1// Copyright 2022 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_${MR}x4c2s4__${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, 8 * 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 do { 93 $for M in range(MR): 94 const __m128i va${M} = _mm_loadl_epi64((const __m128i*) a${M}); 95 $if DATATYPE == "QU8": 96 $if SSE == 4: 97 __m128i vxa${M} = _mm_cvtepu8_epi16(va${M}); 98 $else: 99 __m128i vxa${M} = _mm_unpacklo_epi8(va${M}, vzero); 100 $else: 101 $if SSE == 4: 102 __m128i vxa${M} = _mm_cvtepi8_epi16(va${M}); 103 $else: 104 __m128i vxa${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(vxa${M}, vxb${K}, vacc${M}x0123); 127 $else: 128 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, _mm_madd_epi16(vxa${M}, vxb${K})); 129 vxa${M} = _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 3, 2, 1)); 130 131 $for M in range(MR): 132 $if XOP: 133 vacc${M}x0123 = _mm_maddd_epi16(vxa${M}, vxb${K+1}, vacc${M}x0123); 134 $else: 135 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, _mm_madd_epi16(vxa${M}, vxb${K+1})); 136 $if K + 2 != 4: 137 vxa${M} = _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 3, 2, 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(vxa${M}, vxb${K}, vacc${M}x0123); 164 $else: 165 vacc${M}x0123 = _mm_add_epi32(vacc${M}x0123, _mm_madd_epi16(vxa${M}, vxb${K})); 166 $if K + 1 != 4: 167 vxa${M} = _mm_shuffle_epi32(vxa${M}, _MM_SHUFFLE(0, 3, 2, 1)); 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 } while (k != 0); 175 176 $for M in range(MR): 177 __m128 vscaled${M}x0123 = _mm_cvtepi32_ps(vacc${M}x0123); 178 179 $if DATATYPE == "QC8": 180 const __m128 vscale0123 = _mm_loadu_ps((const float*) w); 181 w = (const void*) ((const float*) w + 4); 182 $for M in range(MR): 183 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale0123); 184 $else: 185 const __m128 vscale = _mm_load_ps(params->${PARAMS_STRUCT}.scale); 186 $for M in range(MR): 187 vscaled${M}x0123 = _mm_mul_ps(vscaled${M}x0123, vscale); 188 189 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->${PARAMS_STRUCT}.output_max_less_zero_point); 190 $for M in range(MR): 191 vscaled${M}x0123 = _mm_min_ps(vscaled${M}x0123, voutput_max_less_zero_point); 192 193 $for M in range(MR): 194 vacc${M}x0123 = _mm_cvtps_epi32(vscaled${M}x0123); 195 196 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point); 197 $for M in range(0, MR, 2): 198 __m128i vacc${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); 199 200 $if DATATYPE == "QU8": 201 $if MR > 2: 202 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 203 $else: 204 __m128i vout = _mm_packus_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 205 206 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 207 $else: 208 $if SSE < 4: 209 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min); 210 $for M in range(0, MR, 2): 211 vacc${M}${min(M+1, MR-1)}x0123 = _mm_max_epi16(vacc${M}${min(M+1, MR-1)}x0123, voutput_min); 212 213 $if MR > 2: 214 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123); 215 $else: 216 __m128i vout = _mm_packs_epi16(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123); 217 218 $if SSE == 4: 219 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min)); 220 221 if (nc >= 4) { 222 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); 223 $for M in range(1, MR): 224 $if SSE == 4: 225 unaligned_store_u32(c${M}, (uint32_t) _mm_extract_epi32(vout, ${M})); 226 $else: 227 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1)); 228 unaligned_store_u32(c${M}, (uint32_t) _mm_cvtsi128_si32(vout)); 229 230 $for M in range(MR): 231 c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride); 232 233 $for M in range(MR): 234 a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc); 235 236 nc -= 4; 237 } else { 238 if (nc & 2) { 239 $for M in range(MR): 240 unaligned_store_u16(c${M}, (uint16_t) _mm_extract_epi16(vout, ${M * 2})); 241 c${M} += 2; 242 vout = _mm_srli_epi32(vout, 16); 243 } 244 if (nc & 1) { 245 $if SSE == 4: 246 $for M in range(MR): 247 *c${M} = (${XINT8_T}) _mm_extract_epi8(vout, ${M * 4}); 248 $else: 249 *c0 = (${XINT8_T}) _mm_cvtsi128_si32(vout); 250 $for M in range(1, MR): 251 *c${M} = (${XINT8_T}) _mm_extract_epi16(vout, ${M * 2}); 252 } 253 254 nc = 0; 255 } 256 } while (nc != 0); 257} 258