File size: 6,195 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 |
// 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 % 16 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <immintrin.h>
$if DATATYPE in ["QC8", "QC4"]:
#include <smmintrin.h>
#include <xnnpack/gemm.h>
#include <xnnpack/intrinsics-polyfill.h>
$ISA = {"F32": "avx512f", "QC8": "avx512skx", "QC4": "avx512skx"}[DATATYPE]
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${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)])
$else:
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
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):
$for N in range(0, NR, 16):
__m512 vacc${M}x${ABC[N:N+16]} = _mm512_load_ps(acc + ${M*NR+N});
acc += ${MR*NR};
$else:
$if DATATYPE == "F32":
__m512 vacc0x${ABC[0:16]} = _mm512_load_ps(w);
$else:
__m512 vacc0x${ABC[0:16]} = _mm512_loadu_ps(w);
$for N in range(16, NR, 16):
$if DATATYPE == "F32":
__m512 vacc0x${ABC[N:N+16]} = _mm512_load_ps(w + ${N});
$else:
__m512 vacc0x${ABC[N:N+16]} = _mm512_loadu_ps((const float*) w + ${N});
$for M in range(1, MR):
$for N in range(0, NR, 16):
__m512 vacc${M}x${ABC[N:N+16]} = vacc0x${ABC[N:N+16]};
$if DATATYPE == "F32":
w += ${NR};
$else:
w = (const float*) w + ${NR};
size_t k = kc;
do {
$if DATATYPE == "F32":
const __m512 vb${ABC[0:16]} = _mm512_load_ps(w);
$for N in range(16, NR, 16):
const __m512 vb${ABC[N:N+16]} = _mm512_loadu_ps(w + ${N});
w += ${NR};
$else:
const __m512i vbi${ABC[0:16]} = _mm512_cvtepi8_epi32(_mm_loadu_epi8(w));
$for N in range(16, NR, 16):
const __m512i vbi${ABC[N:N+16]} = _mm512_cvtepi8_epi32(_mm_loadu_epi8((const int8_t*) w + ${N}));
$for N in range(0, NR, 16):
const __m512 vb${ABC[N:N+16]} = _mm512_cvtepi32_ps(vbi${ABC[N:N+16]});
w = (const int8_t*) w + ${NR};
$for M in range(MR):
const __m512 va${M} = _mm512_set1_ps(*a${M});
$for N in range(0, NR, 16):
vacc${M}x${ABC[N:N+16]} = _mm512_fmadd_ps(va${M}, vb${ABC[N:N+16]}, vacc${M}x${ABC[N:N+16]});
$for M in range(MR):
a${M} += 1;
k -= sizeof(float);
} while (k != 0);
$if DATATYPE in ["QC8", "QC4"]:
$for N in range(0, NR, 16):
const __m512 vscale${ABC[N:N+16]} = _mm512_loadu_ps((const float*) w + ${N});
$for M in range(MR):
vacc${M}x${ABC[N:N+16]} = _mm512_mul_ps(vacc${M}x${ABC[N:N+16]}, vscale${ABC[N:N+16]});
w = (const float*) w + ${NR};
const __m512 vmin = _mm512_set1_ps(params->scalar.min);
$for N in range(0, NR, 16):
$for M in range(MR):
vacc${M}x${ABC[N:N+16]} = _mm512_max_ps(vmin, vacc${M}x${ABC[N:N+16]});
const __m512 vmax = _mm512_set1_ps(params->scalar.max);
$for N in range(0, NR, 16):
$for M in range(MR):
vacc${M}x${ABC[N:N+16]} = _mm512_min_ps(vmax, vacc${M}x${ABC[N:N+16]});
if XNN_LIKELY(nc >= ${NR}) {
$for M in reversed(range(MR)):
_mm512_storeu_ps(c${M}, vacc${M}x${ABC[0:16]});
$for N in range(16, NR, 16):
_mm512_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+16]});
c${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(4, NR.bit_length())):
$if NR != 1 << LOG2N:
if (nc & ${1 << LOG2N}) {
$if LOG2N >= 4:
$for M in reversed(range(MR)):
_mm512_storeu_ps(c${M}, vacc${M}x${ABC[0:16]});
$for N in range(16, 1 << LOG2N, 16):
_mm512_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+16]});
$for M in reversed(range(MR)):
$for N in range(0, NR - (1 << LOG2N), 16):
vacc${M}x${ABC[N:N+16]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+16]};
$for M in reversed(range(MR)):
c${M} += ${1 << LOG2N};
}
$if LOG2N == 4:
if (nc & 15) {
// Prepare mask for valid 32-bit elements (depends on nc).
const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1)));
$for M in reversed(range(MR)):
_mm512_mask_storeu_ps(c${M}, vmask, vacc${M}x${ABC[0:16]});
}
nc = 0;
}
} while (nc != 0);
}
|