1// Copyright 2019 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 <xmmintrin.h>
11
12#include <xnnpack/igemm.h>
13
14
15void xnn_f32_igemm_minmax_ukernel_${MR}x${NR}c4__sse(
16    size_t mr,
17    size_t nc,
18    size_t kc,
19    size_t ks,
20    const float**restrict a,
21    const float*restrict w,
22    float*restrict c,
23    size_t cm_stride,
24    size_t cn_stride,
25    size_t a_offset,
26    const float* zero,
27    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28{
29  assert(mr != 0);
30  assert(mr <= ${MR});
31  assert(nc != 0);
32  assert(kc != 0);
33  assert(kc % sizeof(float) == 0);
34  assert(ks != 0);
35  assert(ks % (${MR} * sizeof(void*)) == 0);
36  assert(a_offset % sizeof(float) == 0);
37  assert(a != NULL);
38  assert(w != NULL);
39  assert(c != NULL);
40
41  float* c0 = c;
42  $for M in range(1, MR):
43    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
44    $if M % 2 == 0:
45      if XNN_UNPREDICTABLE(mr <= ${M}) {
46        c${M} = c${M-1};
47      }
48    $elif M + 1 == MR:
49      if XNN_UNPREDICTABLE(mr != ${M+1}) {
50        c${M} = c${M-1};
51      }
52    $else:
53      if XNN_UNPREDICTABLE(mr < ${M+1}) {
54        c${M} = c${M-1};
55      }
56
57  do {
58    __m128 vacc0x0c4 = _mm_load_ss(w);
59    $for N in range(1, NR):
60      __m128 vacc0x${N}c4 = _mm_load_ss(w + ${N});
61    $for M in range(1, MR):
62      $for N in range(NR):
63        __m128 vacc${M}x${N}c4 = vacc0x${N}c4;
64    w += ${NR};
65
66    size_t p = ks;
67    do {
68      $for M in range(MR):
69        const float* restrict a${M} = a[${M}];
70        assert(a${M} != NULL);
71        if XNN_UNPREDICTABLE(a${M} != zero) {
72          a${M} = (const float*) ((uintptr_t) a${M} + a_offset);
73        }
74      a += ${MR};
75
76      size_t k = kc;
77      for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
78        $for M in range(MR):
79          const __m128 va${M} = _mm_loadu_ps(a${M});
80          a${M} += 4;
81
82        const __m128 vb0 = _mm_loadu_ps(w);
83        $for N in range(1, NR):
84          const __m128 vb${N} = _mm_loadu_ps(w + ${N * 4});
85        w += ${NR * 4};
86
87        $for M in range(MR):
88          $for N in range(NR):
89            vacc${M}x${N}c4 = _mm_add_ps(vacc${M}x${N}c4, _mm_mul_ps(va${M}, vb${N}));
90      }
91      if XNN_UNLIKELY(k != 0) {
92        $for M in range(MR):
93          const __m128 va${M} = _mm_loadu_ps(a${M});
94
95        const __m128 vb0 = _mm_loadu_ps(w);
96        $for N in range(1, NR):
97          const __m128 vb${N} = _mm_loadu_ps(w + ${N * 4});
98        w += ${NR * 4};
99
100        $for N in range(NR):
101          const __m128 vmask${N} = _mm_cmpeq_ps(_mm_setzero_ps(), vb${N});
102
103        $for M in range(MR):
104          $for N in range(NR):
105            vacc${M}x${N}c4 = _mm_add_ps(vacc${M}x${N}c4, _mm_mul_ps(_mm_andnot_ps(vmask${N}, va${M}), vb${N}));
106      }
107      p -= ${MR} * sizeof(void*);
108    } while (p != 0);
109
110    $for M in range(MR):
111      const __m128 vacc${M}x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc${M}x0c4, vacc${M}x1c4), _mm_unpackhi_ps(vacc${M}x0c4, vacc${M}x1c4));
112
113    $for M in range(0, MR, 2):
114      __m128 vacc${M}${M+1}x01 = _mm_add_ps(_mm_movelh_ps(vacc${M}x01c2, vacc${M+1}x01c2), _mm_movehl_ps(vacc${M+1}x01c2, vacc${M}x01c2));
115
116    const __m128 vmax = _mm_load_ps(params->sse.max);
117    $for M in range(0, MR, 2):
118      vacc${M}${M+1}x01 = _mm_min_ps(vacc${M}${M+1}x01, vmax);
119
120    const __m128 vmin = _mm_load_ps(params->sse.min);
121    $for M in range(0, MR, 2):
122      vacc${M}${M+1}x01 = _mm_max_ps(vacc${M}${M+1}x01, vmin);
123
124    if XNN_LIKELY(nc >= ${NR}) {
125      $for M in reversed(range(0, MR, 2)):
126        _mm_storeh_pi((__m64*) c${M+1}, vacc${M}${M+1}x01);
127        c${M+1} = (float*) ((uintptr_t) c${M+1} + cn_stride);
128        _mm_storel_pi((__m64*) c${M}, vacc${M}${M+1}x01);
129        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
130
131      a = (const float**restrict) ((uintptr_t) a - ks);
132      nc -= ${NR};
133    } else {
134      assert(nc == 1);
135      $for M in reversed(range(0, MR, 2)):
136        _mm_store_ss(c${M+1}, _mm_movehl_ps(vacc${M}${M+1}x01, vacc${M}${M+1}x01));
137        _mm_store_ss(c${M}, vacc${M}${M+1}x01);
138
139      nc = 0;
140    }
141  } while (nc != 0);
142}
143