XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright 2019 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 | #pragma once |
| 7 | |
| 8 | #include <gtest/gtest.h> |
| 9 | |
| 10 | #include <algorithm> |
| 11 | #include <cassert> |
| 12 | #include <cmath> |
| 13 | #include <cstddef> |
| 14 | #include <cstdlib> |
| 15 | #include <functional> |
| 16 | #include <random> |
| 17 | #include <vector> |
| 18 | |
| 19 | #include <xnnpack.h> |
| 20 | #include <xnnpack/AlignedAllocator.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 21 | #include <xnnpack/params-init.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 23 | |
| 24 | |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 25 | static inline bool is_fp16_zero(uint16_t x) { |
| 26 | const uint32_t ext_x = x; |
| 27 | const uint32_t two_x = ext_x + ext_x; |
| 28 | return (uint16_t) two_x == 0; |
| 29 | } |
| 30 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 31 | class SpMMMicrokernelTester { |
| 32 | public: |
| 33 | enum class Variant { |
| 34 | Native, |
| 35 | Scalar, |
| 36 | }; |
| 37 | |
| 38 | inline SpMMMicrokernelTester& mr(size_t mr) { |
| 39 | this->mr_ = mr; |
| 40 | return *this; |
| 41 | } |
| 42 | |
| 43 | inline size_t mr() const { |
| 44 | return this->mr_; |
| 45 | } |
| 46 | |
| 47 | inline SpMMMicrokernelTester& nr(size_t nr) { |
| 48 | this->nr_ = nr; |
| 49 | return *this; |
| 50 | } |
| 51 | |
| 52 | inline size_t nr() const { |
| 53 | return this->nr_; |
| 54 | } |
| 55 | |
| 56 | inline SpMMMicrokernelTester& m(size_t m) { |
| 57 | this->m_ = m; |
| 58 | return *this; |
| 59 | } |
| 60 | |
| 61 | inline size_t m() const { |
| 62 | return this->m_; |
| 63 | } |
| 64 | |
| 65 | inline SpMMMicrokernelTester& n(size_t n) { |
| 66 | this->n_ = n; |
| 67 | return *this; |
| 68 | } |
| 69 | |
| 70 | inline size_t n() const { |
| 71 | return this->n_; |
| 72 | } |
| 73 | |
| 74 | inline SpMMMicrokernelTester& k(size_t k) { |
| 75 | this->k_ = k; |
| 76 | return *this; |
| 77 | } |
| 78 | |
| 79 | inline size_t k() const { |
| 80 | return this->k_; |
| 81 | } |
| 82 | |
| 83 | inline SpMMMicrokernelTester& sparsity(float sparsity) { |
| 84 | this->sparsity_ = sparsity; |
| 85 | return *this; |
| 86 | } |
| 87 | |
| 88 | inline float sparsity() const { |
| 89 | return this->sparsity_; |
| 90 | } |
| 91 | |
| 92 | inline SpMMMicrokernelTester& qmin(uint8_t qmin) { |
| 93 | this->qmin_ = qmin; |
| 94 | return *this; |
| 95 | } |
| 96 | |
| 97 | inline uint8_t qmin() const { |
| 98 | return this->qmin_; |
| 99 | } |
| 100 | |
| 101 | inline SpMMMicrokernelTester& qmax(uint8_t qmax) { |
| 102 | this->qmax_ = qmax; |
| 103 | return *this; |
| 104 | } |
| 105 | |
| 106 | inline uint8_t qmax() const { |
| 107 | return this->qmax_; |
| 108 | } |
| 109 | |
| 110 | inline SpMMMicrokernelTester& iterations(size_t iterations) { |
| 111 | this->iterations_ = iterations; |
| 112 | return *this; |
| 113 | } |
| 114 | |
| 115 | inline size_t iterations() const { |
| 116 | return this->iterations_; |
| 117 | } |
| 118 | |
| 119 | void Test(xnn_f32_spmm_ukernel_function spmm, Variant variant = Variant::Native) const { |
| 120 | ASSERT_GE(m(), 1); |
| 121 | ASSERT_GE(n(), 1); |
| 122 | ASSERT_GE(k(), 1); |
| 123 | |
| 124 | std::random_device random_device; |
| 125 | auto rng = std::mt19937(random_device()); |
| 126 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 127 | auto prng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 128 | |
| 129 | std::vector<float, AlignedAllocator<float, 64>> a(k() * m()); |
| 130 | // Think of b as (n/nr + n % nr) x k, expansion happens later. |
| 131 | const size_t ncols = n() / nr() + n() % nr(); |
| 132 | std::vector<float> b(ncols * k()); |
| 133 | std::vector<float> bias(n()); |
| 134 | // Number of non-zero weights per N (output channel). |
| 135 | std::vector<uint32_t> nmap(n()); |
| 136 | // Mapping from index of non-zero weight to increment of K (input channel) following this index. |
| 137 | std::vector<int32_t> dmap(n() * k()); |
| 138 | std::vector<float> w(n() * k() + n()); |
| 139 | std::vector<float> c(n() * m()); |
| 140 | std::vector<float> c_ref(n() * m()); |
| 141 | |
| 142 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 143 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 144 | std::generate(b.begin(), b.end(), std::ref(f32rng)); |
| 145 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 146 | std::fill(c.begin(), c.end(), nanf("")); |
| 147 | std::fill(c_ref.begin(), c_ref.end(), 0.0f); |
| 148 | std::fill(nmap.begin(), nmap.end(), 0); |
| 149 | std::fill(dmap.begin(), dmap.end(), 0); |
| 150 | std::fill(w.begin(), w.end(), 0.0f); |
| 151 | |
| 152 | for (float& b_value : b) { |
| 153 | if (prng() <= sparsity()) { |
| 154 | b_value = 0.0f; |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | uint32_t nnz = 0; |
| 159 | uint32_t wcnt = 0; |
| 160 | size_t last_kk = 0; |
| 161 | bool first_nzz = true; |
| 162 | size_t first_kk = 0; |
| 163 | for (size_t nn = 0; nn < n() / nr(); nn++) { |
| 164 | for (size_t i = 0; i < nr(); ++i) |
| 165 | w[wcnt++] = bias[nr() * nn + i]; |
| 166 | for (size_t kk = 0; kk < k(); kk++) { |
| 167 | if (b[nn * k() + kk] != 0.0f) { |
| 168 | // Every non-zero actually corresponds to nr adjacent non-zeros. |
| 169 | for (size_t i = 0; i < nr(); ++i) |
| 170 | w[wcnt++] = b[nn * k() + kk] + static_cast<float>(i); |
| 171 | // Skip the very first non-zero weight as we record only the difference. |
| 172 | if (first_nzz) { |
| 173 | first_kk = kk; |
| 174 | } else { |
| 175 | const int32_t increment = int32_t(kk - last_kk) * int32_t(m() * sizeof(float)); |
| 176 | dmap[nnz++] = increment; |
| 177 | } |
| 178 | last_kk = kk; |
| 179 | first_nzz = false; |
| 180 | nmap[nn] += 1; |
| 181 | } |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | // now we've constructed the matrix for the blocked part and switch to the |
| 186 | // leftovers, which we do as nr=1 always. |
| 187 | for (size_t nn = n() / nr(); nn < ncols; nn++) { |
| 188 | w[wcnt++] = bias[(n() / nr()) * nr() + (nn - n() / nr())]; |
| 189 | for (size_t kk = 0; kk < k(); kk++) { |
| 190 | if (b[nn * k() + kk] != 0.0f) { |
| 191 | // Every non-zero actually corresponds to nr adjacent non-zeros. |
| 192 | w[wcnt++] = b[nn * k() + kk]; |
| 193 | // Skip the very first non-zero weight as we record only the difference. |
| 194 | if (first_nzz) { |
| 195 | first_kk = kk; |
| 196 | } else { |
| 197 | const int32_t increment = int32_t(kk - last_kk) * int32_t(m() * sizeof(float)); |
| 198 | dmap[nnz++] = increment; |
| 199 | } |
| 200 | last_kk = kk; |
| 201 | first_nzz = false; |
| 202 | nmap[nn] += 1; |
| 203 | } |
| 204 | } |
| 205 | } |
| 206 | // In the end, we must return input pointer to the initial value. |
| 207 | const int64_t increment = int32_t(first_kk - last_kk) * int32_t(m() * sizeof(float)); |
| 208 | dmap[nnz++] = increment; |
| 209 | |
| 210 | // Generate expanded b which will be used in reference calculation. |
| 211 | // Everywhere there is a non-zero in the original we copy it and add an |
| 212 | // adjacent non-zero with incremented weight value. |
| 213 | std::vector<float> b_full(n() * k()); |
| 214 | if (nr() == 1) { |
| 215 | b_full = b; |
| 216 | } |
| 217 | else { |
| 218 | for (size_t nn = 0; nn < n() / nr(); nn++) { |
| 219 | for (size_t kk = 0; kk < k(); kk++) { |
| 220 | if (b[nn * k() + kk] != 0.0f) { |
| 221 | for (size_t i = 0; i < nr(); ++i) |
| 222 | b_full[nr() * nn * k() + i * k() + kk] = b[nn * k() + kk] + static_cast<float>(i); |
| 223 | } |
| 224 | } |
| 225 | } |
| 226 | for (size_t nn = n() / nr(); nn < ncols; nn++) { |
| 227 | for (size_t kk = 0; kk < k(); kk++) { |
| 228 | if (b[nn * k() + kk] != 0.0f) { |
| 229 | b_full[nr() * (n() / nr()) * k() + (nn - n() / nr()) * k() + kk] = b[nn * k() + kk]; |
| 230 | } |
| 231 | } |
| 232 | } |
| 233 | } |
| 234 | |
| 235 | for (size_t oc = 0; oc < n(); oc++) { |
| 236 | for (size_t pxb = 0; pxb < m(); pxb++) { |
| 237 | c_ref[oc * m() + pxb] = bias[oc]; |
| 238 | for (size_t ic = 0; ic < k(); ic++) { |
| 239 | c_ref[oc * m() + pxb] += a[ic * m() + pxb] * b_full[oc * k() + ic]; |
| 240 | } |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | // Micro-kernel can access one element beyond w and dmap for software pipelining. |
| 245 | w.resize(wcnt + 1); |
| 246 | dmap.resize(nnz + 1); |
| 247 | |
| 248 | // Compute clamping parameters. |
| 249 | const float accumulated_min = *std::min_element(c_ref.cbegin(), c_ref.cend()); |
| 250 | const float accumulated_max = *std::max_element(c_ref.cbegin(), c_ref.cend()); |
| 251 | const float c_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 252 | const float c_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 253 | |
| 254 | // Clamp reference results. |
| 255 | for (float& c_value : c_ref) { |
| 256 | c_value = std::min(std::max(c_value, c_min), c_max); |
| 257 | } |
| 258 | |
| 259 | // Prepare output parameters. |
| 260 | xnn_f32_output_params output_params = { }; |
| 261 | switch (variant) { |
| 262 | case Variant::Native: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 263 | output_params = xnn_init_f32_output_params(c_min, c_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 264 | break; |
| 265 | case Variant::Scalar: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 266 | output_params = xnn_init_scalar_f32_output_params(c_min, c_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 267 | break; |
| 268 | } |
| 269 | |
| 270 | spmm(m(), n(), |
| 271 | a.data() + first_kk * m(), w.data(), dmap.data(), nmap.data(), c.data(), |
| 272 | &output_params); |
| 273 | |
| 274 | // Validate micro-kernel outputs. |
| 275 | for (size_t pxb = 0; pxb < n(); pxb++) { |
| 276 | for (size_t oc = 0; oc < m(); oc++) { |
| 277 | ASSERT_NEAR( |
| 278 | c[pxb * m() + oc], |
| 279 | c_ref[pxb * m() + oc], |
| 280 | std::abs(c_ref[pxb * m() + oc]) * 1.0e-6f) |
| 281 | << "at " << pxb << ", " << oc |
| 282 | << ": Mr x Nr x Kr = " << mr() << " x " << nr() |
| 283 | << ", M x N x K = " << m() << " x " << n() << " x " << k(); |
| 284 | } |
| 285 | } |
| 286 | } |
| 287 | } |
| 288 | |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 289 | void Test(xnn_f16_spmm_ukernel_function spmm) const { |
| 290 | ASSERT_GE(m(), 1); |
| 291 | ASSERT_GE(n(), 1); |
| 292 | ASSERT_GE(k(), 1); |
| 293 | |
| 294 | std::random_device random_device; |
| 295 | auto rng = std::mt19937(random_device()); |
| 296 | auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 297 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 298 | auto prng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 299 | |
| 300 | std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> a(k() * m()); |
| 301 | // Think of b as (n/nr + n % nr) x k, expansion happens later. |
| 302 | const size_t ncols = n() / nr() + n() % nr(); |
| 303 | std::vector<uint16_t> b(ncols * k()); |
| 304 | std::vector<uint16_t> bias(n()); |
| 305 | // Number of non-zero weights per N (output channel). |
| 306 | std::vector<uint32_t> nmap(n()); |
| 307 | // Mapping from index of non-zero weight to increment of K (input channel) following this index. |
| 308 | std::vector<int32_t> dmap(n() * k()); |
| 309 | std::vector<uint16_t> w(n() * k() + n()); |
| 310 | std::vector<uint16_t> c(n() * m()); |
| 311 | std::vector<float> c_ref(n() * m()); |
| 312 | |
| 313 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 314 | std::generate(a.begin(), a.end(), std::ref(f16rng)); |
| 315 | std::generate(b.begin(), b.end(), std::ref(f16rng)); |
| 316 | std::generate(bias.begin(), bias.end(), std::ref(f16rng)); |
| 317 | std::fill(c.begin(), c.end(), 0xC000); |
| 318 | std::fill(c_ref.begin(), c_ref.end(), 0.0f); |
| 319 | std::fill(nmap.begin(), nmap.end(), 0); |
| 320 | std::fill(dmap.begin(), dmap.end(), 0); |
| 321 | std::fill(w.begin(), w.end(), 0); |
| 322 | |
| 323 | for (uint16_t& b_value : b) { |
| 324 | if (prng() <= sparsity()) { |
| 325 | b_value = 0; |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | uint32_t nnz = 0; |
| 330 | uint32_t wcnt = 0; |
| 331 | size_t last_kk = 0; |
| 332 | bool first_nzz = true; |
| 333 | size_t first_kk = 0; |
| 334 | for (size_t nn = 0; nn < n() / nr(); nn++) { |
| 335 | for (size_t i = 0; i < nr(); ++i) |
| 336 | w[wcnt++] = bias[nr() * nn + i]; |
| 337 | for (size_t kk = 0; kk < k(); kk++) { |
| 338 | if (!is_fp16_zero(b[nn * k() + kk])) { |
| 339 | // Every non-zero actually corresponds to nr adjacent non-zeros. |
| 340 | for (size_t i = 0; i < nr(); ++i) |
| 341 | w[wcnt++] = fp16_ieee_from_fp32_value(fp16_ieee_to_fp32_value(b[nn * k() + kk]) + static_cast<float>(i)); |
| 342 | // Skip the very first non-zero weight as we record only the difference. |
| 343 | if (first_nzz) { |
| 344 | first_kk = kk; |
| 345 | } else { |
| 346 | const int32_t increment = int32_t(kk - last_kk) * int32_t(m() * sizeof(uint16_t)); |
| 347 | dmap[nnz++] = increment; |
| 348 | } |
| 349 | last_kk = kk; |
| 350 | first_nzz = false; |
| 351 | nmap[nn] += 1; |
| 352 | } |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | // now we've constructed the matrix for the blocked part and switch to the |
| 357 | // leftovers, which we do as nr=1 always. |
| 358 | for (size_t nn = n() / nr(); nn < ncols; nn++) { |
| 359 | w[wcnt++] = bias[(n() / nr()) * nr() + (nn - n() / nr())]; |
| 360 | for (size_t kk = 0; kk < k(); kk++) { |
| 361 | if (!is_fp16_zero(b[nn * k() + kk])) { |
| 362 | // Every non-zero actually corresponds to nr adjacent non-zeros. |
| 363 | w[wcnt++] = b[nn * k() + kk]; |
| 364 | // Skip the very first non-zero weight as we record only the difference. |
| 365 | if (first_nzz) { |
| 366 | first_kk = kk; |
| 367 | } else { |
| 368 | const int32_t increment = int32_t(kk - last_kk) * int32_t(m() * sizeof(uint16_t)); |
| 369 | dmap[nnz++] = increment; |
| 370 | } |
| 371 | last_kk = kk; |
| 372 | first_nzz = false; |
| 373 | nmap[nn] += 1; |
| 374 | } |
| 375 | } |
| 376 | } |
| 377 | // In the end, we must return input pointer to the initial value. |
| 378 | const int64_t increment = int32_t(first_kk - last_kk) * int32_t(m() * sizeof(uint16_t)); |
| 379 | dmap[nnz++] = increment; |
| 380 | |
| 381 | // Generate expanded b which will be used in reference calculation. |
| 382 | // Everywhere there is a non-zero in the original we copy it and add an |
| 383 | // adjacent non-zero with incremented weight value. |
| 384 | std::vector<uint16_t> b_full(n() * k()); |
| 385 | if (nr() == 1) { |
| 386 | b_full = b; |
| 387 | } |
| 388 | else { |
| 389 | for (size_t nn = 0; nn < n() / nr(); nn++) { |
| 390 | for (size_t kk = 0; kk < k(); kk++) { |
| 391 | if (b[nn * k() + kk] != 0.0f) { |
| 392 | for (size_t i = 0; i < nr(); ++i) |
| 393 | b_full[nr() * nn * k() + i * k() + kk] = fp16_ieee_from_fp32_value( |
| 394 | fp16_ieee_to_fp32_value(b[nn * k() + kk]) + static_cast<float>(i)); |
| 395 | } |
| 396 | } |
| 397 | } |
| 398 | for (size_t nn = n() / nr(); nn < ncols; nn++) { |
| 399 | for (size_t kk = 0; kk < k(); kk++) { |
| 400 | if (b[nn * k() + kk] != 0.0f) { |
| 401 | b_full[nr() * (n() / nr()) * k() + (nn - n() / nr()) * k() + kk] = b[nn * k() + kk]; |
| 402 | } |
| 403 | } |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | for (size_t oc = 0; oc < n(); oc++) { |
| 408 | for (size_t pxb = 0; pxb < m(); pxb++) { |
| 409 | c_ref[oc * m() + pxb] = fp16_ieee_to_fp32_value(bias[oc]); |
| 410 | for (size_t ic = 0; ic < k(); ic++) { |
| 411 | c_ref[oc * m() + pxb] += fp16_ieee_to_fp32_value(a[ic * m() + pxb]) * fp16_ieee_to_fp32_value(b_full[oc * k() + ic]); |
| 412 | } |
| 413 | } |
| 414 | } |
| 415 | |
| 416 | // Micro-kernel can access one element beyond w and dmap for software pipelining. |
| 417 | w.resize(wcnt + 1); |
| 418 | dmap.resize(nnz + 1); |
| 419 | |
| 420 | // Compute clamping parameters. |
| 421 | const float accumulated_min = *std::min_element(c_ref.cbegin(), c_ref.cend()); |
| 422 | const float accumulated_max = *std::max_element(c_ref.cbegin(), c_ref.cend()); |
| 423 | const float c_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 424 | const float c_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 425 | |
| 426 | // Clamp reference results. |
| 427 | for (float& c_value : c_ref) { |
| 428 | c_value = std::min(std::max(c_value, c_min), c_max); |
| 429 | } |
| 430 | |
| 431 | // Prepare output parameters. |
| 432 | xnn_f16_output_params output_params; |
| 433 | output_params.scale = UINT16_C(0x3C00) /* 1.0 */; |
| 434 | output_params.max = fp16_ieee_from_fp32_value(c_max); |
| 435 | output_params.min = fp16_ieee_from_fp32_value(c_min); |
| 436 | |
| 437 | spmm(m(), n(), |
| 438 | a.data() + first_kk * m(), w.data(), dmap.data(), nmap.data(), c.data(), |
| 439 | &output_params); |
| 440 | |
| 441 | // Validate micro-kernel outputs. |
| 442 | for (size_t pxb = 0; pxb < n(); pxb++) { |
| 443 | for (size_t oc = 0; oc < m(); oc++) { |
| 444 | ASSERT_NEAR( |
| 445 | fp16_ieee_to_fp32_value(c[pxb * m() + oc]), |
| 446 | c_ref[pxb * m() + oc], |
| 447 | std::abs(c_ref[pxb * m() + oc]) * 1.0e-2f) |
| 448 | << "at " << pxb << ", " << oc |
| 449 | << ": Mr x Nr x Kr = " << mr() << " x " << nr() |
| 450 | << ", M x N x K = " << m() << " x " << n() << " x " << k(); |
| 451 | } |
| 452 | } |
| 453 | } |
| 454 | } |
| 455 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 456 | private: |
| 457 | size_t mr_{1}; |
| 458 | size_t nr_{1}; |
| 459 | size_t m_{1}; |
| 460 | size_t n_{1}; |
| 461 | size_t k_{1}; |
| 462 | float sparsity_{0.5f}; |
| 463 | uint8_t qmin_{0}; |
| 464 | uint8_t qmax_{255}; |
| 465 | size_t iterations_{1}; |
| 466 | }; |