| 1 | // Adapted from |
| 2 | // https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu |
| 3 | |
| 4 | // clang-format off |
| 5 | |
| 6 | #include "common.h" |
| 7 | #include "gemm.h" |
| 8 | #include "moe.h" |
| 9 | |
| 10 | namespace { |
| 11 | |
| 12 | template <typename scalar_t, int BLOCK_N> |
| 13 | inline void silu_and_mul( |
| 14 | scalar_t* __restrict__ C, |
| 15 | const int32_t* __restrict__ C0, // x: x0, x1 |
| 16 | const int32_t* __restrict__ C1, // y: y0, y1 |
| 17 | const float* __restrict__ As, |
| 18 | const float* __restrict__ Bs0, |
| 19 | const float* __restrict__ Bs1, |
| 20 | const int32_t* __restrict__ Bcomp0, |
| 21 | const int32_t* __restrict__ Bcomp1, |
| 22 | int64_t m_size, |
| 23 | int64_t N) { |
| 24 | #if defined(CPU_CAPABILITY_AVX512) |
| 25 | constexpr int COLS = BLOCK_N / 16; |
| 26 | static_assert(COLS % 2 == 0); |
| 27 | |
| 28 | __m512 vc0[COLS]; |
| 29 | __m512 vc1[COLS]; |
| 30 | __m512i vcomp0[COLS]; |
| 31 | __m512i vcomp1[COLS]; |
| 32 | __m512 vas; |
| 33 | __m512 vbs0[COLS]; |
| 34 | __m512 vbs1[COLS]; |
| 35 | |
| 36 | auto load_scale_and_comp = [&](auto col) { |
| 37 | vcomp0[col] = _mm512_loadu_si512(Bcomp0 + col * 16); |
| 38 | vcomp1[col] = _mm512_loadu_si512(Bcomp1 + col * 16); |
| 39 | vbs0[col] = _mm512_loadu_ps(Bs0 + col * 16); |
| 40 | vbs1[col] = _mm512_loadu_ps(Bs1 + col * 16); |
| 41 | }; |
| 42 | Unroll<COLS>{}(load_scale_and_comp); |
| 43 | |
| 44 | auto scalec = [&](auto col, int64_t m) { |
| 45 | // update As |
| 46 | vas = _mm512_set1_ps(As[m]); |
| 47 | // C = As * (C - Bcomp) * Bs |
| 48 | __m512i vc32_0 = _mm512_loadu_si512(C0 + m * BLOCK_N + col * 16); |
| 49 | __m512i vc32_1 = _mm512_loadu_si512(C1 + m * BLOCK_N + col * 16); |
| 50 | vc0[col] = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc32_0, vcomp0[col])); |
| 51 | vc1[col] = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc32_1, vcomp1[col])); |
| 52 | vc0[col] = _mm512_mul_ps(_mm512_mul_ps(vc0[col], vas), vbs0[col]); |
| 53 | vc1[col] = _mm512_mul_ps(_mm512_mul_ps(vc1[col], vas), vbs1[col]); |
| 54 | }; |
| 55 | |
| 56 | auto silu_and_mul = [&](auto col) { |
| 57 | __m512 x = vc0[col]; |
| 58 | __m512 y = vc1[col]; |
| 59 | vc0[col] = _mm512_mul_ps(_mm512_rcp14_silu_ps(x), y); |
| 60 | }; |
| 61 | |
| 62 | using bVec = at::vec::Vectorized<scalar_t>; |
| 63 | using fVec = at::vec::Vectorized<float>; |
| 64 | auto storec = [&](auto col, int64_t m) { |
| 65 | if constexpr (col % 2 == 0) { |
| 66 | fVec x0 = fVec(vc0[col + 0]); |
| 67 | fVec x1 = fVec(vc0[col + 1]); |
| 68 | bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1); |
| 69 | out_vec.store(C + m * N + col * 16); |
| 70 | } |
| 71 | }; |
| 72 | |
| 73 | for (int64_t m = 0; m < m_size; ++m) { |
| 74 | Unroll<COLS>{}(scalec, m); |
| 75 | Unroll<COLS>{}(silu_and_mul); |
| 76 | Unroll<COLS>{}(storec, m); |
| 77 | } |
| 78 | #else |
| 79 | TORCH_CHECK(false, "silu_and_mul: scalar path not implemented!"); |
| 80 | #endif |
| 81 | } |
| 82 | |
| 83 | template <int BLOCK_N> |
| 84 | inline void scale_C( |
| 85 | float* __restrict__ C, |
| 86 | const int32_t* __restrict__ Ctmp, |
| 87 | const float* __restrict__ As, |
| 88 | const float* __restrict__ Bs, |
| 89 | const int32_t* __restrict__ Bcomp, |
| 90 | int64_t m_size) { |
| 91 | #if defined(CPU_CAPABILITY_AVX512) |
| 92 | constexpr int COLS = BLOCK_N / 16; |
| 93 | static_assert(COLS % 2 == 0); |
| 94 | |
| 95 | __m512 vc[COLS]; |
| 96 | __m512i vcomp[COLS]; |
| 97 | __m512 vas; |
| 98 | __m512 vbs[COLS]; |
| 99 | |
| 100 | auto load_scale_and_comp = [&](auto col) { |
| 101 | vcomp[col] = _mm512_loadu_si512(Bcomp + col * 16); |
| 102 | vbs[col] = _mm512_loadu_ps(Bs + col * 16); |
| 103 | }; |
| 104 | Unroll<COLS>{}(load_scale_and_comp); |
| 105 | |
| 106 | auto scalec = [&](auto col, int64_t m) { |
| 107 | // update As |
| 108 | vas = _mm512_set1_ps(As[m]); |
| 109 | // C = As * (C - Bcomp) * Bs |
| 110 | __m512i vc32 = _mm512_loadu_si512(Ctmp + m * BLOCK_N + col * 16); |
| 111 | vc[col] = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc32, vcomp[col])); |
| 112 | vc[col] = _mm512_mul_ps(_mm512_mul_ps(vc[col], vas), vbs[col]); |
| 113 | _mm512_storeu_ps(C + m * BLOCK_N + col * 16, vc[col]); |
| 114 | }; |
| 115 | |
| 116 | for (int64_t m = 0; m < m_size; ++m) { |
| 117 | Unroll<COLS>{}(scalec, m); |
| 118 | } |
| 119 | #else |
| 120 | TORCH_CHECK(false, "scale_C: scalar path not implemented!"); |
| 121 | #endif |
| 122 | } |
| 123 | |
| 124 | /// gemm for w13 |
| 125 | template <typename scalar_t, int BLOCK_M, int BLOCK_N> |
| 126 | struct tinygemm_kernel_vnni { |
| 127 | static inline void apply( |
| 128 | const uint8_t* __restrict__ A, |
| 129 | const int8_t* __restrict__ B0, |
| 130 | const int8_t* __restrict__ B1, |
| 131 | scalar_t* __restrict__ C, |
| 132 | const float* __restrict__ As, |
| 133 | const float* __restrict__ Bs0, |
| 134 | const float* __restrict__ Bs1, |
| 135 | const int32_t* __restrict__ Bcomp0, |
| 136 | const int32_t* __restrict__ Bcomp1, |
| 137 | int64_t K, |
| 138 | int64_t lda, |
| 139 | int64_t ldb, |
| 140 | int64_t ldc) { |
| 141 | TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!"); |
| 142 | } |
| 143 | }; |
| 144 | |
| 145 | #if defined(CPU_CAPABILITY_AVX512) |
| 146 | template <int BLOCK_M, int BLOCK_N> |
| 147 | struct tinygemm_kernel_vnni<at::BFloat16, BLOCK_M, BLOCK_N> { |
| 148 | static inline void apply( |
| 149 | const uint8_t* __restrict__ A, |
| 150 | const int8_t* __restrict__ B0, |
| 151 | const int8_t* __restrict__ B1, |
| 152 | at::BFloat16* __restrict__ C, |
| 153 | const float* __restrict__ As, |
| 154 | const float* __restrict__ Bs0, |
| 155 | const float* __restrict__ Bs1, |
| 156 | const int32_t* __restrict__ Bcomp0, |
| 157 | const int32_t* __restrict__ Bcomp1, |
| 158 | int64_t K, |
| 159 | int64_t lda, |
| 160 | int64_t ldb, |
| 161 | int64_t ldc) { |
| 162 | constexpr int ROWS = BLOCK_M; |
| 163 | constexpr int COLS = BLOCK_N / 16; |
| 164 | static_assert(COLS % 2 == 0); |
| 165 | |
| 166 | __m512i va; |
| 167 | __m512i vb0[COLS]; |
| 168 | __m512i vb1[COLS]; |
| 169 | __m512i vc0[ROWS * COLS]; |
| 170 | __m512i vc1[ROWS * COLS]; |
| 171 | __m512i vcomp0[COLS]; |
| 172 | __m512i vcomp1[COLS]; |
| 173 | __m512 vas; |
| 174 | __m512 vbs0[COLS]; |
| 175 | __m512 vbs1[COLS]; |
| 176 | |
| 177 | auto loadc = [&](auto i) { |
| 178 | vc0[i] = _mm512_set1_epi32(0); |
| 179 | vc1[i] = _mm512_set1_epi32(0); |
| 180 | }; |
| 181 | Unroll<ROWS * COLS>{}(loadc); |
| 182 | |
| 183 | const int64_t K4 = K >> 2; |
| 184 | const int64_t lda4 = lda >> 2; |
| 185 | const int64_t ldb4 = ldb; // ldb * 4 >> 2; |
| 186 | const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A); |
| 187 | const int32_t* b0_ptr = reinterpret_cast<const int32_t*>(B0); |
| 188 | const int32_t* b1_ptr = reinterpret_cast<const int32_t*>(B1); |
| 189 | |
| 190 | auto compute = [&](auto i, int64_t k) { |
| 191 | constexpr int row = i / COLS; |
| 192 | constexpr int col = i % COLS; |
| 193 | |
| 194 | if constexpr (col == 0) { |
| 195 | va = _mm512_set1_epi32(a_ptr[row * lda4 + k]); |
| 196 | } |
| 197 | if constexpr (row == 0) { |
| 198 | vb0[col] = _mm512_loadu_si512(b0_ptr + k * ldb4 + col * 16); |
| 199 | vb1[col] = _mm512_loadu_si512(b1_ptr + k * ldb4 + col * 16); |
| 200 | } |
| 201 | vc0[i] = _mm512_dpbusd_epi32(vc0[i], va, vb0[col]); |
| 202 | vc1[i] = _mm512_dpbusd_epi32(vc1[i], va, vb1[col]); |
| 203 | }; |
| 204 | for (int64_t k = 0; k < K4; ++k) { |
| 205 | Unroll<ROWS * COLS>{}(compute, k); |
| 206 | } |
| 207 | |
| 208 | auto scalec = [&](auto i) { |
| 209 | constexpr int row = i / COLS; |
| 210 | constexpr int col = i % COLS; |
| 211 | |
| 212 | // load a scale |
| 213 | if constexpr (col == 0) { |
| 214 | vas = _mm512_set1_ps(As[row]); |
| 215 | } |
| 216 | // load b scale and vcomp |
| 217 | if constexpr (row == 0) { |
| 218 | vbs0[col] = _mm512_loadu_ps(Bs0 + col * 16); |
| 219 | vbs1[col] = _mm512_loadu_ps(Bs1 + col * 16); |
| 220 | vcomp0[col] = _mm512_loadu_si512(Bcomp0 + col * 16); |
| 221 | vcomp1[col] = _mm512_loadu_si512(Bcomp1 + col * 16); |
| 222 | } |
| 223 | __m512 c0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc0[i], vcomp0[col])); |
| 224 | __m512 c1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc1[i], vcomp1[col])); |
| 225 | vc0[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c0, vas), vbs0[col])); |
| 226 | vc1[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c1, vas), vbs1[col])); |
| 227 | }; |
| 228 | Unroll<ROWS * COLS>{}(scalec); |
| 229 | |
| 230 | auto storec = [&](auto i) { |
| 231 | constexpr int row = i / COLS; |
| 232 | constexpr int col = i % COLS; |
| 233 | // for COLS = 2, 4 use 512bit store |
| 234 | if constexpr (col % 2 == 0) { |
| 235 | __m512 x0 = _mm512_castsi512_ps(vc0[row * COLS + col + 0]); |
| 236 | __m512 x1 = _mm512_castsi512_ps(vc0[row * COLS + col + 1]); |
| 237 | __m512 y0 = _mm512_castsi512_ps(vc1[row * COLS + col + 0]); |
| 238 | __m512 y1 = _mm512_castsi512_ps(vc1[row * COLS + col + 1]); |
| 239 | x0 = _mm512_mul_ps(_mm512_rcp14_silu_ps(x0), y0); |
| 240 | x1 = _mm512_mul_ps(_mm512_rcp14_silu_ps(x1), y1); |
| 241 | |
| 242 | _mm512_storeu_si512( |
| 243 | reinterpret_cast<__m512i*>((C + row * ldc + col * 16)), |
| 244 | (__m512i)(_mm512_cvtne2ps_pbh(__m512(x1), __m512(x0)))); |
| 245 | } |
| 246 | }; |
| 247 | Unroll<ROWS * COLS>{}(storec); |
| 248 | } |
| 249 | }; |
| 250 | #endif |
| 251 | |
| 252 | #define LAUNCH_TINYGEMM_KERNEL_VNNI(MB_SIZE, NB_SIZE) \ |
| 253 | tinygemm_kernel_vnni<scalar_t, MB_SIZE, NB_SIZE>::apply( \ |
| 254 | A + mb_start * lda, \ |
| 255 | B0 + nb_start * 4, \ |
| 256 | B1 + nb_start * 4, \ |
| 257 | C + mb_start * ldc + nb_start, \ |
| 258 | As + mb_start, \ |
| 259 | Bs0 + nb_start, \ |
| 260 | Bs1 + nb_start, \ |
| 261 | Bcomp0 + nb_start, \ |
| 262 | Bcomp1 + nb_start, \ |
| 263 | K, \ |
| 264 | lda, \ |
| 265 | ldb, \ |
| 266 | ldc); |
| 267 | |
| 268 | template <typename scalar_t> |
| 269 | void tinygemm_kernel( |
| 270 | const uint8_t* __restrict__ A, |
| 271 | const int8_t* __restrict__ B0, |
| 272 | const int8_t* __restrict__ B1, |
| 273 | scalar_t* __restrict__ C, |
| 274 | const float* __restrict__ As, |
| 275 | const float* __restrict__ Bs0, |
| 276 | const float* __restrict__ Bs1, |
| 277 | int64_t M, |
| 278 | int64_t N, |
| 279 | int64_t K, |
| 280 | int64_t lda, |
| 281 | int64_t ldb, |
| 282 | int64_t ldc) { |
| 283 | const int32_t* Bcomp0 = reinterpret_cast<const int32_t*>(B0 + block_size_n() * K); |
| 284 | const int32_t* Bcomp1 = reinterpret_cast<const int32_t*>(B1 + block_size_n() * K); |
| 285 | |
| 286 | // pattern: 1-(2+2)-(8+8) |
| 287 | constexpr int64_t BLOCK_M = 4; |
| 288 | constexpr int64_t BLOCK_N = 32; |
| 289 | const int64_t MB = div_up(M, BLOCK_M); |
| 290 | const int64_t NB = div_up(N, BLOCK_N); |
| 291 | for (int mb = 0; mb < MB; ++mb) { |
| 292 | int64_t mb_start = mb * BLOCK_M; |
| 293 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 294 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 295 | int64_t nb_start = nb * BLOCK_N; |
| 296 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 297 | |
| 298 | switch (mb_size << 4 | nb_size >> 4) { |
| 299 | case 0x12: |
| 300 | LAUNCH_TINYGEMM_KERNEL_VNNI(1, 32); |
| 301 | break; |
| 302 | case 0x22: |
| 303 | LAUNCH_TINYGEMM_KERNEL_VNNI(2, 32); |
| 304 | break; |
| 305 | case 0x32: |
| 306 | LAUNCH_TINYGEMM_KERNEL_VNNI(3, 32); |
| 307 | break; |
| 308 | case 0x42: |
| 309 | LAUNCH_TINYGEMM_KERNEL_VNNI(4, 32); |
| 310 | break; |
| 311 | default: |
| 312 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 313 | } |
| 314 | } |
| 315 | } |
| 316 | } |
| 317 | |
| 318 | /// gemm for w2 |
| 319 | template <typename scalar_t, int BLOCK_M, int BLOCK_N> |
| 320 | struct tinygemm_kernel_vnni2 { |
| 321 | static inline void apply( |
| 322 | const uint8_t* __restrict__ A, |
| 323 | const int8_t* __restrict__ B, |
| 324 | float* __restrict__ C, |
| 325 | const float* __restrict__ As, |
| 326 | const float* __restrict__ Bs, |
| 327 | const int32_t* __restrict__ Bcomp, |
| 328 | int64_t K, |
| 329 | int64_t lda, |
| 330 | int64_t ldb, |
| 331 | int64_t ldc) { |
| 332 | TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!"); |
| 333 | } |
| 334 | }; |
| 335 | |
| 336 | #if defined(CPU_CAPABILITY_AVX512) |
| 337 | template <int BLOCK_M, int BLOCK_N> |
| 338 | struct tinygemm_kernel_vnni2<at::BFloat16, BLOCK_M, BLOCK_N> { |
| 339 | static inline void apply( |
| 340 | const uint8_t* __restrict__ A, |
| 341 | const int8_t* __restrict__ B, |
| 342 | float* __restrict__ C, |
| 343 | const float* __restrict__ As, |
| 344 | const float* __restrict__ Bs, |
| 345 | const int32_t* __restrict__ Bcomp, |
| 346 | int64_t K, |
| 347 | int64_t lda, |
| 348 | int64_t ldb, |
| 349 | int64_t ldc) { |
| 350 | constexpr int ROWS = BLOCK_M; |
| 351 | constexpr int COLS = BLOCK_N / 16; |
| 352 | static_assert(COLS % 2 == 0); |
| 353 | |
| 354 | __m512i va; |
| 355 | __m512i vb[COLS]; |
| 356 | __m512i vc[ROWS * COLS]; |
| 357 | __m512i vcomp[COLS]; |
| 358 | __m512 vas; |
| 359 | __m512 vbs[COLS]; |
| 360 | |
| 361 | auto loadc = [&](auto i) { vc[i] = _mm512_set1_epi32(0); }; |
| 362 | Unroll<ROWS * COLS>{}(loadc); |
| 363 | |
| 364 | const int64_t K4 = K >> 2; |
| 365 | const int64_t lda4 = lda >> 2; |
| 366 | const int64_t ldb4 = ldb; // ldb * 4 >> 2; |
| 367 | const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A); |
| 368 | const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B); |
| 369 | |
| 370 | auto compute = [&](auto i, int64_t k) { |
| 371 | constexpr int row = i / COLS; |
| 372 | constexpr int col = i % COLS; |
| 373 | |
| 374 | if constexpr (col == 0) { |
| 375 | va = _mm512_set1_epi32(a_ptr[row * lda4 + k]); |
| 376 | } |
| 377 | if constexpr (row == 0) { |
| 378 | vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16); |
| 379 | } |
| 380 | vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]); |
| 381 | }; |
| 382 | for (int64_t k = 0; k < K4; ++k) { |
| 383 | Unroll<ROWS * COLS>{}(compute, k); |
| 384 | } |
| 385 | |
| 386 | auto storec = [&](auto i) { |
| 387 | constexpr int row = i / COLS; |
| 388 | constexpr int col = i % COLS; |
| 389 | |
| 390 | // load a scale |
| 391 | if constexpr (col == 0) { |
| 392 | vas = _mm512_set1_ps(As[row]); |
| 393 | } |
| 394 | // load b scale and vcomp per 2 vectors |
| 395 | // also load bias if any |
| 396 | if constexpr (row == 0) { |
| 397 | if constexpr (col % 2 == 0) { |
| 398 | vbs[col + 0] = _mm512_loadu_ps(Bs + col * 16); |
| 399 | vbs[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16); |
| 400 | vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16); |
| 401 | vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16); |
| 402 | } |
| 403 | } |
| 404 | __m512 x = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[i], vcomp[col])); |
| 405 | x = _mm512_mul_ps(_mm512_mul_ps(x, vas), vbs[col]); |
| 406 | _mm512_storeu_ps(reinterpret_cast<__m512*>(C + row * ldc + col * 16), x); |
| 407 | }; |
| 408 | Unroll<ROWS * COLS>{}(storec); |
| 409 | } |
| 410 | }; |
| 411 | #endif |
| 412 | |
| 413 | #define LAUNCH_TINYGEMM_KERNEL_VNNI2(MB_SIZE, NB_SIZE) \ |
| 414 | tinygemm_kernel_vnni2<scalar_t, MB_SIZE, NB_SIZE>::apply( \ |
| 415 | A + mb_start * lda, \ |
| 416 | B + nb_start * 4, \ |
| 417 | C + mb_start * ldc + nb_start, \ |
| 418 | As + mb_start, \ |
| 419 | Bs + nb_start, \ |
| 420 | Bcomp + nb_start, \ |
| 421 | K, \ |
| 422 | lda, \ |
| 423 | ldb, \ |
| 424 | ldc); |
| 425 | |
| 426 | template <typename scalar_t> |
| 427 | void tinygemm_kernel( |
| 428 | const uint8_t* __restrict__ A, |
| 429 | const int8_t* __restrict__ B, |
| 430 | float* __restrict__ C, |
| 431 | const float* __restrict__ As, |
| 432 | const float* __restrict__ Bs, |
| 433 | int64_t M, |
| 434 | int64_t N, |
| 435 | int64_t K, |
| 436 | int64_t lda, |
| 437 | int64_t ldb, |
| 438 | int64_t ldc) { |
| 439 | // B compensation |
| 440 | const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K); |
| 441 | |
| 442 | // pattern: 1-4-16 |
| 443 | constexpr int64_t BLOCK_M = 4; |
| 444 | constexpr int64_t BLOCK_N = 64; |
| 445 | const int64_t MB = div_up(M, BLOCK_M); |
| 446 | const int64_t NB = div_up(N, BLOCK_N); |
| 447 | for (int64_t mb = 0; mb < MB; ++mb) { |
| 448 | int64_t mb_start = mb * BLOCK_M; |
| 449 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 450 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 451 | int64_t nb_start = nb * BLOCK_N; |
| 452 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 453 | |
| 454 | switch (mb_size << 4 | nb_size >> 4) { |
| 455 | case 0x12: |
| 456 | LAUNCH_TINYGEMM_KERNEL_VNNI2(1, 32); |
| 457 | break; |
| 458 | case 0x22: |
| 459 | LAUNCH_TINYGEMM_KERNEL_VNNI2(2, 32); |
| 460 | break; |
| 461 | case 0x32: |
| 462 | LAUNCH_TINYGEMM_KERNEL_VNNI2(3, 32); |
| 463 | break; |
| 464 | case 0x42: |
| 465 | LAUNCH_TINYGEMM_KERNEL_VNNI2(4, 32); |
| 466 | break; |
| 467 | default: |
| 468 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 469 | } |
| 470 | } |
| 471 | } |
| 472 | } |
| 473 | |
| 474 | } // anonymous namespace |
| 475 | |
| 476 | template <typename scalar_t> |
| 477 | void fused_experts_int8_kernel_impl( |
| 478 | scalar_t* __restrict__ output, |
| 479 | scalar_t* __restrict__ ic1, |
| 480 | scalar_t* __restrict__ ic2, |
| 481 | uint8_t* __restrict__ A_tmp, |
| 482 | float* __restrict__ C_tmp, |
| 483 | uint8_t* __restrict__ Aq_tmp, |
| 484 | float* __restrict__ As_tmp, |
| 485 | const scalar_t* __restrict__ input, |
| 486 | const int8_t* __restrict__ packed_w1, |
| 487 | const int8_t* __restrict__ packed_w2, |
| 488 | const float* __restrict__ w1s, |
| 489 | const float* __restrict__ w2s, |
| 490 | const float* __restrict__ topk_weights, |
| 491 | const int32_t* __restrict__ sorted_ids, |
| 492 | const int32_t* __restrict__ expert_ids, |
| 493 | const int32_t* __restrict__ offsets, |
| 494 | int64_t M, |
| 495 | int64_t N, |
| 496 | int64_t K, |
| 497 | int64_t E, |
| 498 | int64_t topk, |
| 499 | int64_t num_tokens_post_pad) { |
| 500 | // handle 2 tiles per block |
| 501 | constexpr int64_t BLOCK_M = block_size_m(); |
| 502 | constexpr int64_t BLOCK_N = block_size_n(); |
| 503 | |
| 504 | // stage 0: quantize input to uint8, [M, K] |
| 505 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 506 | for (int64_t m = begin; m < end; ++m) { |
| 507 | quantize_row_int8<scalar_t>(Aq_tmp + m * K, As_tmp[m], input + m * K, K); |
| 508 | } |
| 509 | }); |
| 510 | |
| 511 | // stage 1: intermediate_cache1 = silu(hidden_states @ w1) |
| 512 | const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M); |
| 513 | const int64_t NB = div_up(N, BLOCK_N); |
| 514 | |
| 515 | // strides for w1: [E, 2N, K] |
| 516 | TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N); |
| 517 | |
| 518 | // K and N are packed for int8 |
| 519 | const int64_t packed_K = get_row_size<int8_t>(K); |
| 520 | const int64_t packed_N = get_row_size<int8_t>(N); |
| 521 | |
| 522 | const int64_t stride_e = 2 * N * packed_K; |
| 523 | const int64_t stride_n = packed_K; |
| 524 | |
| 525 | int64_t avg_M = std::max(int64_t(1), M * topk / E); |
| 526 | const bool use_brgemm = can_use_brgemm<int8_t>(avg_M); |
| 527 | |
| 528 | // here we only parallel on half of 2N to fuse silu_and_mul with gemm |
| 529 | parallel_2d(MB, NB, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 530 | // get local pointers |
| 531 | int tid = get_thread_num(); |
| 532 | uint8_t* __restrict__ A = A_tmp + tid * BLOCK_M * K; |
| 533 | int32_t* __restrict__ C0 = reinterpret_cast<int32_t*>(C_tmp) + tid * 2 * BLOCK_M * BLOCK_N; |
| 534 | int32_t* __restrict__ C1 = C0 + BLOCK_M * BLOCK_N; |
| 535 | |
| 536 | alignas(64) float As[BLOCK_M]; |
| 537 | |
| 538 | loop_2d<int8_t>(mb0, mb1, nb0, nb1, BLOCK_N * K * 2, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 539 | // nb_upper from top half and nb_lower from bottom half |
| 540 | int64_t nb_upper = nb, nb_lower = nb + NB; |
| 541 | int64_t n_size = std::min(N - nb * BLOCK_N, BLOCK_N); |
| 542 | |
| 543 | // B shape [K, n_size] in vnni format |
| 544 | int32_t expert_id = expert_ids[mb]; |
| 545 | const int8_t* __restrict__ B0 = packed_w1 + expert_id * stride_e + nb_upper * BLOCK_N * stride_n; |
| 546 | const int8_t* __restrict__ B1 = packed_w1 + expert_id * stride_e + nb_lower * BLOCK_N * stride_n; |
| 547 | const float* __restrict__ Bs0 = w1s + expert_id * 2 * N + nb_upper * BLOCK_N; |
| 548 | const float* __restrict__ Bs1 = w1s + expert_id * 2 * N + nb_lower * BLOCK_N; |
| 549 | |
| 550 | int64_t m_size = offsets[mb + 1] - offsets[mb]; |
| 551 | |
| 552 | if (nb_offset == 0) { |
| 553 | // 1.a load A |
| 554 | const int32_t* A_ids = sorted_ids + mb * BLOCK_M; |
| 555 | for (int64_t m = 0; m < m_size; ++m) { |
| 556 | int32_t index = A_ids[m] / topk; |
| 557 | copy_stub(A + m * K, Aq_tmp + index * K, K); |
| 558 | As[m] = As_tmp[index]; |
| 559 | } |
| 560 | } |
| 561 | |
| 562 | if (use_brgemm) { |
| 563 | // 1.b gemm: C0 = A @ B0 |
| 564 | at::native::cpublas::brgemm( |
| 565 | /* M */ m_size, |
| 566 | /* N */ n_size, |
| 567 | /* K */ K, |
| 568 | /* lda */ K, |
| 569 | /* ldb */ n_size, |
| 570 | /* ldc */ BLOCK_N, |
| 571 | /* add_C */ false, |
| 572 | /* A */ A, |
| 573 | /* B */ B0, |
| 574 | /* C */ C0); |
| 575 | |
| 576 | // 1.c gemm: C1 = A @ B1 |
| 577 | at::native::cpublas::brgemm( |
| 578 | /* M */ m_size, |
| 579 | /* N */ n_size, |
| 580 | /* K */ K, |
| 581 | /* lda */ K, |
| 582 | /* ldb */ n_size, |
| 583 | /* ldc */ BLOCK_N, |
| 584 | /* add_C */ false, |
| 585 | /* A */ A, |
| 586 | /* B */ B1, |
| 587 | /* C */ C1); |
| 588 | |
| 589 | const int32_t* Bcomp0 = reinterpret_cast<const int32_t*>(B0 + block_size_n() * K); |
| 590 | const int32_t* Bcomp1 = reinterpret_cast<const int32_t*>(B1 + block_size_n() * K); |
| 591 | |
| 592 | // 1.d silu and mul |
| 593 | const int64_t offset = offsets[mb]; |
| 594 | silu_and_mul<scalar_t, BLOCK_N>( |
| 595 | ic1 + offset * N + nb * BLOCK_N, C0, C1, As, Bs0, Bs1, Bcomp0, Bcomp1, m_size, N); |
| 596 | } else { |
| 597 | // fused 1.bcd: silu_and_mul(A @ B0, A @ B1) |
| 598 | const int64_t offset = offsets[mb]; |
| 599 | tinygemm_kernel( |
| 600 | /* A */ A, |
| 601 | /* B0 */ B0, |
| 602 | /* B1 */ B1, |
| 603 | /* C */ ic1 + offset * N + nb * BLOCK_N, |
| 604 | /* As */ As, |
| 605 | /* Bs0 */ Bs0, |
| 606 | /* Bs1 */ Bs1, |
| 607 | /* M */ m_size, |
| 608 | /* N */ n_size, |
| 609 | /* K */ K, |
| 610 | /* lda */ K, |
| 611 | /* ldb */ n_size, |
| 612 | /* ldc */ N); |
| 613 | } |
| 614 | }); |
| 615 | |
| 616 | if (use_brgemm) { |
| 617 | at::native::cpublas::brgemm_release(); |
| 618 | } |
| 619 | }); |
| 620 | |
| 621 | // stage 1.5: quantize ic1 to uint8, [M * topk, N] |
| 622 | at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) { |
| 623 | for (int64_t m = begin; m < end; ++m) { |
| 624 | quantize_row_int8<scalar_t>(Aq_tmp + m * N, As_tmp[m], ic1 + m * N, N); |
| 625 | } |
| 626 | }); |
| 627 | |
| 628 | // stage 2: intermediate_cache2 = intermediate_cache1 @ w2 |
| 629 | // w2 : [E, K, N] as [E, OC, IC] |
| 630 | const int64_t OC = K; // rename K as OC |
| 631 | const int64_t IC = N; // rename N as IC |
| 632 | const int64_t MB2 = MB; |
| 633 | const int64_t NB2 = div_up(OC, BLOCK_N); |
| 634 | const int64_t stride_e2 = OC * packed_N; |
| 635 | const int64_t stride_oc = packed_N; |
| 636 | |
| 637 | // parallel on [MB2, NB2] |
| 638 | parallel_2d(MB2, NB2, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 639 | // get local pointers |
| 640 | int tid = get_thread_num(); |
| 641 | float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N; |
| 642 | int32_t* __restrict__ C32 = reinterpret_cast<int32_t*>(C + BLOCK_M * BLOCK_N); |
| 643 | |
| 644 | loop_2d<int8_t>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 645 | int64_t m_size = offsets[mb + 1] - offsets[mb]; |
| 646 | int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N); |
| 647 | |
| 648 | // A ptr from ic1 of [M * topk, N] in sorted order |
| 649 | // so as to avoid copy A to tmp buffer again |
| 650 | const uint8_t* __restrict__ A = Aq_tmp + offsets[mb] * N; |
| 651 | const float* __restrict__ As = As_tmp + offsets[mb]; |
| 652 | const int32_t* A_ids = sorted_ids + mb * BLOCK_M; |
| 653 | |
| 654 | // B shape [IC, n_size] in vnni format |
| 655 | int32_t expert_id = expert_ids[mb]; |
| 656 | const int8_t* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc; |
| 657 | const float* __restrict__ Bs = w2s + expert_id * K + nb * BLOCK_N; |
| 658 | |
| 659 | // 2.a gemm: C = A @ B |
| 660 | if (use_brgemm) { |
| 661 | at::native::cpublas::brgemm( |
| 662 | /* M */ m_size, |
| 663 | /* N */ n_size, |
| 664 | /* K */ IC, |
| 665 | /* lda */ IC, |
| 666 | /* ldb */ n_size, |
| 667 | /* ldc */ BLOCK_N, |
| 668 | /* add_C */ false, |
| 669 | /* A */ A, |
| 670 | /* B */ B, |
| 671 | /* C */ C32); |
| 672 | |
| 673 | // apply scales |
| 674 | const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * IC); |
| 675 | scale_C<BLOCK_N>(C, C32, As, Bs, Bcomp, m_size); |
| 676 | } else { |
| 677 | tinygemm_kernel<scalar_t>( |
| 678 | /* A */ A, |
| 679 | /* B */ B, |
| 680 | /* C */ C, |
| 681 | /* As */ As, |
| 682 | /* Bs */ Bs, |
| 683 | /* M */ m_size, |
| 684 | /* N */ n_size, |
| 685 | /* K */ IC, |
| 686 | /* lda */ IC, |
| 687 | /* ldb */ n_size, |
| 688 | /* ldc */ BLOCK_N); |
| 689 | } |
| 690 | |
| 691 | // 2.b copy from C to ic2 in original order |
| 692 | // and also mul topk_weights in float32 |
| 693 | for (int64_t m = 0; m < m_size; ++m) { |
| 694 | int32_t index = A_ids[m]; |
| 695 | float weight = topk_weights[index]; |
| 696 | copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size); |
| 697 | } |
| 698 | }); |
| 699 | |
| 700 | if (use_brgemm) { |
| 701 | at::native::cpublas::brgemm_release(); |
| 702 | } |
| 703 | }); |
| 704 | |
| 705 | // stage 3: out = intermediate_cache2.sum(dim=1) |
| 706 | // from [M, topk, K] to [M, K] |
| 707 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 708 | for (int64_t m = begin; m < end; ++m) { |
| 709 | sum_stub(output + m * K, ic2 + m * topk * K, topk, K); |
| 710 | } |
| 711 | }); |
| 712 | } |
| 713 | |
| 714 | #define INSTANTIATE_MOE_INT8_TEMPLATE(TYPE) \ |
| 715 | template void fused_experts_int8_kernel_impl<TYPE>( \ |
| 716 | TYPE* __restrict__ output, \ |
| 717 | TYPE* __restrict__ ic1, \ |
| 718 | TYPE* __restrict__ ic2, \ |
| 719 | uint8_t* __restrict__ A_tmp, \ |
| 720 | float* __restrict__ C_tmp, \ |
| 721 | uint8_t* __restrict__ Aq_tmp, \ |
| 722 | float* __restrict__ As_tmp, \ |
| 723 | const TYPE* __restrict__ input, \ |
| 724 | const int8_t* __restrict__ packed_w1, \ |
| 725 | const int8_t* __restrict__ packed_w2, \ |
| 726 | const float* __restrict__ w1s, \ |
| 727 | const float* __restrict__ w2s, \ |
| 728 | const float* __restrict__ topk_weights, \ |
| 729 | const int32_t* __restrict__ sorted_ids, \ |
| 730 | const int32_t* __restrict__ expert_ids, \ |
| 731 | const int32_t* __restrict__ offsets, \ |
| 732 | int64_t M, \ |
| 733 | int64_t N, \ |
| 734 | int64_t K, \ |
| 735 | int64_t E, \ |
| 736 | int64_t topk, \ |
| 737 | int64_t num_tokens_post_pad) |
| 738 | |
| 739 | INSTANTIATE_MOE_INT8_TEMPLATE(at::BFloat16); |
| 740 | INSTANTIATE_MOE_INT8_TEMPLATE(at::Half); |
| 741 | |
| 742 | template <typename scalar_t> |
| 743 | void shared_expert_int8_kernel_impl( |
| 744 | scalar_t* __restrict__ output, |
| 745 | scalar_t* __restrict__ ic1, |
| 746 | float* __restrict__ C_tmp, |
| 747 | uint8_t* __restrict__ Aq_tmp, |
| 748 | float* __restrict__ As_tmp, |
| 749 | const scalar_t* __restrict__ input, |
| 750 | const int8_t* __restrict__ packed_w1, |
| 751 | const int8_t* __restrict__ packed_w2, |
| 752 | const float* __restrict__ w1s, |
| 753 | const float* __restrict__ w2s, |
| 754 | const scalar_t* __restrict__ fused_experts_out, |
| 755 | float routed_scaling_factor, |
| 756 | int64_t M, |
| 757 | int64_t N, |
| 758 | int64_t K) { |
| 759 | // handle 2 tiles per block |
| 760 | constexpr int64_t BLOCK_M = block_size_m(); |
| 761 | constexpr int64_t BLOCK_N = block_size_n(); |
| 762 | |
| 763 | // stage 0: quantize input to uint8, [M, K] |
| 764 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 765 | for (int64_t m = begin; m < end; ++m) { |
| 766 | quantize_row_int8<scalar_t>(Aq_tmp + m * K, As_tmp[m], input + m * K, K); |
| 767 | } |
| 768 | }); |
| 769 | |
| 770 | // stage 1: intermediate_cache1 = silu(hidden_states @ w1) |
| 771 | const int64_t MB = div_up(M, BLOCK_M); |
| 772 | const int64_t NB = div_up(N, BLOCK_N); |
| 773 | |
| 774 | TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N); |
| 775 | |
| 776 | // K and N are packed for int8 |
| 777 | const int64_t packed_K = get_row_size<int8_t>(K); |
| 778 | const int64_t packed_N = get_row_size<int8_t>(N); |
| 779 | const int64_t stride_n = packed_K; |
| 780 | |
| 781 | const bool use_brgemm = can_use_brgemm<int8_t>(M); |
| 782 | const bool apply_scaling_factor = fused_experts_out != nullptr; |
| 783 | |
| 784 | // here we only parallel on half of 2N to fuse silu_and_mul with gemm |
| 785 | parallel_2d(MB, NB, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 786 | // get local pointers |
| 787 | int tid = get_thread_num(); |
| 788 | int32_t* __restrict__ C0 = reinterpret_cast<int32_t*>(C_tmp) + tid * 2 * BLOCK_M * BLOCK_N; |
| 789 | int32_t* __restrict__ C1 = C0 + BLOCK_M * BLOCK_N; |
| 790 | |
| 791 | loop_2d<int8_t>(mb0, mb1, nb0, nb1, BLOCK_N * K * 2, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 792 | // nb_upper from top half and nb_lower from bottom half |
| 793 | int64_t nb_upper = nb, nb_lower = nb + NB; |
| 794 | int64_t n_size = std::min(N - nb * BLOCK_N, BLOCK_N); |
| 795 | int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M); |
| 796 | |
| 797 | // A shape [m_size, K] |
| 798 | const uint8_t* A = Aq_tmp + mb * BLOCK_M * K; |
| 799 | const float* As = As_tmp + mb * BLOCK_M; |
| 800 | |
| 801 | // B shape [K, n_size] in vnni format |
| 802 | const int8_t* __restrict__ B0 = packed_w1 + nb_upper * BLOCK_N * stride_n; |
| 803 | const int8_t* __restrict__ B1 = packed_w1 + nb_lower * BLOCK_N * stride_n; |
| 804 | const float* __restrict__ Bs0 = w1s + nb_upper * BLOCK_N; |
| 805 | const float* __restrict__ Bs1 = w1s + nb_lower * BLOCK_N; |
| 806 | |
| 807 | if (use_brgemm) { |
| 808 | // 1.b gemm: C0 = A @ B0 |
| 809 | at::native::cpublas::brgemm( |
| 810 | /* M */ m_size, |
| 811 | /* N */ n_size, |
| 812 | /* K */ K, |
| 813 | /* lda */ K, |
| 814 | /* ldb */ n_size, |
| 815 | /* ldc */ BLOCK_N, |
| 816 | /* add_C */ false, |
| 817 | /* A */ A, |
| 818 | /* B */ B0, |
| 819 | /* C */ C0); |
| 820 | |
| 821 | // 1.c gemm: C1 = A @ B1 |
| 822 | at::native::cpublas::brgemm( |
| 823 | /* M */ m_size, |
| 824 | /* N */ n_size, |
| 825 | /* K */ K, |
| 826 | /* lda */ K, |
| 827 | /* ldb */ n_size, |
| 828 | /* ldc */ BLOCK_N, |
| 829 | /* add_C */ false, |
| 830 | /* A */ A, |
| 831 | /* B */ B1, |
| 832 | /* C */ C1); |
| 833 | |
| 834 | const int32_t* Bcomp0 = reinterpret_cast<const int32_t*>(B0 + block_size_n() * K); |
| 835 | const int32_t* Bcomp1 = reinterpret_cast<const int32_t*>(B1 + block_size_n() * K); |
| 836 | |
| 837 | // 1.d silu and mul |
| 838 | silu_and_mul<scalar_t, BLOCK_N>( |
| 839 | ic1 + mb * BLOCK_M * N + nb * BLOCK_N, C0, C1, As, Bs0, Bs1, Bcomp0, Bcomp1, m_size, N); |
| 840 | } else { |
| 841 | // fused 1.bcd: silu_and_mul(A @ B0, A @ B1) |
| 842 | tinygemm_kernel( |
| 843 | /* A */ A, |
| 844 | /* B0 */ B0, |
| 845 | /* B1 */ B1, |
| 846 | /* C */ ic1 + mb * BLOCK_M * N + nb * BLOCK_N, |
| 847 | /* As */ As, |
| 848 | /* Bs0 */ Bs0, |
| 849 | /* Bs1 */ Bs1, |
| 850 | /* M */ m_size, |
| 851 | /* N */ n_size, |
| 852 | /* K */ K, |
| 853 | /* lda */ K, |
| 854 | /* ldb */ n_size, |
| 855 | /* ldc */ N); |
| 856 | } |
| 857 | }); |
| 858 | |
| 859 | if (use_brgemm) { |
| 860 | at::native::cpublas::brgemm_release(); |
| 861 | } |
| 862 | }); |
| 863 | |
| 864 | // stage 1.5: quantize ic1 to uint8, [M * topk, N] |
| 865 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 866 | for (int64_t m = begin; m < end; ++m) { |
| 867 | quantize_row_int8<scalar_t>(Aq_tmp + m * N, As_tmp[m], ic1 + m * N, N); |
| 868 | } |
| 869 | }); |
| 870 | |
| 871 | // stage 2: intermediate_cache2 = intermediate_cache1 @ w2 |
| 872 | // w2 : [K, N] as [OC, IC] |
| 873 | const int64_t OC = K; // rename K as OC |
| 874 | const int64_t IC = N; // rename N as IC |
| 875 | const int64_t MB2 = MB; |
| 876 | const int64_t NB2 = div_up(OC, BLOCK_N); |
| 877 | const int64_t stride_oc = packed_N; |
| 878 | |
| 879 | // parallel on [MB2, NB2] |
| 880 | parallel_2d(MB2, NB2, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 881 | // get local pointers |
| 882 | int tid = get_thread_num(); |
| 883 | float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N; |
| 884 | int32_t* __restrict__ C32 = reinterpret_cast<int32_t*>(C + BLOCK_M * BLOCK_N); |
| 885 | |
| 886 | loop_2d<int8_t>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 887 | int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M); |
| 888 | int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N); |
| 889 | |
| 890 | // A shape [m_size, IC] |
| 891 | const uint8_t* __restrict__ A = Aq_tmp + mb * BLOCK_M * N; |
| 892 | const float* __restrict__ As = As_tmp + mb * BLOCK_M; |
| 893 | |
| 894 | // B shape [IC, n_size] in vnni format |
| 895 | const int8_t* __restrict__ B = packed_w2 + nb * BLOCK_N * stride_oc; |
| 896 | const float* __restrict__ Bs = w2s + nb * BLOCK_N; |
| 897 | |
| 898 | if (use_brgemm) { |
| 899 | at::native::cpublas::brgemm( |
| 900 | /* M */ m_size, |
| 901 | /* N */ n_size, |
| 902 | /* K */ IC, |
| 903 | /* lda */ IC, |
| 904 | /* ldb */ n_size, |
| 905 | /* ldc */ BLOCK_N, |
| 906 | /* add_C */ false, |
| 907 | /* A */ A, |
| 908 | /* B */ B, |
| 909 | /* C */ C32); |
| 910 | |
| 911 | // apply scales |
| 912 | const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * IC); |
| 913 | scale_C<BLOCK_N>(C, C32, As, Bs, Bcomp, m_size); |
| 914 | } else { |
| 915 | // 2.a gemm: C = A @ B |
| 916 | tinygemm_kernel<scalar_t>( |
| 917 | /* A */ A, |
| 918 | /* B */ B, |
| 919 | /* C */ C, |
| 920 | /* As */ As, |
| 921 | /* Bs */ Bs, |
| 922 | /* M */ m_size, |
| 923 | /* N */ n_size, |
| 924 | /* K */ IC, |
| 925 | /* lda */ IC, |
| 926 | /* ldb */ n_size, |
| 927 | /* ldc */ BLOCK_N); |
| 928 | } |
| 929 | |
| 930 | // 2.b copy from C to output and add fused_experts_out |
| 931 | scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N; |
| 932 | const scalar_t* __restrict__ fused_out = |
| 933 | apply_scaling_factor ? fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N : nullptr; |
| 934 | for (int64_t m = 0; m < m_size; ++m) { |
| 935 | const scalar_t* __restrict__ fused_out_row = apply_scaling_factor ? (fused_out + m * K) : nullptr; |
| 936 | add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out_row, routed_scaling_factor, n_size); |
| 937 | } |
| 938 | }); |
| 939 | |
| 940 | if (use_brgemm) { |
| 941 | at::native::cpublas::brgemm_release(); |
| 942 | } |
| 943 | }); |
| 944 | } |
| 945 | |
| 946 | #define INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(TYPE) \ |
| 947 | template void shared_expert_int8_kernel_impl<TYPE>( \ |
| 948 | TYPE* __restrict__ output, \ |
| 949 | TYPE* __restrict__ ic1, \ |
| 950 | float* __restrict__ C_tmp, \ |
| 951 | uint8_t* __restrict__ Aq_tmp, \ |
| 952 | float* __restrict__ As_tmp, \ |
| 953 | const TYPE* __restrict__ input, \ |
| 954 | const int8_t* __restrict__ packed_w1, \ |
| 955 | const int8_t* __restrict__ packed_w2, \ |
| 956 | const float* __restrict__ w1s, \ |
| 957 | const float* __restrict__ w2s, \ |
| 958 | const TYPE* __restrict__ fused_experts_out, \ |
| 959 | float routed_scaling_factor, \ |
| 960 | int64_t M, \ |
| 961 | int64_t N, \ |
| 962 | int64_t K) |
| 963 | |
| 964 | INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::BFloat16); |
| 965 | INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::Half); |