1// Copyright 2020 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$assert SSE in [2, 4] 7$assert not XOP or AVX 8$assert not AVX or SSE == 4 9$assert REQUANTIZATION == "FP32" 10$assert DATATYPE in ["QC8", "QS8", "QU8"] 11$assert VARIANT in ["LD64", "LD128", "EXTENDED"] 12$assert MR <= 4 13#include <assert.h> 14 15$if XOP: 16 #if defined(__GNUC__) || defined(__clang__) 17 #include <x86intrin.h> 18 #else 19 #include <immintrin.h> 20 #include <ammintrin.h> 21 #endif 22$else: 23 $SSE_HEADER = {2: "emmintrin.h", 4: "smmintrin.h"}[SSE] 24 #include <${SSE_HEADER}> 25 26#include <xnnpack/gemm.h> 27#include <xnnpack/math.h> 28 29 30 31$LOAD_SUFFIX = {"LD128": "_ld128", "LD64": "_ld64", "EXTENDED": ""}[VARIANT] 32$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else "" 33$PARAMS_UNION = "xnn_qs8_minmax_params" if DATATYPE == "QC8" else "xnn_%s_conv_minmax_params" % DATATYPE.lower() 34$PARAMS_STRUCT = ("" if DATATYPE == "QC8" else "fp32_") + ("sse4" if SSE == 4 and DATATYPE != "QU8" else "sse2") 35$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t" 36$ISA = "xop" if XOP else "avx" if AVX else {2: "sse2", 4: "sse41"}[SSE] 37void xnn_${DATATYPE.lower()}_gemm${GEMM_SUFFIX}_minmax_fp32_ukernel_${MR}x4c2__${ISA}${LOAD_SUFFIX}( 38 size_t mr, 39 size_t nc, 40 size_t kc, 41 const ${XINT8_T}* restrict a, 42 size_t a_stride, 43 const void* restrict w, 44 ${XINT8_T}* restrict c, 45 size_t cm_stride, 46 size_t cn_stride, 47 const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 48{ 49 assert(mr != 0); 50 assert(mr <= ${MR}); 51 assert(nc != 0); 52 assert(kc != 0); 53 assert(kc % sizeof(${XINT8_T}) == 0); 54 assert(a != NULL); 55 assert(w != NULL); 56 assert(c != NULL); 57 58 kc = round_up_po2(kc, 2); 59 const ${XINT8_T}* a0 = a; 60 ${XINT8_T}* c0 = c; 61 $for M in range(1, MR): 62 const ${XINT8_T}* a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M-1} + a_stride); 63 ${XINT8_T}* c${M} = (${XINT8_T}*) ((uintptr_t) c${M-1} + cm_stride); 64 $if M % 2 == 0: 65 if XNN_UNPREDICTABLE(mr <= ${M}) { 66 a${M} = a${M-1}; 67 c${M} = c${M-1}; 68 } 69 $elif M + 1 == MR: 70 if XNN_UNPREDICTABLE(mr != ${M+1}) { 71 a${M} = a${M-1}; 72 c${M} = c${M-1}; 73 } 74 $else: 75 if XNN_UNPREDICTABLE(mr < ${M+1}) { 76 a${M} = a${M-1}; 77 c${M} = c${M-1}; 78 } 79 80 do { 81 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w); 82 $for M in range(1, MR): 83 __m128i vacc${M}x0123 = vacc0x0123; 84 w = (const void*) ((const int32_t*) w + 4); 85 86 size_t k = kc; 87 $if DATATYPE == "QU8": 88 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.kernel_zero_point); 89 $if SSE < 4 or VARIANT == "LD128": 90 const __m128i vzero = _mm_setzero_si128(); 91 while (k >= 8 * sizeof(${XINT8_T})) { 92 $for M in range(MR): 93 const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M}); 94 $if DATATYPE == "QU8": 95 $if SSE == 4: 96 const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M}); 97 $else: 98 const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero); 99 $else: 100 $if SSE == 4: 101 const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M}); 102 $else: 103 const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8); 104 a${M} += 8; 105 106 $if VARIANT == "LD128": 107 $for K in range(0, 4, 2): 108 $if K == 0: 109 const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) w); 110 $else: 111 const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((const ${XINT8_T}*) w + ${K * 8})); 112 $if DATATYPE == "QU8": 113 const __m128i vxb${K} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${K}${K+1}, vzero), vb_zero_point); 114 const __m128i vxb${K+1} = _mm_sub_epi16(_mm_unpackhi_epi8(vb${K}${K+1}, vzero), vb_zero_point); 115 $elif SSE == 4: 116 const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K}${K+1}); 117 const __m128i vxb${K+1} = _mm_srai_epi16(_mm_unpackhi_epi8(vb${K}${K+1}, vb${K}${K+1}), 8); 118 $else: 119 const __m128i vsb${K}${K+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${K}${K+1}); 120 const __m128i vxb${K} = _mm_unpacklo_epi8(vb${K}${K+1}, vsb${K}${K+1}); 121 const __m128i vxb${K+1} = _mm_unpackhi_epi8(vb${K}${K+1}, vsb${K}${K+1}); 122 123 $for M in range(MR): 124 $if XOP: 125 vacc${M}x0123 = _mm_maddd_epi16( 126 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}, vacc${M}x0123); 127 $else: 128 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 129 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K})); 130 131 $for M in range(MR): 132 $if XOP: 133 vacc${M}x0123 = _mm_maddd_epi16( 134 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K+1}, ${K+1}, ${K+1}, ${K+1})), vxb${K+1}, vacc${M}x0123); 135 $else: 136 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 137 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K+1}, ${K+1}, ${K+1}, ${K+1})), vxb${K+1})); 138 $else: 139 $for K in range(4): 140 $if VARIANT == "LD64": 141 $if K == 0: 142 const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) w); 143 $else: 144 const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${K * 8})); 145 $if DATATYPE == "QU8": 146 $if SSE == 4: 147 const __m128i vxb${K} = _mm_sub_epi16(_mm_cvtepu8_epi16(vb${K}), vb_zero_point); 148 $else: 149 const __m128i vxb${K} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${K}, vzero), vb_zero_point); 150 $else: 151 $if SSE == 4: 152 const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K}); 153 $else: 154 const __m128i vxb${K} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${K}, vb${K}), 8); 155 $elif VARIANT == "EXTENDED": 156 $if K == 0: 157 const __m128i vxb${K} = _mm_load_si128((const __m128i*) w); 158 $else: 159 const __m128i vxb${K} = _mm_load_si128((const __m128i*) ((const int16_t*) w + ${K * 8})); 160 161 $for M in range(MR): 162 $if XOP: 163 vacc${M}x0123 = _mm_maddd_epi16( 164 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}, vacc${M}x0123); 165 $else: 166 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 167 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K})); 168 169 $if VARIANT == "EXTENDED": 170 w = (const void*) ((const int16_t*) w + 32); 171 $else: 172 w = (const void*) ((const ${XINT8_T}*) w + 32); 173 k -= 8 * sizeof(${XINT8_T}); 174 } 175 if (k != 0) { 176 $for M in range(MR): 177 const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M}); 178 $if DATATYPE == "QU8": 179 $if SSE == 4: 180 const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M}); 181 $else: 182 const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero); 183 $else: 184 $if SSE == 4: 185 const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M}); 186 $else: 187 const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8); 188 a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + k); 189 190 $if VARIANT == "EXTENDED": 191 const __m128i vxb0 = _mm_load_si128((const __m128i*) w); 192 w = (const void*) ((const int16_t*) w + 8); 193 $else: 194 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); 195 $if DATATYPE == "QU8": 196 $if SSE == 4: 197 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point); 198 $else: 199 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point); 200 $else: 201 $if SSE == 4: 202 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0); 203 $else: 204 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); 205 w = (const void*) ((const ${XINT8_T}*) w + 8); 206 207 $for M in range(MR): 208 $if XOP: 209 vacc${M}x0123 = _mm_maddd_epi16( 210 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc${M}x0123); 211 $else: 212 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 213 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0)); 214 215 if (k > 2 * sizeof(${XINT8_T})) { 216 $if VARIANT == "EXTENDED": 217 const __m128i vxb1 = _mm_load_si128((const __m128i*) w); 218 w = (const void*) ((const int16_t*) w + 8); 219 $else: 220 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w); 221 $if DATATYPE == "QU8": 222 $if SSE == 4: 223 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point); 224 $else: 225 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point); 226 $else: 227 $if SSE == 4: 228 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1); 229 $else: 230 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); 231 w = (const void*) ((const ${XINT8_T}*) w + 8); 232 233 $for M in range(MR): 234 $if XOP: 235 vacc${M}x0123 = _mm_maddd_epi16( 236 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc${M}x0123); 237 $else: 238 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 239 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1)); 240 241 if (k > 4 * sizeof(${XINT8_T})) { 242 $if VARIANT == "EXTENDED": 243 const __m128i vxb2 = _mm_load_si128((const __m128i*) w); 244 w = (const void*) ((const int16_t*) w + 8); 245 $else: 246 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w); 247 $if DATATYPE == "QU8": 248 $if SSE == 4: 249 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point); 250 $else: 251 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point); 252 $else: 253 $if SSE == 4: 254 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2); 255 $else: 256 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); 257 w = (const void*) ((const ${XINT8_T}*) w + 8); 258 259 $for M in range(MR): 260 $if XOP: 261 vacc${M}x0123 = _mm_maddd_epi16( 262 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc${M}x0123); 263 $else: 264 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 265 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(2, 2, 2, 2)), vxb2)); 266 } 267 } 268 } 269 270 $for M in range(MR): 271 __m128 vscaled${M}x0123 = _mm_cvtepi32_ps(vacc${M}x0123); 272 273 $if DATATYPE == "QC8": 274 const __m128 vscale0123 = _mm_loadu_ps((const float*) w); 275 w = (const void*) ((const float*) w + 4); 276 $for M in range(MR): 277 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123); 278 $else: 279 const __m128 vscale = _mm_load_ps(params->${PARAMS_STRUCT}.scale); 280 $for M in range(MR): 281 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale); 282 283 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->${PARAMS_STRUCT}.output_max_less_zero_point); 284 $for M in range(MR): 285 vscaled${M}x0123 = _mm_min_ps(vscaled${M}x0123, voutput_max_less_zero_point); 286 287 $for M in range(MR): 288 vacc${M}x0123 = _mm_cvtps_epi32(vscaled${M}x0123); 289 290 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point); 291 $for M in range(0, MR, 2): 292 __m128i vacc${M}${min(M+1, MR-1)}x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123), voutput_zero_point); 293 294 $if DATATYPE == "QU8": 295 $if MR > 2: 296 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 297 $else: 298 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 299 300 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 301 $else: 302 $if SSE < 4: 303 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min); 304 $for M in range(0, MR, 2): 305 vacc${M}${min(M+1, MR-1)}x0123 = _mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min); 306 307 $if MR > 2: 308 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 309 $else: 310 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 311 312 $if SSE == 4: 313 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 314 315 if (nc >= 4) { 316 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout); 317 $for M in range(1, MR): 318 $if SSE == 4: 319 *((uint32_t*) c${M}) = (uint32_t) _mm_extract_epi32(vout, ${M}); 320 $else: 321 vout = _mm_srli_si128(vout, 4); 322 *((uint32_t*) c${M}) = (uint32_t) _mm_cvtsi128_si32(vout); 323 324 $for M in range(MR): 325 c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride); 326 327 $for M in range(MR): 328 a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc); 329 330 nc -= 4; 331 } else { 332 if (nc & 2) { 333 $for M in range(MR): 334 *((uint16_t*) c${M}) = (uint16_t) _mm_extract_epi16(vout, ${M * 2}); 335 c${M} += 2; 336 vout = _mm_srli_epi32(vout, 16); 337 } 338 if (nc & 1) { 339 $if SSE == 4: 340 $for M in range(MR): 341 *c${M} = (${XINT8_T}) _mm_extract_epi8(vout, ${M * 4}); 342 $else: 343 *c0 = (${XINT8_T}) _mm_cvtsi128_si32(vout); 344 $for M in range(1, MR): 345 *c${M} = (${XINT8_T}) _mm_extract_epi16(vout, ${M * 2}); 346 } 347 348 nc = 0; 349 } 350 } while (nc != 0); 351} 352