File size: 8,587 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 |
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert DATATYPE in ["F32", "QC4", "QC8"]
$assert NR % 8 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <immintrin.h>
$if DATATYPE in ["QC8", "QC4"]:
#include <smmintrin.h>
#include <xnnpack/gemm.h>
$if DATATYPE in ["QC8", "QC4"]:
$ISA = "avx2"
$else:
$ISA = {0: "avx", 3: "fma3"}[FMA]
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}s4__${ISA}_broadcast(
size_t mr,
size_t nc,
size_t kc,
const float* restrict a,
size_t a_stride,
$if DATATYPE == "F32":
const float* restrict w,
$else:
const void* restrict w,
float* restrict c,
size_t cm_stride,
size_t cn_stride,
$if INC:
const float* restrict acc,
$if DATATYPE == "QC4":
const union xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
$else:
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(mr != 0);
assert(mr <= ${MR});
assert(nc != 0);
assert(kc != 0);
assert(kc % sizeof(float) == 0);
assert(a != NULL);
assert(w != NULL);
assert(c != NULL);
$if INC:
assert(acc != NULL);
const float* a0 = a;
float* c0 = c;
$for M in range(1, MR):
const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
$if M % 2 == 0:
if XNN_UNPREDICTABLE(mr <= ${M}) {
a${M} = a${M-1};
c${M} = c${M-1};
}
$elif M + 1 == MR:
if XNN_UNPREDICTABLE(mr != ${M+1}) {
a${M} = a${M-1};
c${M} = c${M-1};
}
$else:
if XNN_UNPREDICTABLE(mr < ${M+1}) {
a${M} = a${M-1};
c${M} = c${M-1};
}
$if DATATYPE == "QC4":
const __m128i vminus_kernel_zero_point = _mm_load_si128((const __m128i *) params->sse.minus_kernel_zero_point);
const __m128i vmask = _mm_load_si128((const __m128i *) params->sse.mask);
do {
$if INC:
$for M in range(MR):
$for N in range(0, NR, 8):
__m256 vacc${M}x${ABC[N:N+8]} = _mm256_load_ps(acc + ${M*NR+N});
acc += ${MR*NR};
$else:
$for N in range(0, NR, 8):
$if DATATYPE == "F32":
__m256 vacc0x${ABC[N:N+8]} = _mm256_load_ps(w + ${N});
$else:
__m256 vacc0x${ABC[N:N+8]} = _mm256_loadu_ps((const float*) w + ${N});
$for M in range(1, MR):
$for N in range(0, NR, 8):
__m256 vacc${M}x${ABC[N:N+8]} = vacc0x${ABC[N:N+8]};
$if DATATYPE == "F32":
w += ${NR};
$else:
w = (const float*) w + ${NR};
size_t k = kc;
while (k >= 4 * sizeof(float)) {
$for M in range(MR):
__m256 va${M} = _mm256_broadcast_ps((const __m128*) a${M});
a${M} += 4;
$for L in range(4):
$if DATATYPE == "F32":
$for N in range(0, NR, 8):
const __m256 vb${ABC[N:N+8]}c${L} = _mm256_load_ps(w + ${L * NR + N});
$else:
$for N in range(0, NR, 8):
const __m256i vbi${ABC[N:N+8]}c${L} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const void*) ((const int8_t*) w + ${L * NR + N})));
$for N in range(0, NR, 8):
const __m256 vb${ABC[N:N+8]}c${L} = _mm256_cvtepi32_ps(vbi${ABC[N:N+8]}c${L});
$for N in range(0, NR, 8):
$for M in range(MR):
$if FMA == 3:
vacc${M}x${ABC[N:N+8]} = _mm256_fmadd_ps(va${M}, vb${ABC[N:N+8]}c${L}, vacc${M}x${ABC[N:N+8]});
$else:
vacc${M}x${ABC[N:N+8]} = _mm256_add_ps(vacc${M}x${ABC[N:N+8]}, _mm256_mul_ps(va${M}, vb${ABC[N:N+8]}c${L}));
$if L + 1 != 4:
$for M in range(MR):
va${M} = _mm256_permute_ps(va${M}, _MM_SHUFFLE(0, 3, 2, 1));
$if DATATYPE == "F32":
w += ${NR * 4};
$elif DATATYPE == "QC4":
w = (const int8_t*) w + ${NR * 4 // 2};
$else:
w = (const int8_t*) w + ${NR * 4};
k -= 4 * sizeof(float);
}
if XNN_UNLIKELY(k != 0) {
$for M in range(MR):
__m256 va${M} = _mm256_broadcast_ps((const __m128*) a${M});
a${M} = (const float*) ((uintptr_t) a${M} + k);
const __m256 vzero = _mm256_setzero_ps();
$for L in range(4):
$if DATATYPE == "F32":
$for N in range(0, NR, 8):
const __m256 vb${ABC[N:N+8]}c${L} = _mm256_load_ps(w + ${L * NR + N});
$else:
$for N in range(0, NR, 8):
const __m256i vbi${ABC[N:N+8]}c${L} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const void*) ((const int8_t*) w + ${L * NR + N})));
$for N in range(0, NR, 8):
const __m256 vb${ABC[N:N+8]}c${L} = _mm256_cvtepi32_ps(vbi${ABC[N:N+8]}c${L});
$for N in range(0, NR, 8):
$for M in range(MR):
$if FMA == 3:
vacc${M}x${ABC[N:N+8]} = _mm256_fmadd_ps(_mm256_and_ps(va${M}, _mm256_cmp_ps(vb${ABC[N:N+8]}c${L}, vzero, _CMP_NEQ_OQ)), vb${ABC[N:N+8]}c${L}, vacc${M}x${ABC[N:N+8]});
$else:
vacc${M}x${ABC[N:N+8]} = _mm256_add_ps(vacc${M}x${ABC[N:N+8]}, _mm256_mul_ps(_mm256_and_ps(va${M}, _mm256_cmp_ps(vb${ABC[N:N+8]}c${L}, vzero, _CMP_NEQ_OQ)), vb${ABC[N:N+8]}c${L}));
$if L + 1 != 4:
$for M in range(MR):
va${M} = _mm256_permute_ps(va${M}, _MM_SHUFFLE(0, 3, 2, 1));
$if DATATYPE == "F32":
w += ${NR * 4};
$elif DATATYPE == "QC4":
w = (const int8_t*) w + ${NR * 4 // 2};
$else:
w = (const int8_t*) w + ${NR * 4};
}
$if DATATYPE in ["QC8", "QC4"]:
$for N in range(0, NR, 8):
const __m256 vscale${ABC[N:N+8]} = _mm256_loadu_ps((const float*) w + ${N});
$for M in range(MR):
vacc${M}x${ABC[N:N+8]} = _mm256_mul_ps(vacc${M}x${ABC[N:N+8]}, vscale${ABC[N:N+8]});
w = (const float*) w + ${NR};
const __m256 vmin = _mm256_load_ps(params->avx.min);
$for N in range(0, NR, 8):
$for M in range(MR):
vacc${M}x${ABC[N:N+8]} = _mm256_max_ps(vmin, vacc${M}x${ABC[N:N+8]});
const __m256 vmax = _mm256_load_ps(params->avx.max);
$for N in range(0, NR, 8):
$for M in range(MR):
vacc${M}x${ABC[N:N+8]} = _mm256_min_ps(vmax, vacc${M}x${ABC[N:N+8]});
if XNN_LIKELY(nc >= ${NR}) {
$for M in reversed(range(MR)):
_mm256_storeu_ps(c${M}, vacc${M}x${ABC[0:8]});
$for N in range(8, NR, 8):
_mm256_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+8]});
c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
$for M in reversed(range(MR)):
a${M} = (const float*) ((uintptr_t) a${M} - kc);
nc -= ${NR};
} else {
$for LOG2N in reversed(range(NR.bit_length())):
$if NR != 1 << LOG2N:
if (nc & ${1 << LOG2N}) {
$if LOG2N >= 3:
$for M in reversed(range(MR)):
_mm256_storeu_ps(c${M}, vacc${M}x${ABC[0:8]});
$for N in range(8, 1 << LOG2N, 8):
_mm256_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+8]});
$for M in reversed(range(MR)):
$for N in range(0, NR - (1 << LOG2N), 8):
vacc${M}x${ABC[N:N+8]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+8]};
$for M in reversed(range(MR)):
c${M} += ${1 << LOG2N};
$elif LOG2N == 2:
$for M in reversed(range(MR)):
_mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]});
$for M in reversed(range(MR)):
vacc${M}x${ABC[0:4]} = _mm256_extractf128_ps(vacc${M}x${ABC[0:8]}, 1);
$for M in reversed(range(MR)):
c${M} += 4;
$elif LOG2N == 1:
$for M in reversed(range(MR)):
_mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]});
$for M in reversed(range(MR)):
vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]});
$for M in reversed(range(MR)):
c${M} += 2;
$elif LOG2N == 0:
$for M in reversed(range(MR)):
_mm_store_ss(c${M}, vacc${M}x${ABC[0:4]});
}
$if LOG2N == 3:
$for M in reversed(range(MR)):
__m128 vacc${M}x${ABC[0:4]} = _mm256_castps256_ps128(vacc${M}x${ABC[0:8]});
nc = 0;
}
} while (nc != 0);
}
|