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$assert SSE in [2, 3, 4] 7$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"] 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 $SSE_HEADER = {2: "emmintrin.h", 3: "tmmintrin.h", 4: "smmintrin.h"}[SSE] 24 #include <${SSE_HEADER}> 25 26#include <xnnpack/igemm.h> 27#include <xnnpack/math.h> 28#include <xnnpack/unaligned.h> 29 30 31$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("sse4" if SSE == 4 and DATATYPE != "QU8" else "sse2") 32$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower() 33$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t" 34$ISA = "xop" if XOP else "avx" if AVX else {2: "sse2", 3: "ssse3", 4: "sse41"}[SSE] 35void xnn_${DATATYPE.lower()}_igemm_minmax_fp32_ukernel_${MR}x4c8__${ISA}_${VARIANT.lower()}( 36 size_t mr, 37 size_t nc, 38 size_t kc, 39 size_t ks, 40 const ${XINT8_T}** restrict a, 41 const void* restrict w, 42 ${XINT8_T}* restrict c, 43 size_t cm_stride, 44 size_t cn_stride, 45 size_t a_offset, 46 const ${XINT8_T}* zero, 47 const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 48{ 49 assert(mr != 0); 50 assert(mr <= ${MR}); 51 assert(nc != 0); 52 assert(kc != 0); 53 assert(ks != 0); 54 assert(ks % (${MR} * sizeof(void*)) == 0); 55 assert(a_offset % sizeof(${XINT8_T}) == 0); 56 assert(a != NULL); 57 assert(w != NULL); 58 assert(c != NULL); 59 60 kc = round_up_po2(kc, 8); 61 ${XINT8_T}* c0 = c; 62 $for M in range(1, MR): 63 ${XINT8_T}* c${M} = (${XINT8_T}*) ((uintptr_t) c${M-1} + cm_stride); 64 $if M % 2 == 0: 65 if XNN_UNPREDICTABLE(mr <= ${M}) { 66 c${M} = c${M-1}; 67 } 68 $elif M + 1 == MR: 69 if XNN_UNPREDICTABLE(mr != ${M+1}) { 70 c${M} = c${M-1}; 71 } 72 $else: 73 if XNN_UNPREDICTABLE(mr < ${M+1}) { 74 c${M} = c${M-1}; 75 } 76 77 do { 78 $for N in range(4): 79 __m128i vacc0x${N} = _mm_cvtsi32_si128(((const int*) w)[${N}]); 80 $for M in range(1, MR): 81 $for N in range(4): 82 __m128i vacc${M}x${N} = vacc0x${N}; 83 w = (const int32_t*) w + 4; 84 85 size_t p = ks; 86 do { 87 $for M in range(MR): 88 const ${XINT8_T}* restrict a${M} = a[${M}]; 89 if XNN_UNPREDICTABLE(a${M} != zero) { 90 a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + a_offset); 91 } 92 a += ${MR}; 93 94 size_t k = 0; 95 $if DATATYPE == "QU8": 96 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.kernel_zero_point); 97 $if SSE < 4 or VARIANT == "LD128": 98 const __m128i vzero = _mm_setzero_si128(); 99 while (k < kc) { 100 $for M in range(MR): 101 const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M}); 102 $if DATATYPE == "QU8": 103 $if SSE == 4: 104 const __m128i vxa${M} = _mm_cvtepu8_epi16(va${M}); 105 $else: 106 const __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero); 107 $else: 108 $if SSE == 4: 109 const __m128i vxa${M} = _mm_cvtepi8_epi16(va${M}); 110 $else: 111 const __m128i vxa${M} = _mm_srai_epi16(_mm_unpacklo_epi8(va${M}, va${M}), 8); 112 a${M} += 8; 113 114 $if VARIANT == "LD128": 115 $for N in range(0, 4, 2): 116 $if N == 0: 117 const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) w); 118 $else: 119 const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8})); 120 $if DATATYPE == "QU8": 121 const __m128i vxb${N} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${N}${N+1}, vzero), vb_zero_point); 122 const __m128i vxb${N+1} = _mm_sub_epi16(_mm_unpackhi_epi8(vb${N}${N+1}, vzero), vb_zero_point); 123 $elif SSE == 4: 124 const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N}${N+1}); 125 const __m128i vxb${N+1} = _mm_srai_epi16(_mm_unpackhi_epi8(vb${N}${N+1}, vb${N}${N+1}), 8); 126 $else: 127 const __m128i vsb${N}${N+1} = _mm_cmpgt_epi8(_mm_setzero_si128(), vb${N}${N+1}); 128 const __m128i vxb${N} = _mm_unpacklo_epi8(vb${N}${N+1}, vsb${N}${N+1}); 129 const __m128i vxb${N+1} = _mm_unpackhi_epi8(vb${N}${N+1}, vsb${N}${N+1}); 130 131 $for M in range(MR): 132 $if XOP: 133 vacc${M}x${N} = _mm_maddd_epi16(vxa${M}, vxb${N}, vacc${M}x${N}); 134 vacc${M}x${N+1} = _mm_maddd_epi16(vxa${M}, vxb${N+1}, vacc${M}x${N+1}); 135 $else: 136 vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N})); 137 vacc${M}x${N+1} = _mm_add_epi32(vacc${M}x${N+1}, _mm_madd_epi16(vxa${M}, vxb${N+1})); 138 $else: 139 $for N in range(4): 140 $if N == 0: 141 const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) w); 142 $else: 143 const __m128i vb${N} = _mm_loadl_epi64((const __m128i*) ((const ${XINT8_T}*) w + ${N * 8})); 144 $if DATATYPE == "QU8": 145 $if SSE == 4: 146 const __m128i vxb${N} = _mm_sub_epi16(_mm_cvtepu8_epi16(vb${N}), vb_zero_point); 147 $else: 148 const __m128i vxb${N} = _mm_sub_epi16(_mm_unpacklo_epi8(vb${N}, vzero), vb_zero_point); 149 $else: 150 $if SSE == 4: 151 const __m128i vxb${N} = _mm_cvtepi8_epi16(vb${N}); 152 $else: 153 const __m128i vxb${N} = _mm_srai_epi16(_mm_unpacklo_epi8(vb${N}, vb${N}), 8); 154 155 $for M in range(MR): 156 $if XOP: 157 vacc${M}x${N} = _mm_maddd_epi16(vxa${M}, vxb${N}, vacc${M}x${N}); 158 $else: 159 vacc${M}x${N} = _mm_add_epi32(vacc${M}x${N}, _mm_madd_epi16(vxa${M}, vxb${N})); 160 161 w = (const void*) ((const ${XINT8_T}*) w + 32); 162 k += 8 * sizeof(${XINT8_T}); 163 } 164 p -= ${MR} * sizeof(void*); 165 } while (p != 0); 166 167 $if SSE >= 3: 168 $for M in range(MR): 169 const __m128i vacc${M}x01 = _mm_hadd_epi32(vacc${M}x0, vacc${M}x1); 170 const __m128i vacc${M}x23 = _mm_hadd_epi32(vacc${M}x2, vacc${M}x3); 171 172 $for M in range(MR): 173 __m128i vacc${M}x0123 = _mm_hadd_epi32(vacc${M}x01, vacc${M}x23); 174 $else: 175 $for M in range(MR): 176 const __m128i vacc${M}x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc${M}x0, vacc${M}x2), _mm_unpackhi_epi32(vacc${M}x0, vacc${M}x2)); 177 const __m128i vacc${M}x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc${M}x1, vacc${M}x3), _mm_unpackhi_epi32(vacc${M}x1, vacc${M}x3)); 178 179 $for M in range(MR): 180 __m128i vacc${M}x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc${M}x02, vacc${M}x13), _mm_unpackhi_epi32(vacc${M}x02, vacc${M}x13)); 181 182 $for M in range(MR): 183 __m128 vscaled${M}x0123 = _mm_cvtepi32_ps(vacc${M}x0123); 184 185 $if DATATYPE == "QC8": 186 const __m128 vscale0123 = _mm_load_ps((const float*) w); 187 w = (const void*) ((const float*) w + 4); 188 $for M in range(MR): 189 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123); 190 $else: 191 const __m128 vscale = _mm_load_ps(params->${PARAMS_STRUCT}.scale); 192 $for M in range(MR): 193 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale); 194 195 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->${PARAMS_STRUCT}.output_max_less_zero_point); 196 $for M in range(MR): 197 vscaled${M}x0123 = _mm_min_ps(vscaled${M}x0123, voutput_max_less_zero_point); 198 199 $for M in range(MR): 200 vacc${M}x0123 = _mm_cvtps_epi32(vscaled${M}x0123); 201 202 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point); 203 $for M in range(0, MR, 2): 204 __m128i vacc${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); 205 206 $if DATATYPE == "QU8": 207 $if MR > 2: 208 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 209 $else: 210 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 211 212 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 213 $else: 214 $if SSE < 4: 215 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min); 216 $for M in range(0, MR, 2): 217 vacc${M}${min(M+1, MR-1)}x0123 = _mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min); 218 219 $if MR > 2: 220 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 221 $else: 222 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 223 224 $if SSE == 4: 225 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 226 227 if (nc >= 4) { 228 $for M in reversed(range(1, MR)): 229 $if SSE == 4: 230 unaligned_store_u32(c${M}, (uint32_t) _mm_extract_epi32(vout, ${M})); 231 $else: 232 unaligned_store_u32(c${M}, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(${M}, ${M}, ${M}, ${M})))); 233 c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride); 234 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); 235 c0 = (${XINT8_T}*) ((uintptr_t) c0 + cn_stride); 236 237 a = (const ${XINT8_T}**restrict) ((uintptr_t) a - ks); 238 239 nc -= 4; 240 } else { 241 if (nc & 2) { 242 $for M in reversed(range(MR)): 243 unaligned_store_u16(c${M}, (uint16_t) _mm_extract_epi16(vout, ${M * 2})); 244 c${M} += 2; 245 vout = _mm_srli_epi32(vout, 16); 246 } 247 if (nc & 1) { 248 $if SSE == 4: 249 $for M in reversed(range(MR)): 250 *c${M} = (${XINT8_T}) _mm_extract_epi8(vout, ${M * 4}); 251 $else: 252 $for M in reversed(range(1, MR)): 253 *c${M} = (${XINT8_T}) _mm_extract_epi16(vout, ${M * 2}); 254 *c0 = (${XINT8_T}) _mm_cvtsi128_si32(vout); 255 } 256 257 nc = 0; 258 } 259 } while (nc != 0); 260} 261