Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame^] | 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 | $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" |
| 8 | #include <assert.h> |
| 9 | |
| 10 | #include <wasm_simd128.h> |
| 11 | |
| 12 | #include <xnnpack/gemm.h> |
| 13 | |
| 14 | |
| 15 | void xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}s4__wasmsimd_${"x86" if X86 else "arm"}( |
| 16 | size_t mr, |
| 17 | size_t nc, |
| 18 | size_t kc, |
| 19 | const float*restrict a, |
| 20 | size_t a_stride, |
| 21 | const float*restrict w, |
| 22 | float*restrict c, |
| 23 | size_t cm_stride, |
| 24 | size_t cn_stride, |
| 25 | $if INC: |
| 26 | const float*restrict acc, |
| 27 | const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 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(a != NULL); |
| 35 | assert(w != NULL); |
| 36 | assert(c != NULL); |
| 37 | $if INC: |
| 38 | assert(acc != NULL); |
| 39 | |
| 40 | const float* a0 = a; |
| 41 | float* c0 = c; |
| 42 | $for M in range(1, MR): |
| 43 | const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride); |
| 44 | float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); |
| 45 | $if M % 2 == 0: |
| 46 | if XNN_UNPREDICTABLE(mr <= ${M}) { |
| 47 | a${M} = a${M-1}; |
| 48 | c${M} = c${M-1}; |
| 49 | } |
| 50 | $elif M + 1 == MR: |
| 51 | if XNN_UNPREDICTABLE(mr != ${M+1}) { |
| 52 | a${M} = a${M-1}; |
| 53 | c${M} = c${M-1}; |
| 54 | } |
| 55 | $else: |
| 56 | if XNN_UNPREDICTABLE(mr < ${M+1}) { |
| 57 | a${M} = a${M-1}; |
| 58 | c${M} = c${M-1}; |
| 59 | } |
| 60 | |
| 61 | do { |
| 62 | $if INC: |
| 63 | $for M in range(MR): |
| 64 | $for N in range(0, NR, 4): |
| 65 | v128_t vacc${M}x${ABC[N:N+4]} = wasm_v128_load(acc + ${M*NR+N}); |
| 66 | acc += ${MR*NR}; |
| 67 | $else: |
| 68 | $for N in range(0, NR, 4): |
| 69 | v128_t vacc0x${ABC[N:N+4]} = wasm_v128_load(w + ${N}); |
| 70 | $for M in range(1, MR): |
| 71 | $for N in range(0, NR, 4): |
| 72 | v128_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]}; |
| 73 | w += ${NR}; |
| 74 | |
| 75 | size_t k = kc; |
| 76 | while (k >= 4 * sizeof(float)) { |
| 77 | $for M in range(MR): |
| 78 | v128_t va${M} = wasm_v128_load(a${M}); |
| 79 | a${M} += 4; |
| 80 | |
| 81 | $for L in range(4): |
| 82 | |
| 83 | $for N in range(0, NR, 4): |
| 84 | const v128_t vb${ABC[N:N+4]}c${L} = wasm_v128_load(w + ${L * NR + N}); |
| 85 | |
| 86 | $for N in range(0, NR, 4): |
| 87 | $for M in range(MR): |
| 88 | vacc${M}x${ABC[N:N+4]} = wasm_f32x4_add(vacc${M}x${ABC[N:N+4]}, wasm_f32x4_mul(va${M}, vb${ABC[N:N+4]}c${L})); |
| 89 | |
| 90 | $if L + 1 != 4: |
| 91 | $for M in range(MR): |
| 92 | va${M} = wasm_v32x4_shuffle(va${M}, va${M}, 1, 2, 3, 0); |
| 93 | |
| 94 | w += ${4 * NR}; |
| 95 | k -= 4 * sizeof(float); |
| 96 | } |
| 97 | if XNN_UNLIKELY(k != 0) { |
| 98 | do { |
| 99 | $for M in range(MR): |
| 100 | const v128_t va${M} = wasm_v32x4_load_splat(a${M}); |
| 101 | a${M} += 1; |
| 102 | |
| 103 | const v128_t vb${ABC[0:4]} = wasm_v128_load(w); |
| 104 | $for N in range(4, NR, 4): |
| 105 | const v128_t vb${ABC[N:N+4]} = wasm_v128_load(w + ${N}); |
| 106 | w += ${NR}; |
| 107 | |
| 108 | $for N in range(0, NR, 4): |
| 109 | $for M in range(MR): |
| 110 | vacc${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]})); |
| 111 | |
| 112 | k -= sizeof(float); |
| 113 | } while (k != 0); |
| 114 | } |
| 115 | |
| 116 | const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max); |
| 117 | $for N in range(0, NR, 4): |
| 118 | $for M in range(MR): |
| 119 | $if X86: |
| 120 | vacc${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)); |
| 121 | $else: |
| 122 | vacc${M}x${ABC[N:N+4]} = wasm_f32x4_min(vacc${M}x${ABC[N:N+4]}, vmax); |
| 123 | |
| 124 | const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); |
| 125 | $for N in range(0, NR, 4): |
| 126 | $for M in range(MR): |
| 127 | $if X86: |
| 128 | vacc${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 | $else: |
| 130 | vacc${M}x${ABC[N:N+4]} = wasm_f32x4_max(vacc${M}x${ABC[N:N+4]}, vmin); |
| 131 | |
| 132 | if XNN_LIKELY(nc >= ${NR}) { |
| 133 | $for M in reversed(range(MR)): |
| 134 | wasm_v128_store(c${M}, vacc${M}x${ABC[0:4]}); |
| 135 | $for N in range(4, NR, 4): |
| 136 | wasm_v128_store(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); |
| 137 | c${M} = (float*) ((uintptr_t) c${M} + cn_stride); |
| 138 | |
| 139 | $for M in reversed(range(MR)): |
| 140 | a${M} = (const float*) ((uintptr_t) a${M} - kc); |
| 141 | |
| 142 | nc -= ${NR}; |
| 143 | } else { |
| 144 | $for LOG2N in reversed(range(NR.bit_length())): |
| 145 | $if NR != 1 << LOG2N: |
| 146 | if (nc & ${1 << LOG2N}) { |
| 147 | $if LOG2N >= 2: |
| 148 | $for M in reversed(range(MR)): |
| 149 | wasm_v128_store(c${M}, vacc${M}x${ABC[0:4]}); |
| 150 | $for N in range(4, 1 << LOG2N, 4): |
| 151 | wasm_v128_store(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); |
| 152 | |
| 153 | $for M in reversed(range(MR)): |
| 154 | $for N in range(0, 1 << (LOG2N - 1), 4): |
| 155 | vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; |
| 156 | |
| 157 | $for M in reversed(range(MR)): |
| 158 | c${M} += ${1 << LOG2N}; |
| 159 | $elif LOG2N == 1: |
| 160 | $for M in reversed(range(MR)): |
| 161 | *((double*) c${M}) = wasm_f64x2_extract_lane(vacc${M}x${ABC[0:4]}, 0); |
| 162 | |
| 163 | $for M in reversed(range(MR)): |
| 164 | vacc${M}x${ABC[0:4]} = wasm_v32x4_shuffle(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}, 2, 3, 2, 3); |
| 165 | |
| 166 | $for M in reversed(range(MR)): |
| 167 | c${M} += 2; |
| 168 | $elif LOG2N == 0: |
| 169 | $for M in reversed(range(MR)): |
| 170 | *c${M} = wasm_f32x4_extract_lane(vacc${M}x${ABC[0:4]}, 0); |
| 171 | } |
| 172 | |
| 173 | nc = 0; |
| 174 | } |
| 175 | } while (nc != 0); |
| 176 | } |