Marat Dukhan | d24301d | 2021-12-02 00:13:45 -0800 | [diff] [blame] | 1 | // Copyright 2021 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 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 28 | TEST(CVT__NEON, positive_normal) { |
| 29 | TEST_REQUIRES_ARM_NEON; |
| 30 | |
| 31 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 32 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 33 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 34 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 35 | zero_point++) |
| 36 | { |
| 37 | const uint32_t max_input = fp32_to_bits((float) (std::numeric_limits<int8_t>::max() - zero_point)); |
| 38 | for (uint32_t n = 0; n < max_input; n += kBlockSize) { |
| 39 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 40 | inputs[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input)); |
| 41 | } |
| 42 | xnn_math_f32_qs8_cvt__neon(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 43 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 44 | long reference_output = std::lrintf(inputs[i]) + long(zero_point); |
| 45 | if (inputs[i] >= float(std::numeric_limits<long>::max())) { |
| 46 | reference_output = std::numeric_limits<int8_t>::max(); |
| 47 | } else if (inputs[i] <= float(std::numeric_limits<long>::min())) { |
| 48 | reference_output = std::numeric_limits<int8_t>::min(); |
| 49 | } |
| 50 | ASSERT_EQ(reference_output, long(outputs[i])) |
| 51 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 52 | << ", reference = " << std::dec << reference_output |
| 53 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 54 | << ", zero point = " << std::dec << zero_point; |
| 55 | } |
| 56 | } |
| 57 | } |
| 58 | } |
| 59 | |
| 60 | TEST(CVT__NEON, negative_normal) { |
| 61 | TEST_REQUIRES_ARM_NEON; |
| 62 | |
| 63 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 64 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 65 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 66 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 67 | zero_point++) |
| 68 | { |
| 69 | const uint32_t max_input = fp32_to_bits((float) (zero_point - std::numeric_limits<int8_t>::min())); |
| 70 | for (uint32_t n = 0; n < max_input; n += kBlockSize) { |
| 71 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 72 | inputs[i] = fp32_from_bits(UINT32_C(0x80000000) | std::min<uint32_t>(n + i, max_input)); |
| 73 | } |
| 74 | xnn_math_f32_qs8_cvt__neon(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 75 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 76 | long reference_output = std::lrintf(inputs[i]) + long(zero_point); |
| 77 | if (inputs[i] >= float(std::numeric_limits<long>::max())) { |
| 78 | reference_output = std::numeric_limits<int8_t>::max(); |
| 79 | } else if (inputs[i] <= float(std::numeric_limits<long>::min())) { |
| 80 | reference_output = std::numeric_limits<int8_t>::min(); |
| 81 | } |
| 82 | ASSERT_EQ(reference_output, long(outputs[i])) |
| 83 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 84 | << ", reference = " << std::dec << reference_output |
| 85 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 86 | << ", zero point = " << std::dec << zero_point; |
| 87 | } |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | TEST(CVT__NEON, positive_saturation) { |
| 93 | TEST_REQUIRES_ARM_NEON; |
| 94 | |
| 95 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 96 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 97 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 98 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 99 | zero_point++) |
| 100 | { |
| 101 | const uint32_t min_input = fp32_to_bits((float) (std::numeric_limits<int8_t>::max() - zero_point)); |
| 102 | const uint32_t max_input = UINT32_C(0x7F800000); |
| 103 | for (uint32_t n = min_input; n < max_input; n += kBlockSize) { |
| 104 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 105 | inputs[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input)); |
| 106 | } |
| 107 | xnn_math_f32_qs8_cvt__neon(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 108 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 109 | const int32_t reference_output = std::numeric_limits<int8_t>::max(); |
| 110 | ASSERT_EQ(reference_output, int32_t(outputs[i])) |
| 111 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 112 | << ", reference = " << std::dec << reference_output |
| 113 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 114 | << ", zero point = " << std::dec << zero_point; |
| 115 | } |
| 116 | } |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | TEST(CVT__NEON, negative_saturation) { |
| 121 | TEST_REQUIRES_ARM_NEON; |
| 122 | |
| 123 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 124 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 125 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 126 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 127 | zero_point++) |
| 128 | { |
| 129 | const uint32_t min_input = fp32_to_bits((float) (zero_point - std::numeric_limits<int8_t>::min())); |
| 130 | const uint32_t max_input = UINT32_C(0x7F800000); |
| 131 | for (uint32_t n = min_input; n < max_input; n += kBlockSize) { |
| 132 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 133 | inputs[i] = fp32_from_bits(UINT32_C(0x80000000) | std::min<uint32_t>(n + i, max_input)); |
| 134 | } |
| 135 | xnn_math_f32_qs8_cvt__neon(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 136 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 137 | const int32_t reference_output = std::numeric_limits<int8_t>::min(); |
| 138 | ASSERT_EQ(reference_output, int32_t(outputs[i])) |
| 139 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 140 | << ", reference = " << std::dec << reference_output |
| 141 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 142 | << ", zero point = " << std::dec << zero_point; |
| 143 | } |
| 144 | } |
| 145 | } |
| 146 | } |
| 147 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 148 | |
| 149 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 150 | TEST(CVT__NEONV8, positive_normal) { |
| 151 | TEST_REQUIRES_ARM_NEON_V8; |
| 152 | |
| 153 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 154 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 155 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 156 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 157 | zero_point++) |
| 158 | { |
| 159 | const uint32_t max_input = fp32_to_bits((float) (std::numeric_limits<int8_t>::max() - zero_point)); |
| 160 | for (uint32_t n = 0; n < max_input; n += kBlockSize) { |
| 161 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 162 | inputs[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input)); |
| 163 | } |
| 164 | xnn_math_f32_qs8_cvt__neonv8(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 165 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 166 | long reference_output = std::lrintf(inputs[i]) + long(zero_point); |
| 167 | if (inputs[i] >= float(std::numeric_limits<long>::max())) { |
| 168 | reference_output = std::numeric_limits<int8_t>::max(); |
| 169 | } else if (inputs[i] <= float(std::numeric_limits<long>::min())) { |
| 170 | reference_output = std::numeric_limits<int8_t>::min(); |
| 171 | } |
| 172 | ASSERT_EQ(reference_output, long(outputs[i])) |
| 173 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 174 | << ", reference = " << std::dec << reference_output |
| 175 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 176 | << ", zero point = " << std::dec << zero_point; |
| 177 | } |
| 178 | } |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | TEST(CVT__NEONV8, negative_normal) { |
| 183 | TEST_REQUIRES_ARM_NEON_V8; |
| 184 | |
| 185 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 186 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 187 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 188 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 189 | zero_point++) |
| 190 | { |
| 191 | const uint32_t max_input = fp32_to_bits((float) (zero_point - std::numeric_limits<int8_t>::min())); |
| 192 | for (uint32_t n = 0; n < max_input; n += kBlockSize) { |
| 193 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 194 | inputs[i] = fp32_from_bits(UINT32_C(0x80000000) | std::min<uint32_t>(n + i, max_input)); |
| 195 | } |
| 196 | xnn_math_f32_qs8_cvt__neonv8(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 197 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 198 | long reference_output = std::lrintf(inputs[i]) + long(zero_point); |
| 199 | if (inputs[i] >= float(std::numeric_limits<long>::max())) { |
| 200 | reference_output = std::numeric_limits<int8_t>::max(); |
| 201 | } else if (inputs[i] <= float(std::numeric_limits<long>::min())) { |
| 202 | reference_output = std::numeric_limits<int8_t>::min(); |
| 203 | } |
| 204 | ASSERT_EQ(reference_output, long(outputs[i])) |
| 205 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 206 | << ", reference = " << std::dec << reference_output |
| 207 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 208 | << ", zero point = " << std::dec << zero_point; |
| 209 | } |
| 210 | } |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | TEST(CVT__NEONV8, positive_saturation) { |
| 215 | TEST_REQUIRES_ARM_NEON_V8; |
| 216 | |
| 217 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 218 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 219 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 220 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 221 | zero_point++) |
| 222 | { |
| 223 | const uint32_t min_input = fp32_to_bits((float) (std::numeric_limits<int8_t>::max() - zero_point)); |
| 224 | const uint32_t max_input = UINT32_C(0x7F800000); |
| 225 | for (uint32_t n = min_input; n < max_input; n += kBlockSize) { |
| 226 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 227 | inputs[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input)); |
| 228 | } |
| 229 | xnn_math_f32_qs8_cvt__neonv8(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 230 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 231 | const int32_t reference_output = std::numeric_limits<int8_t>::max(); |
| 232 | ASSERT_EQ(reference_output, int32_t(outputs[i])) |
| 233 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 234 | << ", reference = " << std::dec << reference_output |
| 235 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 236 | << ", zero point = " << std::dec << zero_point; |
| 237 | } |
| 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | TEST(CVT__NEONV8, negative_saturation) { |
| 243 | TEST_REQUIRES_ARM_NEON_V8; |
| 244 | |
| 245 | std::vector<float, AlignedAllocator<float, 64>> inputs(kBlockSize); |
| 246 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> outputs(kBlockSize); |
| 247 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 248 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 249 | zero_point++) |
| 250 | { |
| 251 | const uint32_t min_input = fp32_to_bits((float) (zero_point - std::numeric_limits<int8_t>::min())); |
| 252 | const uint32_t max_input = UINT32_C(0x7F800000); |
| 253 | for (uint32_t n = min_input; n < max_input; n += kBlockSize) { |
| 254 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 255 | inputs[i] = fp32_from_bits(UINT32_C(0x80000000) | std::min<uint32_t>(n + i, max_input)); |
| 256 | } |
| 257 | xnn_math_f32_qs8_cvt__neonv8(kBlockSize * sizeof(int8_t), inputs.data(), outputs.data(), int8_t(zero_point)); |
| 258 | for (uint32_t i = 0; i < kBlockSize; i++) { |
| 259 | const int32_t reference_output = std::numeric_limits<int8_t>::min(); |
| 260 | ASSERT_EQ(reference_output, int32_t(outputs[i])) |
| 261 | << "input = 0x" << std::hex << std::setw(8) << std::setfill('0') << fp32_to_bits(inputs[i]) |
| 262 | << ", reference = " << std::dec << reference_output |
| 263 | << ", optimized = " << std::dec << int32_t(outputs[i]) |
| 264 | << ", zero point = " << std::dec << zero_point; |
| 265 | } |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |