 | const float reference_output = 0.0f; |
| 1244 | ASSERT_EQ(reference_output, outputs[0]) |
| 1245 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1246 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1247 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1248 | } |
| 1249 | |
| 1250 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, negative_saturation) { |
| 1251 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1252 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1253 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1254 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1255 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1256 | } |
| 1257 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1258 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1259 | const float reference_output = -1.0f; |
| 1260 | ASSERT_EQ(reference_output, outputs[i]) |
| 1261 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1262 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1263 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1264 | } |
| 1265 | } |
| 1266 | } |
| 1267 | |
| 1268 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, positive_nan) { |
| 1269 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1270 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1271 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1272 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1273 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1274 | } |
| 1275 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1276 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1277 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1278 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1279 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1280 | } |
| 1281 | } |
| 1282 | } |
| 1283 | |
| 1284 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_ANDNOT, negative_nan) { |
| 1285 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1286 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1287 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1288 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1289 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1290 | } |
| 1291 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1292 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1293 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1294 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1295 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1296 | } |
| 1297 | } |
| 1298 | } |
| 1299 | #endif // XNN_ARCH_WASMSIMD |
| 1300 | |
| 1301 | |
| 1302 | #if XNN_ARCH_WASMSIMD |
| 1303 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, negative_zero) { |
| 1304 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1305 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1306 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1307 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1308 | const float reference_output = 0.0f; |
| 1309 | ASSERT_EQ(reference_output, outputs[0]) |
| 1310 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1311 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1312 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1313 | } |
| 1314 | |
| 1315 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, negative_saturation) { |
| 1316 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1317 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1318 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1319 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1320 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1321 | } |
| 1322 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1323 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1324 | const float reference_output = -1.0f; |
| 1325 | ASSERT_EQ(reference_output, outputs[i]) |
| 1326 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1327 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1328 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1329 | } |
| 1330 | } |
| 1331 | } |
| 1332 | |
| 1333 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, positive_nan) { |
| 1334 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1335 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1336 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1337 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1338 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1339 | } |
| 1340 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1341 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1342 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1343 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1344 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1345 | } |
| 1346 | } |
| 1347 | } |
| 1348 | |
| 1349 | TEST(EXPM1MINUS__WASMSIMD_RR2_P6_MAX, negative_nan) { |
| 1350 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1351 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1352 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1353 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1354 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1355 | } |
| 1356 | xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1357 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1358 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1359 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1360 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1361 | } |
| 1362 | } |
| 1363 | } |
| 1364 | #endif // XNN_ARCH_WASMSIMD |
| 1365 | |
| 1366 | |
| 1367 | TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, negative_zero) { |
| 1368 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1369 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1370 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1371 | xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1372 | const float reference_output = 0.0f; |
| 1373 | ASSERT_EQ(reference_output, outputs[0]) |
| 1374 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1375 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1376 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1377 | } |
| 1378 | |
| 1379 | TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, negative_saturation) { |
| 1380 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1381 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1382 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1383 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1384 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1385 | } |
| 1386 | xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1387 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1388 | const float reference_output = -1.0f; |
| 1389 | ASSERT_EQ(reference_output, outputs[i]) |
| 1390 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1391 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1392 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1393 | } |
| 1394 | } |
| 1395 | } |
| 1396 | |
| 1397 | TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, positive_nan) { |
| 1398 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1399 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1400 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1401 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1402 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1403 | } |
| 1404 | xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1405 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1406 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1407 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1408 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1409 | } |
| 1410 | } |
| 1411 | } |
| 1412 | |
| 1413 | TEST(EXPM1MINUS__SCALAR_RR2_LUT4_P4, negative_nan) { |
| 1414 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1415 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1416 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1417 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1418 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1419 | } |
| 1420 | xnn_math_f32_expm1minus__scalar_rr2_lut4_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1421 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1422 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1423 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1424 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1425 | } |
| 1426 | } |
| 1427 | } |
| 1428 | |
| 1429 | |
| 1430 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, negative_zero) { |
| 1431 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1432 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1433 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1434 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1435 | const float reference_output = 0.0f; |
| 1436 | ASSERT_EQ(reference_output, outputs[0]) |
| 1437 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1438 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1439 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1440 | } |
| 1441 | |
| 1442 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, negative_saturation) { |
| 1443 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1444 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1445 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1446 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1447 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1448 | } |
| 1449 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1450 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1451 | const float reference_output = -1.0f; |
| 1452 | ASSERT_EQ(reference_output, outputs[i]) |
| 1453 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1454 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1455 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1456 | } |
| 1457 | } |
| 1458 | } |
| 1459 | |
| 1460 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, positive_nan) { |
| 1461 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1462 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1463 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1464 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1465 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1466 | } |
| 1467 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1468 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1469 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1470 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1471 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1472 | } |
| 1473 | } |
| 1474 | } |
| 1475 | |
| 1476 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P3, negative_nan) { |
| 1477 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1478 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1479 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1480 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1481 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1482 | } |
| 1483 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1484 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1485 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1486 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1487 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1488 | } |
| 1489 | } |
| 1490 | } |
| 1491 | |
| 1492 | |
| 1493 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, negative_zero) { |
| 1494 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1495 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1496 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1497 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1498 | const float reference_output = 0.0f; |
| 1499 | ASSERT_EQ(reference_output, outputs[0]) |
| 1500 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1501 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1502 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1503 | } |
| 1504 | |
| 1505 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, negative_saturation) { |
| 1506 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1507 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1508 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1509 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1510 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1511 | } |
| 1512 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1513 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1514 | const float reference_output = -1.0f; |
| 1515 | ASSERT_EQ(reference_output, outputs[i]) |
| 1516 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1517 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1518 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1519 | } |
| 1520 | } |
| 1521 | } |
| 1522 | |
| 1523 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, positive_nan) { |
| 1524 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1525 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1526 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1527 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1528 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1529 | } |
| 1530 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1531 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1532 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1533 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1534 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1535 | } |
| 1536 | } |
| 1537 | } |
| 1538 | |
| 1539 | TEST(EXPM1MINUS__SCALAR_RR2_LUT8_P4, negative_nan) { |
| 1540 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1541 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1542 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1543 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1544 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1545 | } |
| 1546 | xnn_math_f32_expm1minus__scalar_rr2_lut8_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1547 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1548 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1549 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1550 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1551 | } |
| 1552 | } |
| 1553 | } |
| 1554 | |
| 1555 | |
| 1556 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, negative_zero) { |
| 1557 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1558 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1559 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1560 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1561 | const float reference_output = 0.0f; |
| 1562 | ASSERT_EQ(reference_output, outputs[0]) |
| 1563 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1564 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1565 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1566 | } |
| 1567 | |
| 1568 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, negative_saturation) { |
| 1569 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1570 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1571 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1572 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1573 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1574 | } |
| 1575 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1576 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1577 | const float reference_output = -1.0f; |
| 1578 | ASSERT_EQ(reference_output, outputs[i]) |
| 1579 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1580 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1581 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1582 | } |
| 1583 | } |
| 1584 | } |
| 1585 | |
| 1586 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, positive_nan) { |
| 1587 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1588 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1589 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1590 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1591 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1592 | } |
| 1593 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1594 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1595 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1596 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1597 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1598 | } |
| 1599 | } |
| 1600 | } |
| 1601 | |
| 1602 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P3, negative_nan) { |
| 1603 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1604 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1605 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1606 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1607 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1608 | } |
| 1609 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p3(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1610 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1611 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1612 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1613 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1614 | } |
| 1615 | } |
| 1616 | } |
| 1617 | |
| 1618 | |
| 1619 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, negative_zero) { |
| 1620 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1621 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1622 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1623 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1624 | const float reference_output = 0.0f; |
| 1625 | ASSERT_EQ(reference_output, outputs[0]) |
| 1626 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1627 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1628 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1629 | } |
| 1630 | |
| 1631 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, negative_saturation) { |
| 1632 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1633 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1634 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1635 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1636 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1637 | } |
| 1638 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1639 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1640 | const float reference_output = -1.0f; |
| 1641 | ASSERT_EQ(reference_output, outputs[i]) |
| 1642 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1643 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1644 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1645 | } |
| 1646 | } |
| 1647 | } |
| 1648 | |
| 1649 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, positive_nan) { |
| 1650 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1651 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1652 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1653 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1654 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1655 | } |
| 1656 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1657 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1658 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1659 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1660 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1661 | } |
| 1662 | } |
| 1663 | } |
| 1664 | |
| 1665 | TEST(EXPM1MINUS__SCALAR_RR2_LUT16_P4, negative_nan) { |
| 1666 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1667 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1668 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1669 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1670 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1671 | } |
| 1672 | xnn_math_f32_expm1minus__scalar_rr2_lut16_p4(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1673 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1674 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1675 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1676 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1677 | } |
| 1678 | } |
| 1679 | } |
| 1680 | |
| 1681 | |
| 1682 | TEST(EXPM1MINUS__SCALAR_RR2_P5, negative_zero) { |
| 1683 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1684 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1685 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1686 | xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1687 | const float reference_output = 0.0f; |
| 1688 | ASSERT_EQ(reference_output, outputs[0]) |
| 1689 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1690 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1691 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1692 | } |
| 1693 | |
| 1694 | TEST(EXPM1MINUS__SCALAR_RR2_P5, negative_saturation) { |
| 1695 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1696 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1697 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1698 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1699 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1700 | } |
| 1701 | xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1702 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1703 | const float reference_output = -1.0f; |
| 1704 | ASSERT_EQ(reference_output, outputs[i]) |
| 1705 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1706 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1707 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1708 | } |
| 1709 | } |
| 1710 | } |
| 1711 | |
| 1712 | TEST(EXPM1MINUS__SCALAR_RR2_P5, positive_nan) { |
| 1713 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1714 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1715 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1716 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1717 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1718 | } |
| 1719 | xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1720 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1721 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1722 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1723 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1724 | } |
| 1725 | } |
| 1726 | } |
| 1727 | |
| 1728 | TEST(EXPM1MINUS__SCALAR_RR2_P5, negative_nan) { |
| 1729 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1730 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1731 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1732 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1733 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1734 | } |
| 1735 | xnn_math_f32_expm1minus__scalar_rr2_p5(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1736 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1737 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1738 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1739 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1740 | } |
| 1741 | } |
| 1742 | } |
| 1743 | |
| 1744 | |
| 1745 | TEST(EXPM1MINUS__SCALAR_RR2_P6, negative_zero) { |
| 1746 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1747 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1748 | std::fill(inputs.begin(), inputs.end(), -0.0f); |
| 1749 | xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1750 | const float reference_output = 0.0f; |
| 1751 | ASSERT_EQ(reference_output, outputs[0]) |
| 1752 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[0]) |
| 1753 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1754 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[0]); |
| 1755 | } |
| 1756 | |
| 1757 | TEST(EXPM1MINUS__SCALAR_RR2_P6, negative_saturation) { |
| 1758 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1759 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1760 | for (uint32_t n = UINT32_C(0xC18AA123); n <= UINT32_C(0xFF800000); n += kBlockSize) { |
| 1761 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1762 | inputs[i] = fp32_from_bits(std::min(n + i, UINT32_C(0xFF800000))); |
| 1763 | } |
| 1764 | xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1765 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1766 | const float reference_output = -1.0f; |
| 1767 | ASSERT_EQ(reference_output, outputs[i]) |
| 1768 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1769 | << ", reference = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(reference_output) |
| 1770 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1771 | } |
| 1772 | } |
| 1773 | } |
| 1774 | |
| 1775 | TEST(EXPM1MINUS__SCALAR_RR2_P6, positive_nan) { |
| 1776 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1777 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1778 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1779 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1780 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), n + i)); |
| 1781 | } |
| 1782 | xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1783 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1784 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1785 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1786 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1787 | } |
| 1788 | } |
| 1789 | } |
| 1790 | |
| 1791 | TEST(EXPM1MINUS__SCALAR_RR2_P6, negative_nan) { |
| 1792 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 1793 | std::vector<float, AlignedAllocator<float, 64>> outputs(kBlockSize); |
| 1794 | for (uint32_t n = UINT32_C(0x7F800001); n < UINT32_C(0x80000000); n += kBlockSize) { |
| 1795 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1796 | inputs[i] = fp32_from_bits(std::min(UINT32_C(0x7FFFFFFF), UINT32_C(0x80000000) | (n + i))); |
| 1797 | } |
| 1798 | xnn_math_f32_expm1minus__scalar_rr2_p6(kBlockSize * sizeof(float), inputs.data(), outputs.data()); |
| 1799 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 1800 | ASSERT_TRUE(std::isnan(outputs[i])) |
| 1801 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 1802 | << ", optimized = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(outputs[i]); |
| 1803 | } |
| 1804 | } |
| 1805 | } |