| 1 | // Adapted from |
| 2 | // https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu |
| 3 | |
| 4 | // clang-format off |
| 5 | |
| 6 | // To use the transpose functions |
| 7 | #include <ATen/native/cpu/utils.h> |
| 8 | |
| 9 | #include "vec.h" |
| 10 | |
| 11 | namespace { |
| 12 | |
| 13 | using namespace at::vec; |
| 14 | |
| 15 | template <typename index_t> |
| 16 | inline index_t get_index(index_t* ind, int i) { |
| 17 | return (ind == nullptr) ? (index_t)i : ind[i]; |
| 18 | } |
| 19 | |
| 20 | #if defined(CPU_CAPABILITY_AVX512) |
| 21 | // key: from [N, 32] to [32/2, N, 2] |
| 22 | template <typename scalar_t, typename index_t> |
| 23 | inline void pack_vnni_Nx32( |
| 24 | scalar_t* __restrict__ dst, |
| 25 | const scalar_t* __restrict__ src, |
| 26 | const index_t* __restrict__ ind, |
| 27 | int N, |
| 28 | int ld_src, |
| 29 | int ld_dst) { |
| 30 | __m512i vinputs[16]; |
| 31 | |
| 32 | int n = 0; |
| 33 | for (; n < N; ++n) { |
| 34 | index_t index = get_index(ind, n); |
| 35 | vinputs[n] = _mm512_loadu_si512(src + index * ld_src); |
| 36 | } |
| 37 | // padding with zero to avoid uninitialized vectors |
| 38 | for (; n < 16; ++n) { |
| 39 | vinputs[n] = _mm512_set1_epi32(0); |
| 40 | } |
| 41 | |
| 42 | // pack key |
| 43 | transpose_16x16_32bit(vinputs); |
| 44 | |
| 45 | const __mmask16 vmask = (1 << N) - 1; |
| 46 | for (int k = 0; k < 16; ++k) { |
| 47 | _mm512_mask_storeu_epi32(dst + k * ld_dst * 2, vmask, vinputs[k]); |
| 48 | } |
| 49 | } |
| 50 | |
| 51 | template <typename scalar_t, typename index_t> |
| 52 | inline void pack_vnni_N_remainder( |
| 53 | scalar_t* __restrict__ dst, |
| 54 | const scalar_t* __restrict__ src, |
| 55 | const index_t* __restrict__ ind, |
| 56 | int N, |
| 57 | int K, |
| 58 | int ld_src, |
| 59 | int ld_dst) { |
| 60 | __m512i vinputs[16]; |
| 61 | |
| 62 | int K2 = K >> 1; |
| 63 | const __mmask16 vmask = (1 << K2) - 1; |
| 64 | |
| 65 | int n = 0; |
| 66 | for (; n < N; ++n) { |
| 67 | index_t index = get_index(ind, n); |
| 68 | vinputs[n] = _mm512_maskz_loadu_epi32(vmask, src + index * ld_src); |
| 69 | } |
| 70 | // padding with zero to avoid uninitialized vectors |
| 71 | for (; n < 16; ++n) { |
| 72 | vinputs[n] = _mm512_set1_epi32(0); |
| 73 | } |
| 74 | |
| 75 | // pack key |
| 76 | transpose_16x16_32bit(vinputs); |
| 77 | |
| 78 | const __mmask16 vmask2 = (1 << N) - 1; |
| 79 | for (int k = 0; k < K2; ++k) { |
| 80 | _mm512_mask_storeu_epi32(dst + k * ld_dst * 2, vmask2, vinputs[k]); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | // value: from [K, 32] to [K/2, 32, 2] |
| 85 | template <typename scalar_t, typename index_t> |
| 86 | inline void pack_vnni_Kx32( |
| 87 | scalar_t* __restrict__ dst, |
| 88 | const scalar_t* __restrict__ src, |
| 89 | const index_t* __restrict__ ind, |
| 90 | int K, |
| 91 | int ld_src, |
| 92 | int ld_dst) { |
| 93 | __m512i vinputs[2]; |
| 94 | |
| 95 | int k = 0; |
| 96 | for (; k < K; ++k) { |
| 97 | index_t index = get_index(ind, k); |
| 98 | vinputs[k] = _mm512_loadu_si512(src + index * ld_src); |
| 99 | } |
| 100 | // padding with zero to avoid uninitialized vectors |
| 101 | for (; k < 2; ++k) { |
| 102 | vinputs[k] = _mm512_set1_epi32(0); |
| 103 | } |
| 104 | |
| 105 | // pack value |
| 106 | __m512i d0, d1; |
| 107 | std::tie(d0, d1) = transpose_2x32_16bit(vinputs[0], vinputs[1]); |
| 108 | _mm512_storeu_si512(dst + 0 * ld_dst * 2, d0); |
| 109 | _mm512_storeu_si512(dst + 0 * ld_dst * 2 + 32, d1); |
| 110 | } |
| 111 | |
| 112 | template <typename scalar_t, typename index_t> |
| 113 | inline void pack_vnni_K_remainder( |
| 114 | scalar_t* __restrict__ dst, |
| 115 | const scalar_t* __restrict__ src, |
| 116 | const index_t* __restrict__ ind, |
| 117 | int K, |
| 118 | int N, |
| 119 | int ld_src, |
| 120 | int ld_dst) { |
| 121 | __m512i vinputs[2]; |
| 122 | |
| 123 | const __mmask32 vmask = (1 << N) - 1; |
| 124 | |
| 125 | int k = 0; |
| 126 | for (; k < K; ++k) { |
| 127 | index_t index = get_index(ind, k); |
| 128 | vinputs[k] = _mm512_maskz_loadu_epi16(vmask, src + index * ld_src); |
| 129 | } |
| 130 | // padding with zero to avoid uninitialized vectors |
| 131 | for (; k < 2; ++k) { |
| 132 | vinputs[k] = _mm512_set1_epi32(0); |
| 133 | } |
| 134 | |
| 135 | // pack value |
| 136 | __m512i d0, d1; |
| 137 | std::tie(d0, d1) = transpose_2x32_16bit(vinputs[0], vinputs[1]); |
| 138 | |
| 139 | if (N <= 16) { |
| 140 | // 2N * 16bits: N * 32bits |
| 141 | const __mmask16 vmask2 = (1 << N) - 1; |
| 142 | _mm512_mask_storeu_epi32(dst + 0 * ld_dst * 2, vmask2, d0); |
| 143 | } else { |
| 144 | // 2(N-16) * 16bits: (N-16) * 32bits |
| 145 | const __mmask16 vmask2 = (1 << (N - 16)) - 1; |
| 146 | _mm512_storeu_epi32(dst + 0 * ld_dst * 2, d0); |
| 147 | _mm512_mask_storeu_epi32(dst + 0 * ld_dst * 2 + 32, vmask2, d1); |
| 148 | } |
| 149 | } |
| 150 | #endif |
| 151 | |
| 152 | // convert to vnni format |
| 153 | // from [N, K/2, 2] to [K/2, N, 2] for bfloat16 and float16 |
| 154 | template <typename scalar_t, typename index_t, bool is_indexed> |
| 155 | void pack_vnni( |
| 156 | scalar_t* __restrict__ dst, |
| 157 | const scalar_t* __restrict__ src, |
| 158 | const index_t* __restrict__ ind, |
| 159 | int N, |
| 160 | int K, |
| 161 | int ld_src, |
| 162 | int ld_dst) { |
| 163 | #if defined(CPU_CAPABILITY_AVX512) |
| 164 | const int NB = div_up(N, 16); |
| 165 | const int KB = K / 32; |
| 166 | const int K_remainder = K - KB * 32; |
| 167 | |
| 168 | for (int nb = 0; nb < NB; ++nb) { |
| 169 | int nb_size = std::min(N - nb * 16, 16); |
| 170 | for (int kb = 0; kb < KB; ++kb) { |
| 171 | // handle 16x512bits each block |
| 172 | pack_vnni_Nx32<scalar_t, index_t>( |
| 173 | /* dst */ dst + ((kb * 32) >> 1) * ld_dst * 2 + nb * 16 * 2, |
| 174 | /* src */ src + kb * 32 + (is_indexed ? 0 : nb * 16 * ld_src), |
| 175 | /* ind */ is_indexed ? ind + nb * 16 : nullptr, |
| 176 | /* N */ nb_size, |
| 177 | /* ld_src */ ld_src, |
| 178 | /* ld_dst */ ld_dst); |
| 179 | } |
| 180 | if (K_remainder > 0) { |
| 181 | pack_vnni_N_remainder<scalar_t, index_t>( |
| 182 | /* dst */ dst + ((KB * 32) >> 1) * ld_dst * 2 + nb * 16 * 2, |
| 183 | /* src */ src + KB * 32 + (is_indexed ? 0 : nb * 16 * ld_src), |
| 184 | /* ind */ is_indexed ? ind + nb * 16 : nullptr, |
| 185 | /* N */ nb_size, |
| 186 | /* K */ K_remainder, |
| 187 | /* ld_src */ ld_src, |
| 188 | /* ld_dst */ ld_dst); |
| 189 | } |
| 190 | } |
| 191 | #else |
| 192 | for (int n = 0; n < N; ++n) { |
| 193 | index_t index = get_index(ind, n); |
| 194 | for (int k = 0; k < K / 2; ++k) { |
| 195 | for (int d = 0; d < 2; ++d) { |
| 196 | dst[k * ld_dst * 2 + n * 2 + d] = src[index * ld_src + k * 2 + d]; |
| 197 | } |
| 198 | } |
| 199 | } |
| 200 | #endif |
| 201 | } |
| 202 | |
| 203 | template <typename scalar_t> |
| 204 | void pack_vnni(scalar_t* __restrict__ dst, const scalar_t* __restrict__ src, int N, int K, int ld_src, int ld_dst) { |
| 205 | pack_vnni<scalar_t, int32_t, false>(dst, src, nullptr, N, K, ld_src, ld_dst); |
| 206 | } |
| 207 | |
| 208 | template <typename scalar_t, typename index_t> |
| 209 | void pack_vnni( |
| 210 | scalar_t* __restrict__ dst, |
| 211 | const scalar_t* __restrict__ src, |
| 212 | const index_t* __restrict__ ind, |
| 213 | int N, |
| 214 | int K, |
| 215 | int ld_src, |
| 216 | int ld_dst) { |
| 217 | assert(ind != nullptr); |
| 218 | pack_vnni<scalar_t, index_t, true>(dst, src, ind, N, K, ld_src, ld_dst); |
| 219 | } |
| 220 | |
| 221 | // convert to vnni format |
| 222 | // from [K/2, 2, N] to [K/2, N, 2] for bfloat16 and float16 |
| 223 | template <typename scalar_t, typename index_t, bool is_indexed> |
| 224 | void pack_vnni2( |
| 225 | scalar_t* __restrict__ dst, |
| 226 | const scalar_t* __restrict__ src, |
| 227 | const index_t* __restrict__ ind, |
| 228 | int K, |
| 229 | int N, |
| 230 | int ld_src, |
| 231 | int ld_dst) { |
| 232 | #if defined(CPU_CAPABILITY_AVX512) |
| 233 | const int KB = div_up(K, 2); |
| 234 | const int NB = N / 32; |
| 235 | const int N_remainder = N - NB * 32; |
| 236 | |
| 237 | for (int kb = 0; kb < KB; ++kb) { |
| 238 | int kb_size = std::min(K - kb * 2, 2); |
| 239 | for (int nb = 0; nb < NB; ++nb) { |
| 240 | // handle 2x512bits each block |
| 241 | pack_vnni_Kx32<scalar_t, index_t>( |
| 242 | /* dst */ dst + ((kb * 2) >> 1) * ld_dst * 2 + nb * 32 * 2, |
| 243 | /* src */ src + (is_indexed ? 0 : kb * 2 * ld_src) + nb * 32, |
| 244 | /* ind */ is_indexed ? ind + kb * 2 : nullptr, |
| 245 | /* K */ kb_size, |
| 246 | /* ld_src */ ld_src, |
| 247 | /* ld_dst */ ld_dst); |
| 248 | } |
| 249 | if (N_remainder > 0) { |
| 250 | pack_vnni_K_remainder( |
| 251 | /* dst */ dst + ((kb * 2) >> 1) * ld_dst * 2 + NB * 32 * 2, |
| 252 | /* src */ src + (is_indexed ? 0 : kb * 2 * ld_src) + NB * 32, |
| 253 | /* ind */ is_indexed ? ind + kb * 2 : nullptr, |
| 254 | /* K */ kb_size, |
| 255 | /* N */ N_remainder, |
| 256 | /* ld_src */ ld_src, |
| 257 | /* ld_dst */ ld_dst); |
| 258 | } |
| 259 | } |
| 260 | #else |
| 261 | int k = 0; |
| 262 | for (; k < (K >> 1) * 2; k += 2) { |
| 263 | index_t index0 = get_index(ind, k + 0); |
| 264 | index_t index1 = get_index(ind, k + 1); |
| 265 | for (int n = 0; n < N; ++n) { |
| 266 | dst[(k >> 1) * ld_dst * 2 + n * 2 + 0] = src[index0 * ld_src + n]; |
| 267 | dst[(k >> 1) * ld_dst * 2 + n * 2 + 1] = src[index1 * ld_src + n]; |
| 268 | } |
| 269 | } |
| 270 | if (K % 2 != 0) { |
| 271 | index_t index = get_index(ind, K - 1); |
| 272 | for (int n = 0; n < N; ++n) { |
| 273 | dst[(K >> 1) * ld_dst * 2 + n * 2 + 0] = src[index * ld_src + n]; |
| 274 | dst[(K >> 1) * ld_dst * 2 + n * 2 + 1] = 0; |
| 275 | } |
| 276 | k += 2; |
| 277 | } |
| 278 | #endif |
| 279 | } |
| 280 | |
| 281 | template <typename scalar_t> |
| 282 | void pack_vnni2(scalar_t* __restrict__ dst, const scalar_t* __restrict__ src, int K, int N, int ld_src, int ld_dst) { |
| 283 | pack_vnni2<scalar_t, int32_t, false>(dst, src, nullptr, K, N, ld_src, ld_dst); |
| 284 | } |
| 285 | |
| 286 | template <typename scalar_t, typename index_t> |
| 287 | void pack_vnni2( |
| 288 | scalar_t* __restrict__ dst, |
| 289 | const scalar_t* __restrict__ src, |
| 290 | const index_t* __restrict__ ind, |
| 291 | int K, |
| 292 | int N, |
| 293 | int ld_src, |
| 294 | int ld_dst) { |
| 295 | assert(ind != nullptr); |
| 296 | pack_vnni2<scalar_t, index_t, true>(dst, src, ind, K, N, ld_src, ld_dst); |
| 297 | } |
| 298 | |
| 299 | } // anonymous namespace |