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$assert NR == 2
7$assert MR % 2 == 0
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${NR}c4__wasmsimd${ARCH_SUFFIX}(
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)]) XNN_DISABLE_TSAN
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 vacc0x0c4 = wasm_f32x4_replace_lane(wasm_f32x4_splat(0.0f), 0, w[0]);
66    $for N in range(1, NR):
67      v128_t vacc0x${N}c4 = wasm_f32x4_replace_lane(vacc0x0c4, 0, w[${N}]);
68    $for M in range(1, MR):
69      $for N in range(NR):
70        v128_t vacc${M}x${N}c4 = vacc0x${N}c4;
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      for (; k >= 4 * sizeof(float); 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        const v128_t vb0 = wasm_v128_load(w);
90        $for N in range(1, NR):
91          const v128_t vb${N} = wasm_v128_load(w + ${N * 4});
92        w += ${NR * 4};
93
94        $for M in range(MR):
95          $for N in range(NR):
96            vacc${M}x${N}c4 = wasm_f32x4_add(vacc${M}x${N}c4, wasm_f32x4_mul(va${M}, vb${N}));
97      }
98      if XNN_UNLIKELY(k != 0) {
99        $for M in range(MR):
100          const v128_t va${M} = wasm_v128_load(a${M});
101
102        const v128_t vb0 = wasm_v128_load(w);
103        $for N in range(1, NR):
104          const v128_t vb${N} = wasm_v128_load(w + ${N * 4});
105        w += ${NR * 4};
106
107        const v128_t vzero = wasm_f32x4_splat(0.0f);
108        $for N in range(NR):
109          const v128_t vmask${N} = wasm_f32x4_eq(vb${N}, vzero);
110
111        $for M in range(MR):
112          $for N in range(NR):
113            vacc${M}x${N}c4 = wasm_f32x4_add(vacc${M}x${N}c4, wasm_f32x4_mul(wasm_v128_andnot(va${M}, vmask${N}), vb${N}));
114      }
115      p -= ${MR} * sizeof(void*);
116    } while (p != 0);
117
118    $for M in range(MR):
119      const v128_t vacc${M}x01c2 = wasm_f32x4_add(
120        wasm_v32x4_shuffle(vacc${M}x0c4, vacc${M}x1c4, 0, 4, 1, 5),
121        wasm_v32x4_shuffle(vacc${M}x0c4, vacc${M}x1c4, 2, 6, 3, 7));
122
123    $for M in range(0, MR, 2):
124      v128_t vacc${M}${M+1}x01 = wasm_f32x4_add(
125        wasm_v32x4_shuffle(vacc${M}x01c2, vacc${M+1}x01c2, 0, 1, 4, 5),
126        wasm_v32x4_shuffle(vacc${M}x01c2, vacc${M+1}x01c2, 2, 3, 6, 7));
127
128    $if ACTIVATION == "MINMAX":
129      $if X86:
130        const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
131        $for M in range(0, MR, 2):
132          vacc${M}${M+1}x01 = wasm_v128_bitselect(vmin, vacc${M}${M+1}x01, wasm_f32x4_lt(vacc${M}${M+1}x01, vmin));
133      $else:
134        $for M in range(0, MR, 2):
135          vacc${M}${M+1}x01 = wasm_f32x4_max(vacc${M}${M+1}x01, vmin);
136
137      $if X86:
138        const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
139        $for M in range(0, MR, 2):
140          vacc${M}${M+1}x01 = wasm_v128_bitselect(vacc${M}${M+1}x01, vmax, wasm_f32x4_le(vacc${M}${M+1}x01, vmax));
141      $else:
142        $for M in range(0, MR, 2):
143          vacc${M}${M+1}x01 = wasm_f32x4_min(vacc${M}${M+1}x01, vmax);
144    $elif ACTIVATION == "RELU":
145      const v128_t vzero = wasm_f32x4_splat(0.0f);
146      $for M in range(0, MR, 2):
147        vacc${M}${M+1}x01 = wasm_i32x4_max(vacc${M}${M+1}x01, vzero);
148
149    if XNN_LIKELY(nc >= ${NR}) {
150      $for M in reversed(range(0, MR, 2)):
151        *((double*) c${M+1}) = wasm_f64x2_extract_lane(vacc${M}${M+1}x01, 1);
152        c${M+1} = (float*) ((uintptr_t) c${M+1} + cn_stride);
153        *((double*) c${M}) = wasm_f64x2_extract_lane(vacc${M}${M+1}x01, 0);
154        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
155
156      a = (const float**restrict) ((uintptr_t) a - ks);
157      nc -= ${NR};
158    } else {
159      assert(nc == 1);
160      $for M in reversed(range(0, MR, 2)):
161        *c${M+1} = wasm_f32x4_extract_lane(vacc${M}${M+1}x01, 2);
162        *c${M} = wasm_f32x4_extract_lane(vacc${M}${M+1}x01, 0);
163
164      nc = 0;
165    }
166  } while (nc != 0);
167}
168