| 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 "vec.h" |
| 9 | |
| 10 | namespace { |
| 11 | |
| 12 | template <typename scalar_t, bool has_bias, int BLOCK_N> |
| 13 | struct scale_C { |
| 14 | static inline void apply( |
| 15 | scalar_t* __restrict__ C, |
| 16 | const int32_t* __restrict__ Ctmp, |
| 17 | const int32_t* __restrict__ Bcomp, |
| 18 | const float* __restrict__ bias, |
| 19 | float As, |
| 20 | const float* __restrict__ Bs) { |
| 21 | TORCH_CHECK(false, "scale_C: scalar path not implemented!"); |
| 22 | } |
| 23 | }; |
| 24 | |
| 25 | #if defined(CPU_CAPABILITY_AVX512) |
| 26 | template <bool has_bias, int BLOCK_N> |
| 27 | struct scale_C<at::BFloat16, has_bias, BLOCK_N> { |
| 28 | static inline void apply( |
| 29 | at::BFloat16* __restrict__ C, |
| 30 | const int32_t* __restrict__ Ctmp, |
| 31 | const int32_t* __restrict__ Bcomp, |
| 32 | const float* __restrict__ bias, |
| 33 | float As, |
| 34 | const float* __restrict__ Bs) { |
| 35 | constexpr int COLS = BLOCK_N / 16; |
| 36 | static_assert(COLS % 2 == 0); |
| 37 | |
| 38 | __m512 vc[COLS]; |
| 39 | __m512 vd0 = _mm512_set1_ps(As); |
| 40 | |
| 41 | auto compute = [&](auto col) { |
| 42 | __m512 vd1 = _mm512_loadu_ps(Bs + col * 16); |
| 43 | __m512i vcomp = _mm512_loadu_si512(Bcomp + col * 16); |
| 44 | __m512i vc32 = _mm512_loadu_si512(Ctmp + col * 16); |
| 45 | vc[col] = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc32, vcomp)); |
| 46 | if constexpr (has_bias) { |
| 47 | __m512 vbias = _mm512_loadu_ps(bias + col * 16); |
| 48 | vc[col] = _mm512_fmadd_ps(_mm512_mul_ps(vc[col], vd0), vd1, vbias); |
| 49 | } else { |
| 50 | vc[col] = _mm512_mul_ps(_mm512_mul_ps(vc[col], vd0), vd1); |
| 51 | } |
| 52 | }; |
| 53 | Unroll<COLS>{}(compute); |
| 54 | |
| 55 | auto storec = [&](auto col) { |
| 56 | // for COLS = 2, 4 use 512bit store |
| 57 | if constexpr (col % 2 == 0) { |
| 58 | _mm512_storeu_si512( |
| 59 | reinterpret_cast<__m512i*>((C + col * 16)), (__m512i)(_mm512_cvtne2ps_pbh(vc[col + 1], vc[col + 0]))); |
| 60 | } |
| 61 | }; |
| 62 | Unroll<COLS>{}(storec); |
| 63 | } |
| 64 | }; |
| 65 | #endif |
| 66 | |
| 67 | template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N> |
| 68 | struct tinygemm_kernel_nn { |
| 69 | static inline void apply( |
| 70 | const uint8_t* __restrict__ A, |
| 71 | const int8_t* __restrict__ B, |
| 72 | scalar_t* __restrict__ C, |
| 73 | const float* __restrict__ As, |
| 74 | const float* __restrict__ Bs, |
| 75 | const int32_t* __restrict__ Bcomp, |
| 76 | const float* __restrict__ bias, |
| 77 | int64_t K, |
| 78 | int64_t lda, |
| 79 | int64_t ldb, |
| 80 | int64_t ldc) { |
| 81 | TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!"); |
| 82 | } |
| 83 | }; |
| 84 | |
| 85 | #if defined(CPU_CAPABILITY_AVX512) |
| 86 | template <bool has_bias, int BLOCK_M, int BLOCK_N> |
| 87 | struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> { |
| 88 | static inline void apply( |
| 89 | const uint8_t* __restrict__ A, |
| 90 | const int8_t* __restrict__ B, |
| 91 | at::BFloat16* __restrict__ C, |
| 92 | const float* __restrict__ As, |
| 93 | const float* __restrict__ Bs, |
| 94 | const int32_t* __restrict__ Bcomp, |
| 95 | const float* __restrict__ bias, |
| 96 | int64_t K, |
| 97 | int64_t lda, |
| 98 | int64_t ldb, |
| 99 | int64_t ldc) { |
| 100 | constexpr int ROWS = BLOCK_M; |
| 101 | constexpr int COLS = BLOCK_N / 16; |
| 102 | static_assert(COLS % 2 == 0); |
| 103 | |
| 104 | // prefetch distance |
| 105 | constexpr int PREFETCH_SIZE_K = 0; |
| 106 | |
| 107 | __m512i va; |
| 108 | __m512i vb[COLS]; |
| 109 | __m512i vc[ROWS * COLS]; |
| 110 | __m512i vcomp[COLS]; |
| 111 | __m512 vd0; |
| 112 | __m512 vd1[COLS]; |
| 113 | |
| 114 | // oops! 4x4 spills but we use 4x2 |
| 115 | __m512 vbias[COLS]; |
| 116 | |
| 117 | // [NOTE]: s8s8 igemm compensation in avx512-vnni |
| 118 | // |
| 119 | // avx512-vnni has no s8s8, so we need to change s8s8 to u8s8 with compensate: |
| 120 | // |
| 121 | // a * b = (a + 128) * b - 128 * b |
| 122 | // s s u s u s |
| 123 | // |
| 124 | // 1) 128 * b is pre-computed when packing B to vnni formats |
| 125 | // 2) a + 128 is fused when dynamically quantize A |
| 126 | // |
| 127 | auto loadc = [&](auto i) { vc[i] = _mm512_set1_epi32(0); }; |
| 128 | Unroll<ROWS * COLS>{}(loadc); |
| 129 | |
| 130 | const int64_t K4 = K >> 2; |
| 131 | const int64_t lda4 = lda >> 2; |
| 132 | const int64_t ldb4 = ldb; // ldb * 4 >> 2; |
| 133 | const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A); |
| 134 | const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B); |
| 135 | |
| 136 | auto compute = [&](auto i, int64_t k) { |
| 137 | constexpr int row = i / COLS; |
| 138 | constexpr int col = i % COLS; |
| 139 | |
| 140 | if constexpr (col == 0) { |
| 141 | va = _mm512_set1_epi32(a_ptr[row * lda4 + k]); |
| 142 | } |
| 143 | if constexpr (row == 0) { |
| 144 | vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16); |
| 145 | if constexpr (PREFETCH_SIZE_K > 0) { |
| 146 | _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb4 + col * 16, _MM_HINT_T0); |
| 147 | } |
| 148 | } |
| 149 | vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]); |
| 150 | }; |
| 151 | for (int64_t k = 0; k < K4; ++k) { |
| 152 | Unroll<ROWS * COLS>{}(compute, k); |
| 153 | } |
| 154 | |
| 155 | auto storec = [&](auto i) { |
| 156 | constexpr int row = i / COLS; |
| 157 | constexpr int col = i % COLS; |
| 158 | |
| 159 | // load a scale |
| 160 | if constexpr (col == 0) { |
| 161 | vd0 = _mm512_set1_ps(As[row]); |
| 162 | } |
| 163 | // load b scale and vcomp per 2 vectors |
| 164 | // also load bias if any |
| 165 | if constexpr (row == 0) { |
| 166 | if constexpr (col % 2 == 0) { |
| 167 | vd1[col + 0] = _mm512_loadu_ps(Bs + col * 16); |
| 168 | vd1[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16); |
| 169 | vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16); |
| 170 | vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16); |
| 171 | if constexpr (has_bias) { |
| 172 | vbias[col + 0] = _mm512_loadu_ps(bias + col * 16); |
| 173 | vbias[col + 1] = _mm512_loadu_ps(bias + col * 16 + 16); |
| 174 | } |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | // for COLS = 2, 4 use 512bit store |
| 179 | if constexpr (col % 2 == 0) { |
| 180 | __m512 vc0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 0], vcomp[col + 0])); |
| 181 | __m512 vc1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 1], vcomp[col + 1])); |
| 182 | if constexpr (has_bias) { |
| 183 | vc0 = _mm512_fmadd_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0], vbias[col + 0]); |
| 184 | vc1 = _mm512_fmadd_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1], vbias[col + 1]); |
| 185 | } else { |
| 186 | vc0 = _mm512_mul_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0]); |
| 187 | vc1 = _mm512_mul_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1]); |
| 188 | } |
| 189 | |
| 190 | _mm512_storeu_si512( |
| 191 | reinterpret_cast<__m512i*>((C + row * ldc + col * 16)), (__m512i)(_mm512_cvtne2ps_pbh(vc1, vc0))); |
| 192 | } |
| 193 | }; |
| 194 | Unroll<ROWS * COLS>{}(storec); |
| 195 | } |
| 196 | }; |
| 197 | #endif |
| 198 | |
| 199 | #define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE) \ |
| 200 | tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply( \ |
| 201 | A + mb_start * lda, \ |
| 202 | B + nb_start * 4, \ |
| 203 | C + mb_start * ldc + nb_start, \ |
| 204 | As + mb_start, \ |
| 205 | Bs + nb_start, \ |
| 206 | Bcomp + nb_start, \ |
| 207 | has_bias ? bias + nb_start : nullptr, \ |
| 208 | K, \ |
| 209 | lda, \ |
| 210 | ldb, \ |
| 211 | ldc); |
| 212 | |
| 213 | template <typename scalar_t, bool has_bias> |
| 214 | void tinygemm_kernel( |
| 215 | const uint8_t* __restrict__ A, |
| 216 | const int8_t* __restrict__ B, |
| 217 | scalar_t* __restrict__ C, |
| 218 | int32_t* __restrict__ Ctmp, |
| 219 | const float* __restrict__ As, |
| 220 | const float* __restrict__ Bs, |
| 221 | const float* __restrict__ bias, |
| 222 | int64_t M, |
| 223 | int64_t N, |
| 224 | int64_t K, |
| 225 | int64_t lda, |
| 226 | int64_t ldb, |
| 227 | int64_t ldc, |
| 228 | bool brg) { |
| 229 | // B compensation |
| 230 | const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K); |
| 231 | |
| 232 | if (brg) { |
| 233 | constexpr int BLOCK_N = block_size_n(); |
| 234 | at::native::cpublas::brgemm(M, N, K, lda, ldb, BLOCK_N, /* add_C */ false, A, B, Ctmp); |
| 235 | |
| 236 | // apply compensation and scale |
| 237 | for (int64_t m = 0; m < M; ++m) { |
| 238 | scale_C<scalar_t, has_bias, BLOCK_N>::apply(C + m * ldc, Ctmp + m * BLOCK_N, Bcomp, bias, As[m], Bs); |
| 239 | } |
| 240 | return; |
| 241 | } |
| 242 | |
| 243 | // pattern: 1-4-16 |
| 244 | constexpr int64_t BLOCK_M = 4; |
| 245 | constexpr int64_t BLOCK_N = 64; |
| 246 | const int64_t MB = div_up(M, BLOCK_M); |
| 247 | const int64_t NB = div_up(N, BLOCK_N); |
| 248 | for (int64_t mb = 0; mb < MB; ++mb) { |
| 249 | int64_t mb_start = mb * BLOCK_M; |
| 250 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 251 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 252 | int64_t nb_start = nb * BLOCK_N; |
| 253 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 254 | |
| 255 | switch (mb_size << 4 | nb_size >> 4) { |
| 256 | // mb_size = 1 |
| 257 | case 0x12: |
| 258 | LAUNCH_TINYGEMM_KERNEL_NN(1, 32); |
| 259 | break; |
| 260 | case 0x14: |
| 261 | LAUNCH_TINYGEMM_KERNEL_NN(1, 64); |
| 262 | break; |
| 263 | // mb_size = 2 |
| 264 | case 0x22: |
| 265 | LAUNCH_TINYGEMM_KERNEL_NN(2, 32); |
| 266 | break; |
| 267 | case 0x24: |
| 268 | LAUNCH_TINYGEMM_KERNEL_NN(2, 64); |
| 269 | break; |
| 270 | // mb_size = 3 |
| 271 | case 0x32: |
| 272 | LAUNCH_TINYGEMM_KERNEL_NN(3, 32); |
| 273 | break; |
| 274 | case 0x34: |
| 275 | LAUNCH_TINYGEMM_KERNEL_NN(3, 64); |
| 276 | break; |
| 277 | // mb_size = 4 |
| 278 | case 0x42: |
| 279 | LAUNCH_TINYGEMM_KERNEL_NN(4, 32); |
| 280 | break; |
| 281 | case 0x44: |
| 282 | LAUNCH_TINYGEMM_KERNEL_NN(4, 64); |
| 283 | break; |
| 284 | default: |
| 285 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 286 | } |
| 287 | } |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | template <typename scalar_t> |
| 292 | void int8_scaled_mm_kernel_impl( |
| 293 | scalar_t* __restrict__ out, |
| 294 | const uint8_t* __restrict__ mat1, |
| 295 | const int8_t* __restrict__ mat2, |
| 296 | const float* __restrict__ scales1, |
| 297 | const float* __restrict__ scales2, |
| 298 | const float* __restrict__ bias, |
| 299 | int64_t M, |
| 300 | int64_t N, |
| 301 | int64_t K) { |
| 302 | constexpr int64_t BLOCK_M = block_size_m(); |
| 303 | constexpr int64_t BLOCK_N = block_size_n(); |
| 304 | const int64_t MB = div_up(M, BLOCK_M); |
| 305 | const int64_t NB = div_up(N, BLOCK_N); |
| 306 | |
| 307 | const bool use_brgemm = can_use_brgemm<int8_t>(M); |
| 308 | |
| 309 | // K + 4 after compensation |
| 310 | const int64_t packed_row_size = get_row_size<int8_t>(K); |
| 311 | |
| 312 | AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] { |
| 313 | parallel_2d(MB, NB, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 314 | // for brgemm, use int32_t for accumulate |
| 315 | alignas(64) int32_t Ctmp[BLOCK_M * BLOCK_N]; |
| 316 | |
| 317 | loop_2d<int8_t>(mb0, mb1, nb0, nb1, BLOCK_N * K, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 318 | int mb_start = mb * BLOCK_M; |
| 319 | int mb_size = std::min(M - mb_start, BLOCK_M); |
| 320 | int nb_start = nb * BLOCK_N; |
| 321 | int nb_size = std::min(N - nb_start, BLOCK_N); |
| 322 | |
| 323 | tinygemm_kernel<scalar_t, has_bias>( |
| 324 | /* A */ mat1 + mb_start * K, |
| 325 | /* B */ mat2 + nb_start * packed_row_size /* nb * BLOCK_N * (K + 4) */, |
| 326 | /* C */ out + mb_start * N + nb_start, |
| 327 | /* Ctmp*/ Ctmp, |
| 328 | /* As */ scales1 + mb_start, |
| 329 | /* Bs */ scales2 + nb_start, |
| 330 | /* bias*/ bias + nb_start, |
| 331 | /* M */ mb_size, |
| 332 | /* N */ nb_size, |
| 333 | /* K */ K, |
| 334 | /* lda */ K, |
| 335 | /* ldb */ nb_size, |
| 336 | /* ldc */ N, |
| 337 | /* brg */ use_brgemm); |
| 338 | }); |
| 339 | |
| 340 | if (use_brgemm) { |
| 341 | at::native::cpublas::brgemm_release(); |
| 342 | } |
| 343 | }); |
| 344 | }); |
| 345 | } |
| 346 | |
| 347 | } // anonymous namespace |
| 348 | |
| 349 | // tinygemm interface |
| 350 | template <typename scalar_t> |
| 351 | void tinygemm_kernel( |
| 352 | const uint8_t* __restrict__ A, |
| 353 | const int8_t* __restrict__ B, |
| 354 | scalar_t* __restrict__ C, |
| 355 | int32_t* __restrict__ Ctmp, |
| 356 | const float* __restrict__ As, |
| 357 | const float* __restrict__ Bs, |
| 358 | int64_t M, |
| 359 | int64_t N, |
| 360 | int64_t K, |
| 361 | int64_t lda, |
| 362 | int64_t ldb, |
| 363 | int64_t ldc, |
| 364 | bool brg) { |
| 365 | tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, As, Bs, nullptr, M, N, K, lda, ldb, ldc, brg); |
| 366 | } |
| 367 | |
| 368 | #define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE) \ |
| 369 | template void tinygemm_kernel<TYPE>( \ |
| 370 | const uint8_t* __restrict__ A, \ |
| 371 | const int8_t* __restrict__ B, \ |
| 372 | TYPE* __restrict__ C, \ |
| 373 | int32_t* __restrict__ Ctmp, \ |
| 374 | const float* __restrict__ As, \ |
| 375 | const float* __restrict__ Bs, \ |
| 376 | int64_t M, \ |
| 377 | int64_t N, \ |
| 378 | int64_t K, \ |
| 379 | int64_t lda, \ |
| 380 | int64_t ldb, \ |
| 381 | int64_t ldc, \ |
| 382 | bool brg) |
| 383 | |
| 384 | INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16); |
| 385 | INSTANTIATE_TINYGEMM_TEMPLATE(at::Half); |
| 386 | |
| 387 | std::tuple<at::Tensor, at::Tensor> per_token_quant_int8_cpu(at::Tensor& A) { |
| 388 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(A); |
| 389 | CHECK_DIM(2, A); |
| 390 | |
| 391 | int64_t M = A.size(0); |
| 392 | int64_t K = A.size(1); |
| 393 | int64_t lda = A.stride(0); |
| 394 | |
| 395 | const auto st = A.scalar_type(); |
| 396 | TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf, "per_token_quant_int8: expect A to be bfloat16 or half."); |
| 397 | |
| 398 | auto Aq = at::empty({M, K}, A.options().dtype(at::kByte)); |
| 399 | auto As = at::empty({M}, A.options().dtype(at::kFloat)); |
| 400 | |
| 401 | AT_DISPATCH_REDUCED_FLOATING_TYPES(st, "per_token_quant_int8", [&] { |
| 402 | uint8_t* __restrict__ Aq_data = Aq.data_ptr<uint8_t>(); |
| 403 | float* __restrict__ As_data = As.data_ptr<float>(); |
| 404 | const scalar_t* __restrict__ A_data = A.data_ptr<scalar_t>(); |
| 405 | |
| 406 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 407 | for (int64_t m = begin; m < end; ++m) { |
| 408 | quantize_row_int8<scalar_t>(Aq_data + m * K, As_data[m], A_data + m * lda, K); |
| 409 | } |
| 410 | }); |
| 411 | }); |
| 412 | return std::make_tuple(Aq, As); |
| 413 | } |
| 414 | |
| 415 | // weight : static, per-channel, symmetric |
| 416 | // activation : dynamic, per-token, symmetric |
| 417 | // |
| 418 | // mat1 : [M, K] |
| 419 | // mat2 : [N, K] |
| 420 | // scales1 : [M] |
| 421 | // scales2 : [N] |
| 422 | // bias : [N] |
| 423 | // out : [M, N] |
| 424 | // |
| 425 | at::Tensor int8_scaled_mm_cpu( |
| 426 | at::Tensor& mat1, |
| 427 | at::Tensor& mat2, |
| 428 | at::Tensor& scales1, |
| 429 | at::Tensor& scales2, |
| 430 | const std::optional<at::Tensor>& bias, |
| 431 | at::ScalarType out_dtype, |
| 432 | bool is_vnni) { |
| 433 | auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2); |
| 434 | |
| 435 | CHECK_INPUT(mat1); |
| 436 | CHECK_INPUT(mat2); |
| 437 | CHECK_INPUT(scales1); |
| 438 | CHECK_INPUT(scales2); |
| 439 | CHECK_DIM(2, mat1); |
| 440 | CHECK_DIM(2, mat2); |
| 441 | |
| 442 | int64_t M = mat1.size(0); |
| 443 | int64_t N = mat2.size(0); |
| 444 | int64_t K = mat1.size(1); |
| 445 | |
| 446 | // see [NOTE]: s8s8 igemm compensation in avx512-vnni |
| 447 | CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K)); |
| 448 | CHECK_EQ(scales1.numel(), M); |
| 449 | CHECK_EQ(scales2.numel(), N); |
| 450 | |
| 451 | TORCH_CHECK(mat1.scalar_type() == at::kByte, "int8_scaled_mm: expect mat1 to be uint8."); |
| 452 | TORCH_CHECK(mat2.scalar_type() == at::kChar, "int8_scaled_mm: expect mat2 to be int8."); |
| 453 | TORCH_CHECK( |
| 454 | scales1.scalar_type() == at::kFloat && scales2.scalar_type() == at::kFloat, |
| 455 | "int8_scaled_mm: expect scales to be float32."); |
| 456 | |
| 457 | auto out = at::empty({M, N}, mat1.options().dtype(out_dtype)); |
| 458 | |
| 459 | const bool has_bias = bias.has_value(); |
| 460 | const float* bias_data = nullptr; |
| 461 | if (has_bias) { |
| 462 | CHECK_EQ(bias.value().size(0), N); |
| 463 | bias_data = bias.value().data_ptr<float>(); |
| 464 | } |
| 465 | |
| 466 | AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_kernel_impl", [&] { |
| 467 | int8_scaled_mm_kernel_impl<scalar_t>( |
| 468 | out.data_ptr<scalar_t>(), |
| 469 | mat1.data_ptr<uint8_t>(), |
| 470 | packed_w.data_ptr<int8_t>(), |
| 471 | scales1.data_ptr<float>(), |
| 472 | scales2.data_ptr<float>(), |
| 473 | bias_data, |
| 474 | M, |
| 475 | N, |
| 476 | K); |
| 477 | }); |
| 478 | return out; |
| 479 | } |
| 480 | |
| 481 | // fused `per_token_quant_int8_cpu` and `int8_scaled_mm_cpu` |
| 482 | at::Tensor int8_scaled_mm_with_quant( |
| 483 | at::Tensor& mat1, |
| 484 | at::Tensor& mat2, |
| 485 | at::Tensor& scales2, |
| 486 | const std::optional<at::Tensor>& bias, |
| 487 | at::ScalarType out_dtype, |
| 488 | bool is_vnni) { |
| 489 | auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2); |
| 490 | |
| 491 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1); |
| 492 | CHECK_INPUT(mat2); |
| 493 | CHECK_INPUT(scales2); |
| 494 | CHECK_DIM(2, mat1); |
| 495 | CHECK_DIM(2, mat2); |
| 496 | |
| 497 | int64_t M = mat1.size(0); |
| 498 | int64_t N = mat2.size(0); |
| 499 | int64_t K = mat1.size(1); |
| 500 | int64_t lda = mat1.stride(0); |
| 501 | |
| 502 | // see [NOTE]: s8s8 igemm compensation in avx512-vnni |
| 503 | CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K)); |
| 504 | CHECK_EQ(scales2.numel(), N); |
| 505 | |
| 506 | const auto st = mat1.scalar_type(); |
| 507 | TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf, "int8_scaled_mm_with_quant: expect A to be bfloat16 or half."); |
| 508 | TORCH_CHECK(st == out_dtype, "int8_scaled_mm_with_quant: expect A has same dtype with out_dtype."); |
| 509 | TORCH_CHECK(mat2.scalar_type() == at::kChar, "int8_scaled_mm_with_quant: expect mat2 to be int8."); |
| 510 | TORCH_CHECK(scales2.scalar_type() == at::kFloat, "int8_scaled_mm_with_quant: expect scales to be float32."); |
| 511 | |
| 512 | const int64_t buffer_size = M * K + M * sizeof(float); |
| 513 | auto buffer = at::empty({buffer_size}, mat1.options().dtype(at::kByte)); |
| 514 | auto out = at::empty({M, N}, mat1.options().dtype(out_dtype)); |
| 515 | |
| 516 | const bool has_bias = bias.has_value(); |
| 517 | const float* bias_data = nullptr; |
| 518 | if (has_bias) { |
| 519 | CHECK_EQ(bias.value().size(0), N); |
| 520 | bias_data = bias.value().data_ptr<float>(); |
| 521 | } |
| 522 | |
| 523 | AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_with_quant_kernel_impl", [&] { |
| 524 | uint8_t* __restrict__ Aq_data = buffer.data_ptr<uint8_t>(); |
| 525 | float* __restrict__ As_data = (float*)((void*)(Aq_data + M * K)); |
| 526 | const scalar_t* __restrict__ A_data = mat1.data_ptr<scalar_t>(); |
| 527 | |
| 528 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 529 | for (int64_t m = begin; m < end; ++m) { |
| 530 | quantize_row_int8<scalar_t>(Aq_data + m * K, As_data[m], A_data + m * lda, K); |
| 531 | } |
| 532 | }); |
| 533 | |
| 534 | int8_scaled_mm_kernel_impl<scalar_t>( |
| 535 | out.data_ptr<scalar_t>(), |
| 536 | Aq_data, |
| 537 | packed_w.data_ptr<int8_t>(), |
| 538 | As_data, |
| 539 | scales2.data_ptr<float>(), |
| 540 | bias_data, |
| 541 | M, |
| 542 | N, |
| 543 | K); |
| 544 | }); |
| 545 | return out; |
| 546 | } |