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$assert NR % 4 == 0
7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
8#include <assert.h>
9
10#include <wasm_simd128.h>
11
12#include <xnnpack/igemm.h>
13
14
15$assert ACTIVATION in ["LINEAR", "RELU", "MINMAX"]
16$ACTIVATION_SUFFIX = {"LINEAR": ""}.get(ACTIVATION, "_" + ACTIVATION.lower())
17$ARCH_SUFFIX = "" if ACTIVATION in ["LINEAR", "RELU"] else "_x86" if X86 else "_arm"
18$PARAMS = {"LINEAR": "xnn_f32_default_params", "RELU": "xnn_f32_relu_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION]
19void xnn_f32_igemm${ACTIVATION_SUFFIX}_ukernel_${MR}x${NR}__wasmsimd${ARCH_SUFFIX}_splat(
20    size_t mr,
21    size_t nc,
22    size_t kc,
23    size_t ks,
24    const float**restrict a,
25    const float*restrict w,
26    float*restrict c,
27    size_t cm_stride,
28    size_t cn_stride,
29    size_t a_offset,
30    const float* zero,
31    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)])
32{
33  assert(mr != 0);
34  assert(mr <= ${MR});
35  assert(nc != 0);
36  assert(kc != 0);
37  assert(kc % sizeof(float) == 0);
38  assert(ks != 0);
39  assert(ks % (${MR} * sizeof(void*)) == 0);
40  assert(a_offset % sizeof(float) == 0);
41  assert(a != NULL);
42  assert(w != NULL);
43  assert(c != NULL);
44
45  float* c0 = c;
46  $for M in range(1, MR):
47    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
48    $if M % 2 == 0:
49      if XNN_UNPREDICTABLE(mr <= ${M}) {
50        c${M} = c${M-1};
51      }
52    $elif M + 1 == MR:
53      if XNN_UNPREDICTABLE(mr != ${M+1}) {
54        c${M} = c${M-1};
55      }
56    $else:
57      if XNN_UNPREDICTABLE(mr < ${M+1}) {
58        c${M} = c${M-1};
59      }
60
61  $if ACTIVATION == "MINMAX" and not X86:
62    const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
63    const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
64  do {
65    v128_t vacc0x${ABC[0:4]} = wasm_v128_load(w);
66    $for N in range(4, NR, 4):
67      v128_t vacc0x${ABC[N:N+4]} = wasm_v128_load(w + ${N});
68    $for M in range(1, MR):
69      $for N in range(0, NR, 4):
70        v128_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
71    w += ${NR};
72
73    size_t p = ks;
74    do {
75      $for M in range(MR):
76        const float* restrict a${M} = a[${M}];
77        assert(a${M} != NULL);
78        if XNN_UNPREDICTABLE(a${M} != zero) {
79          a${M} = (const float*) ((uintptr_t) a${M} + a_offset);
80        }
81      a += ${MR};
82
83      size_t k = kc;
84      while (k >= 4 * sizeof(float)) {
85        $for M in range(MR):
86          const v128_t va${M} = wasm_v128_load(a${M});
87          a${M} += 4;
88
89        $for L in range(4):
90          $for M in range(MR):
91            const v128_t va${M}c${L} = wasm_v32x4_shuffle(va${M}, va${M}, ${L}, ${L}, ${L}, ${L});
92
93          $for N in range(0, NR, 4):
94            const v128_t vb${ABC[N:N+4]}c${L} = wasm_v128_load(w + ${L * NR + N});
95
96          $for N in range(0, NR, 4):
97            $for M in range(MR):
98              vacc${M}x${ABC[N:N+4]} = wasm_f32x4_add(vacc${M}x${ABC[N:N+4]}, wasm_f32x4_mul(va${M}c${L}, vb${ABC[N:N+4]}c${L}));
99
100        w += ${4 * NR};
101        k -= 4 * sizeof(float);
102      }
103      if XNN_UNLIKELY(k != 0) {
104        do {
105          const v128_t vb${ABC[0:4]} = wasm_v128_load(w);
106          $for N in range(4, NR, 4):
107            const v128_t vb${ABC[N:N+4]} = wasm_v128_load(w + ${N});
108          w += ${NR};
109
110          $for M in range(MR):
111            const v128_t va${M} = wasm_v32x4_load_splat(a${M});
112            a${M} += 1;
113
114          $for M in range(MR):
115            $for N in range(0, NR, 4):
116              vacc${M}x${ABC[N:N+4]} = wasm_f32x4_add(vacc${M}x${ABC[N:N+4]}, wasm_f32x4_mul(va${M}, vb${ABC[N:N+4]}));
117          k -= sizeof(float);
118        } while (k != 0);
119      }
120      p -= ${MR} * sizeof(void*);
121    } while (p != 0);
122
123    $if ACTIVATION == "MINMAX":
124      $if X86:
125        const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
126        $for N in range(0, NR, 4):
127          $for M in range(MR):
128            vacc${M}x${ABC[N:N+4]} = wasm_v128_bitselect(vmin, vacc${M}x${ABC[N:N+4]}, wasm_f32x4_lt(vacc${M}x${ABC[N:N+4]}, vmin));
129
130        const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
131        $for N in range(0, NR, 4):
132          $for M in range(MR):
133            vacc${M}x${ABC[N:N+4]} = wasm_v128_bitselect(vacc${M}x${ABC[N:N+4]}, vmax, wasm_f32x4_le(vacc${M}x${ABC[N:N+4]}, vmax));
134      $else:
135        $for N in range(0, NR, 4):
136          $for M in range(MR):
137            vacc${M}x${ABC[N:N+4]} = wasm_f32x4_max(vacc${M}x${ABC[N:N+4]}, vmin);
138
139        $for N in range(0, NR, 4):
140          $for M in range(MR):
141            vacc${M}x${ABC[N:N+4]} = wasm_f32x4_min(vacc${M}x${ABC[N:N+4]}, vmax);
142    $elif ACTIVATION == "RELU":
143      const v128_t vzero = wasm_f32x4_splat(0.0f);
144      $for N in range(0, NR, 4):
145        $for M in range(MR):
146          vacc${M}x${ABC[N:N+4]} = wasm_i32x4_max(vacc${M}x${ABC[N:N+4]}, vzero);
147
148    if XNN_LIKELY(nc >= ${NR}) {
149      $for M in reversed(range(MR)):
150        wasm_v128_store(c${M}, vacc${M}x${ABC[0:4]});
151        $for N in range(4, NR, 4):
152          wasm_v128_store(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
153        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
154
155      a = (const float**restrict) ((uintptr_t) a - ks);
156      nc -= ${NR};
157    } else {
158      $for LOG2N in reversed(range(NR.bit_length())):
159        $if NR != 1 << LOG2N:
160          if (nc & ${1 << LOG2N}) {
161            $if LOG2N >= 2:
162              $for M in reversed(range(MR)):
163                wasm_v128_store(c${M}, vacc${M}x${ABC[0:4]});
164                $for N in range(4, 1 << LOG2N, 4):
165                  wasm_v128_store(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
166
167              $for M in reversed(range(MR)):
168                $for N in range(0, 1 << (LOG2N - 1), 4):
169                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]};
170
171              $for M in reversed(range(MR)):
172                c${M} += ${1 << LOG2N};
173            $elif LOG2N == 1:
174              $for M in reversed(range(MR)):
175                *((double*) c${M}) = wasm_f64x2_extract_lane(vacc${M}x${ABC[0:4]}, 0);
176
177              $for M in reversed(range(MR)):
178                vacc${M}x${ABC[0:4]} = wasm_v32x4_shuffle(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}, 2, 3, 2, 3);
179
180              $for M in reversed(range(MR)):
181                c${M} += 2;
182            $elif LOG2N == 0:
183              $for M in reversed(range(MR)):
184                *c${M} = wasm_f32x4_extract_lane(vacc${M}x${ABC[0:4]}, 0);
185          }
186
187      nc = 0;
188    }
189  } while (nc != 0);
190}
191