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#include <xnnpack/unaligned.h> 29 30 31 32$LOAD_SUFFIX = {"LD128": "_ld128", "LD64": "_ld64", "EXTENDED": ""}[VARIANT] 33$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else "" 34$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("sse4" if SSE == 4 and DATATYPE != "QU8" else "sse2") 35$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower() 36$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t" 37$ISA = "xop" if XOP else "avx" if AVX else {2: "sse2", 4: "sse41"}[SSE] 38void xnn_${DATATYPE.lower()}_gemm${GEMM_SUFFIX}_minmax_fp32_ukernel_${MR}x4c2__${ISA}${LOAD_SUFFIX}( 39 size_t mr, 40 size_t nc, 41 size_t kc, 42 const ${XINT8_T}* restrict a, 43 size_t a_stride, 44 const void* restrict w, 45 ${XINT8_T}* restrict c, 46 size_t cm_stride, 47 size_t cn_stride, 48 const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 49{ 50 assert(mr != 0); 51 assert(mr <= ${MR}); 52 assert(nc != 0); 53 assert(kc != 0); 54 assert(kc % sizeof(${XINT8_T}) == 0); 55 assert(a != NULL); 56 assert(w != NULL); 57 assert(c != NULL); 58 59 kc = round_up_po2(kc, 2 * sizeof(${XINT8_T})); 60 const ${XINT8_T}* a0 = a; 61 ${XINT8_T}* c0 = c; 62 $for M in range(1, MR): 63 const ${XINT8_T}* a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M-1} + a_stride); 64 ${XINT8_T}* c${M} = (${XINT8_T}*) ((uintptr_t) c${M-1} + cm_stride); 65 $if M % 2 == 0: 66 if XNN_UNPREDICTABLE(mr <= ${M}) { 67 a${M} = a${M-1}; 68 c${M} = c${M-1}; 69 } 70 $elif M + 1 == MR: 71 if XNN_UNPREDICTABLE(mr != ${M+1}) { 72 a${M} = a${M-1}; 73 c${M} = c${M-1}; 74 } 75 $else: 76 if XNN_UNPREDICTABLE(mr < ${M+1}) { 77 a${M} = a${M-1}; 78 c${M} = c${M-1}; 79 } 80 81 do { 82 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w); 83 $for M in range(1, MR): 84 __m128i vacc${M}x0123 = vacc0x0123; 85 w = (const void*) ((const int32_t*) w + 4); 86 87 size_t k = kc; 88 $if DATATYPE == "QU8": 89 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.kernel_zero_point); 90 $if SSE < 4 or VARIANT == "LD128": 91 const __m128i vzero = _mm_setzero_si128(); 92 while (k >= 8 * sizeof(${XINT8_T})) { 93 $for M in range(MR): 94 const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M}); 95 $if DATATYPE == "QU8": 96 $if SSE == 4: 97 const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M}); 98 $else: 99 const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero); 100 $else: 101 $if SSE == 4: 102 const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M}); 103 $else: 104 const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8); 105 a${M} += 8; 106 107 $if VARIANT == "LD128": 108 $for K in range(0, 4, 2): 109 $if K == 0: 110 const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) w); 111 $else: 112 const __m128i vb${K}${K+1} = _mm_loadu_si128((const __m128i*) ((const ${XINT8_T}*) w + ${K * 8})); 113 $if DATATYPE == "QU8": 114 const __m128i vxb${K} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${K}${K+1}, vzero), vb_zero_point); 115 const __m128i vxb${K+1} = _mm_sub_epi16(_mm_unpackhi_epi8(vb${K}${K+1}, vzero), vb_zero_point); 116 $elif SSE == 4: 117 const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K}${K+1}); 118 const __m128i vxb${K+1} = _mm_srai_epi16(_mm_unpackhi_epi8(vb${K}${K+1}, vb${K}${K+1}), 8); 119 $else: 120 const __m128i vsb${K}${K+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${K}${K+1}); 121 const __m128i vxb${K} = _mm_unpacklo_epi8(vb${K}${K+1}, vsb${K}${K+1}); 122 const __m128i vxb${K+1} = _mm_unpackhi_epi8(vb${K}${K+1}, vsb${K}${K+1}); 123 124 $for M in range(MR): 125 $if XOP: 126 vacc${M}x0123 = _mm_maddd_epi16( 127 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}, vacc${M}x0123); 128 $else: 129 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 130 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K})); 131 132 $for M in range(MR): 133 $if XOP: 134 vacc${M}x0123 = _mm_maddd_epi16( 135 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K+1}, ${K+1}, ${K+1}, ${K+1})), vxb${K+1}, vacc${M}x0123); 136 $else: 137 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 138 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K+1}, ${K+1}, ${K+1}, ${K+1})), vxb${K+1})); 139 $else: 140 $for K in range(4): 141 $if VARIANT == "LD64": 142 $if K == 0: 143 const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) w); 144 $else: 145 const __m128i vb${K} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${K * 8})); 146 $if DATATYPE == "QU8": 147 $if SSE == 4: 148 const __m128i vxb${K} = _mm_sub_epi16(_mm_cvtepu8_epi16(vb${K}), vb_zero_point); 149 $else: 150 const __m128i vxb${K} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${K}, vzero), vb_zero_point); 151 $else: 152 $if SSE == 4: 153 const __m128i vxb${K} = _mm_cvtepi8_epi16(vb${K}); 154 $else: 155 const __m128i vxb${K} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${K}, vb${K}), 8); 156 $elif VARIANT == "EXTENDED": 157 $if K == 0: 158 const __m128i vxb${K} = _mm_load_si128((const __m128i*) w); 159 $else: 160 const __m128i vxb${K} = _mm_load_si128((const __m128i*) ((const int16_t*) w + ${K * 8})); 161 162 $for M in range(MR): 163 $if XOP: 164 vacc${M}x0123 = _mm_maddd_epi16( 165 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K}, vacc${M}x0123); 166 $else: 167 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 168 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(${K}, ${K}, ${K}, ${K})), vxb${K})); 169 170 $if VARIANT == "EXTENDED": 171 w = (const void*) ((const int16_t*) w + 32); 172 $else: 173 w = (const void*) ((const ${XINT8_T}*) w + 32); 174 k -= 8 * sizeof(${XINT8_T}); 175 } 176 if (k != 0) { 177 $for M in range(MR): 178 const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M}); 179 $if DATATYPE == "QU8": 180 $if SSE == 4: 181 const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M}); 182 $else: 183 const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero); 184 $else: 185 $if SSE == 4: 186 const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M}); 187 $else: 188 const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8); 189 a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + k); 190 191 $if VARIANT == "EXTENDED": 192 const __m128i vxb0 = _mm_load_si128((const __m128i*) w); 193 w = (const void*) ((const int16_t*) w + 8); 194 $else: 195 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); 196 $if DATATYPE == "QU8": 197 $if SSE == 4: 198 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point); 199 $else: 200 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point); 201 $else: 202 $if SSE == 4: 203 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0); 204 $else: 205 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); 206 w = (const void*) ((const ${XINT8_T}*) w + 8); 207 208 $for M in range(MR): 209 $if XOP: 210 vacc${M}x0123 = _mm_maddd_epi16( 211 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc${M}x0123); 212 $else: 213 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 214 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 0, 0, 0)), vxb0)); 215 216 if (k > 2 * sizeof(${XINT8_T})) { 217 $if VARIANT == "EXTENDED": 218 const __m128i vxb1 = _mm_load_si128((const __m128i*) w); 219 w = (const void*) ((const int16_t*) w + 8); 220 $else: 221 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w); 222 $if DATATYPE == "QU8": 223 $if SSE == 4: 224 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point); 225 $else: 226 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point); 227 $else: 228 $if SSE == 4: 229 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1); 230 $else: 231 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); 232 w = (const void*) ((const ${XINT8_T}*) w + 8); 233 234 $for M in range(MR): 235 $if XOP: 236 vacc${M}x0123 = _mm_maddd_epi16( 237 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc${M}x0123); 238 $else: 239 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 240 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(1, 1, 1, 1)), vxb1)); 241 242 if (k > 4 * sizeof(${XINT8_T})) { 243 $if VARIANT == "EXTENDED": 244 const __m128i vxb2 = _mm_load_si128((const __m128i*) w); 245 w = (const void*) ((const int16_t*) w + 8); 246 $else: 247 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w); 248 $if DATATYPE == "QU8": 249 $if SSE == 4: 250 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point); 251 $else: 252 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point); 253 $else: 254 $if SSE == 4: 255 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2); 256 $else: 257 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); 258 w = (const void*) ((const ${XINT8_T}*) w + 8); 259 260 $for M in range(MR): 261 $if XOP: 262 vacc${M}x0123 = _mm_maddd_epi16( 263 _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc${M}x0123); 264 $else: 265 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, 266 _mm_madd_epi16(_mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(2, 2, 2, 2)), vxb2)); 267 } 268 } 269 } 270 271 $for M in range(MR): 272 __m128 vscaled${M}x0123 = _mm_cvtepi32_ps(vacc${M}x0123); 273 274 $if DATATYPE == "QC8": 275 const __m128 vscale0123 = _mm_loadu_ps((const float*) w); 276 w = (const void*) ((const float*) w + 4); 277 $for M in range(MR): 278 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123); 279 $else: 280 const __m128 vscale = _mm_load_ps(params->${PARAMS_STRUCT}.scale); 281 $for M in range(MR): 282 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale); 283 284 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->${PARAMS_STRUCT}.output_max_less_zero_point); 285 $for M in range(MR): 286 vscaled${M}x0123 = _mm_min_ps(vscaled${M}x0123, voutput_max_less_zero_point); 287 288 $for M in range(MR): 289 vacc${M}x0123 = _mm_cvtps_epi32(vscaled${M}x0123); 290 291 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point); 292 $for M in range(0, MR, 2): 293 __m128i vacc${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); 294 295 $if DATATYPE == "QU8": 296 $if MR > 2: 297 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 298 $else: 299 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 300 301 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 302 $else: 303 $if SSE < 4: 304 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min); 305 $for M in range(0, MR, 2): 306 vacc${M}${min(M+1, MR-1)}x0123 = _mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min); 307 308 $if MR > 2: 309 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 310 $else: 311 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 312 313 $if SSE == 4: 314 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 315 316 if (nc >= 4) { 317 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); 318 $for M in range(1, MR): 319 $if SSE == 4: 320 unaligned_store_u32(c${M}, (uint32_t) _mm_extract_epi32(vout, ${M})); 321 $else: 322 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1)); 323 unaligned_store_u32(c${M}, (uint32_t) _mm_cvtsi128_si32(vout)); 324 325 $for M in range(MR): 326 c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride); 327 328 $for M in range(MR): 329 a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc); 330 331 nc -= 4; 332 } else { 333 if (nc & 2) { 334 $for M in range(MR): 335 unaligned_store_u16(c${M}, (uint16_t) _mm_extract_epi16(vout, ${M * 2})); 336 c${M} += 2; 337 vout = _mm_srli_epi32(vout, 16); 338 } 339 if (nc & 1) { 340 $if SSE == 4: 341 $for M in range(MR): 342 *c${M} = (${XINT8_T}) _mm_extract_epi8(vout, ${M * 4}); 343 $else: 344 *c0 = (${XINT8_T}) _mm_cvtsi128_si32(vout); 345 $for M in range(1, MR): 346 *c${M} = (${XINT8_T}) _mm_extract_epi16(vout, ${M * 2}); 347 } 348 349 nc = 0; 350 } 351 } while (nc != 0); 352} 353