File size: 7,276 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
// Copyright 2020 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", "QC8"]
$assert NR == 2
$assert MR % 2 == 0
$assert ACTIVATION in ["LINEAR", "RELU", "MINMAX"]
$assert ACTIVATION != "MINMAX" or ARCH in ["ARM", "X86", "RELAXED"]
$assert not FMA or ARCH == "RELAXED"
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/gemm.h>


$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w"}[DATATYPE]
$if ACTIVATION == "MINMAX":
$  WASM_F32X4_MIN={"ARM": "wasm_f32x4_min", "X86": "wasm_f32x4_pmin", "RELAXED": "wasm_f32x4_relaxed_min"}[ARCH]
$  WASM_F32X4_MAX={"ARM": "wasm_f32x4_max", "X86": "wasm_f32x4_pmax", "RELAXED": "wasm_f32x4_relaxed_max"}[ARCH]
$ACTIVATION_SUFFIX = {"LINEAR": ""}.get(ACTIVATION, "_" + ACTIVATION.lower())
$ISA = "wasmsimd" if not FMA and (ACTIVATION in ["LINEAR", "RELU"] or ARCH != "RELAXED") else "wasmrelaxedsimd"
$ARCH_SUFFIX = "" if not FMA and (ACTIVATION in ["LINEAR", "RELU"] or ARCH == "RELAXED") else "_" + ("fma" if FMA else ARCH.lower())
$PARAMS = {"LINEAR": "xnn_f32_default_params", "RELU": "xnn_f32_relu_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION]
void xnn_${DATATYPE_SPEC}_gemm${ACTIVATION_SUFFIX}_ukernel_${MR}x${NR}c4__${ISA}${ARCH_SUFFIX}(
    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,
    const union ${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);

  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 ACTIVATION == "MINMAX":
    const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
    const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
  do {
    v128_t vacc0x0c4 = wasm_v128_load32_zero(w);
    $for N in range(1, NR):
      $if DATATYPE == "F32":
        v128_t vacc0x${N}c4 = wasm_v128_load32_zero(w + ${N});
      $else:
        v128_t vacc0x${N}c4 = wasm_v128_load32_zero((const float*) w + ${N});
    $for M in range(1, MR):
      $for N in range(NR):
        v128_t vacc${M}x${N}c4 = vacc0x${N}c4;
    $if DATATYPE == "F32":
      w += ${NR};
    $else:
      w = (const float*) w + ${NR};

    size_t k = kc;
    for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
      $for M in range(MR):
        const v128_t va${M} = wasm_v128_load(a${M});
        a${M} += 4;

      $if DATATYPE == "F32":
        const v128_t vb0 = wasm_v128_load(w);
      $else:
        const v128_t vb0 = wasm_f32x4_convert_i32x4(wasm_i32x4_extend_low_i16x8(wasm_i16x8_extend_low_i8x16(wasm_v128_load32_splat((const int8_t*) w))));
      $for N in range(1, NR):
        $if DATATYPE == "F32":
          const v128_t vb${N} = wasm_v128_load(w + ${N * 4});
        $else:
          const v128_t vb${N} = wasm_f32x4_convert_i32x4(wasm_i32x4_extend_low_i16x8(wasm_i16x8_extend_low_i8x16(wasm_v128_load32_splat((const int8_t*) w + ${N * 4}))));
      $if DATATYPE == "F32":
        w += ${NR * 4};
      $else:
        w = (const int8_t*) w + ${NR * 4};

      $for M in range(MR):
        $for N in range(NR):
          $if FMA:
            vacc${M}x${N}c4 = wasm_f32x4_relaxed_madd(va${M}, vb${N}, vacc${M}x${N}c4);
          $else:
            vacc${M}x${N}c4 = wasm_f32x4_add(wasm_f32x4_mul(va${M}, vb${N}), vacc${M}x${N}c4);
    }
    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const v128_t va${M} = wasm_v128_load(a${M});
        a${M} = (const float*) ((uintptr_t) a${M} + k);

      $if DATATYPE == "F32":
        const v128_t vb0 = wasm_v128_load(w);
      $else:
        const v128_t vb0 = wasm_f32x4_convert_i32x4(wasm_i32x4_extend_low_i16x8(wasm_i16x8_extend_low_i8x16(wasm_v128_load32_splat((const int8_t*) w))));
      $for N in range(1, NR):
        $if DATATYPE == "F32":
          const v128_t vb${N} = wasm_v128_load(w + ${N * 4});
        $else:
          const v128_t vb${N} = wasm_f32x4_convert_i32x4(wasm_i32x4_extend_low_i16x8(wasm_i16x8_extend_low_i8x16(wasm_v128_load32_splat((const int8_t*) w + ${N * 4}))));
      $if DATATYPE == "F32":
        w += ${NR * 4};
      $else:
        w = (const int8_t*) w + ${NR * 4};

      const v128_t vzero = wasm_f32x4_const_splat(0.0f);
      $for N in range(NR):
        const v128_t vmask${N} = wasm_f32x4_eq(vb${N}, vzero);

      $for M in range(MR):
        $for N in range(NR):
          $if FMA:
            vacc${M}x${N}c4 = wasm_f32x4_relaxed_madd(wasm_v128_andnot(va${M}, vmask${N}), vb${N}, vacc${M}x${N}c4);
          $else:
            vacc${M}x${N}c4 = wasm_f32x4_add(wasm_f32x4_mul(wasm_v128_andnot(va${M}, vmask${N}), vb${N}), vacc${M}x${N}c4);
    }

    $for M in range(MR):
      const v128_t vacc${M}x01c2 = wasm_f32x4_add(
        wasm_v32x4_shuffle(vacc${M}x0c4, vacc${M}x1c4, 0, 4, 1, 5),
        wasm_v32x4_shuffle(vacc${M}x0c4, vacc${M}x1c4, 2, 6, 3, 7));

    $for M in range(0, MR, 2):
      v128_t vacc${M}${M+1}x01 = wasm_f32x4_add(
        wasm_v32x4_shuffle(vacc${M}x01c2, vacc${M+1}x01c2, 0, 1, 4, 5),
        wasm_v32x4_shuffle(vacc${M}x01c2, vacc${M+1}x01c2, 2, 3, 6, 7));

    $if DATATYPE == "QC8":
      const v128_t vscalex01 = wasm_v128_load64_splat(w);
      w = (const float*) w + 2;
      $for M in range(0, MR, 2):
        vacc${M}${M+1}x01 = wasm_f32x4_mul(vacc${M}${M+1}x01, vscalex01);
    $if ACTIVATION == "MINMAX":
      $for M in range(0, MR, 2):
        vacc${M}${M+1}x01 = ${WASM_F32X4_MAX}(vmin, vacc${M}${M+1}x01);

      $for M in range(0, MR, 2):
        vacc${M}${M+1}x01 = ${WASM_F32X4_MIN}(vmax, vacc${M}${M+1}x01);
    $elif ACTIVATION == "RELU":
      const v128_t vzero = wasm_i32x4_const_splat(0);
      $for M in range(0, MR, 2):
        vacc${M}${M+1}x01 = wasm_i32x4_max(vacc${M}${M+1}x01, vzero);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(0, MR, 2)):
        wasm_v128_store64_lane(c${M}, vacc${M}${M+1}x01, 0);
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
        a${M} = (const float*) ((uintptr_t) a${M} - kc);
        wasm_v128_store64_lane(c${M+1}, vacc${M}${M+1}x01, 1);
        c${M+1} = (float*) ((uintptr_t) c${M+1} + cn_stride);
        a${M+1} = (const float*) ((uintptr_t) a${M+1} - kc);

      nc -= ${NR};
    } else {
      assert(nc == 1);
      $for M in reversed(range(0, MR, 2)):
        wasm_v128_store32_lane(c${M}, vacc${M}${M+1}x01, 0);
        wasm_v128_store32_lane(c${M+1}, vacc${M}${M+1}x01, 2);

      nc = 0;
    }
  } while (nc != 0);
}