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$assert NR % 4 == 0 7$assert ACTIVATION != "MINMAX" or ARCH in ["ARM", "X86", "RELAXED"] 8$assert not FMA or ARCH == "RELAXED" 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <wasm_simd128.h> 13 14#include <xnnpack/gemm.h> 15 16 17$assert ACTIVATION in ["LINEAR", "RELU", "MINMAX"] 18$if ACTIVATION == "MINMAX": 19$ WASM_F32X4_MIN={"ARM": "wasm_f32x4_min", "X86": "wasm_f32x4_pmin", "RELAXED": "__builtin_wasm_relaxed_min_f32x4"}[ARCH] 20$ WASM_F32X4_MAX={"ARM": "wasm_f32x4_max", "X86": "wasm_f32x4_pmax", "RELAXED": "__builtin_wasm_relaxed_max_f32x4"}[ARCH] 21$ACTIVATION_SUFFIX = {"LINEAR": ""}.get(ACTIVATION, "_" + ACTIVATION.lower()) 22$ISA = "wasmsimd" if not FMA and (ACTIVATION in ["LINEAR", "RELU"] or ARCH != "RELAXED") else "wasmrelaxedsimd" 23$ARCH_SUFFIX = "" if not FMA and (ACTIVATION in ["LINEAR", "RELU"] or ARCH == "RELAXED") else "_" + ("fma" if FMA else ARCH.lower()) 24$PARAMS = {"LINEAR": "xnn_f32_default_params", "RELU": "xnn_f32_relu_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION] 25void xnn_f32_gemm${"inc" if INC else ""}${ACTIVATION_SUFFIX}_ukernel_${MR}x${NR}__${ISA}${ARCH_SUFFIX}_splat( 26 size_t mr, 27 size_t nc, 28 size_t kc, 29 const float*restrict a, 30 size_t a_stride, 31 const float*restrict w, 32 float*restrict c, 33 size_t cm_stride, 34 size_t cn_stride, 35 $if INC: 36 const float*restrict acc, 37 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) 38{ 39 assert(mr != 0); 40 assert(mr <= ${MR}); 41 assert(nc != 0); 42 assert(kc != 0); 43 assert(kc % sizeof(float) == 0); 44 assert(a != NULL); 45 assert(w != NULL); 46 assert(c != NULL); 47 $if INC: 48 assert(acc != NULL); 49 50 const float* a0 = a; 51 float* c0 = c; 52 $for M in range(1, MR): 53 const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride); 54 float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); 55 $if M % 2 == 0: 56 if XNN_UNPREDICTABLE(mr <= ${M}) { 57 a${M} = a${M-1}; 58 c${M} = c${M-1}; 59 } 60 $elif M + 1 == MR: 61 if XNN_UNPREDICTABLE(mr != ${M+1}) { 62 a${M} = a${M-1}; 63 c${M} = c${M-1}; 64 } 65 $else: 66 if XNN_UNPREDICTABLE(mr < ${M+1}) { 67 a${M} = a${M-1}; 68 c${M} = c${M-1}; 69 } 70 71 $if ACTIVATION == "MINMAX": 72 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min); 73 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max); 74 do { 75 $if INC: 76 $for M in range(MR): 77 $for N in range(0, NR, 4): 78 v128_t vacc${M}x${ABC[N:N+4]} = wasm_v128_load(acc + ${M*NR+N}); 79 acc += ${MR*NR}; 80 $else: 81 $for N in range(0, NR, 4): 82 v128_t vacc0x${ABC[N:N+4]} = wasm_v128_load(w + ${N}); 83 $for M in range(1, MR): 84 $for N in range(0, NR, 4): 85 v128_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]}; 86 w += ${NR}; 87 88 size_t k = kc; 89 while (k >= 4 * sizeof(float)) { 90 $for M in range(MR): 91 const v128_t va${M} = wasm_v128_load(a${M}); 92 a${M} += 4; 93 94 $for L in range(4): 95 $for M in range(MR): 96 const v128_t va${M}c${L} = wasm_v32x4_shuffle(va${M}, va${M}, ${L}, ${L}, ${L}, ${L}); 97 98 $for N in range(0, NR, 4): 99 const v128_t vb${ABC[N:N+4]}c${L} = wasm_v128_load(w + ${L * NR + N}); 100 101 $for N in range(0, NR, 4): 102 $for M in range(MR): 103 vacc${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})); 104 105 w += ${4 * NR}; 106 k -= 4 * sizeof(float); 107 } 108 if XNN_UNLIKELY(k != 0) { 109 do { 110 $for M in range(MR): 111 const v128_t va${M} = wasm_v128_load32_splat(a${M}); 112 a${M} += 1; 113 114 const v128_t vb${ABC[0:4]} = wasm_v128_load(w); 115 $for N in range(4, NR, 4): 116 const v128_t vb${ABC[N:N+4]} = wasm_v128_load(w + ${N}); 117 w += ${NR}; 118 119 $for N in range(0, NR, 4): 120 $for M in range(MR): 121 $if FMA: 122 vacc${M}x${ABC[N:N+4]} = __builtin_wasm_fma_f32x4(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]}); 123 $else: 124 vacc${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]})); 125 126 k -= sizeof(float); 127 } while (k != 0); 128 } 129 130 $if ACTIVATION == "MINMAX": 131 $for N in range(0, NR, 4): 132 $for M in range(MR): 133 vacc${M}x${ABC[N:N+4]} = ${WASM_F32X4_MAX}(vmin, vacc${M}x${ABC[N:N+4]}); 134 135 $for N in range(0, NR, 4): 136 $for M in range(MR): 137 vacc${M}x${ABC[N:N+4]} = ${WASM_F32X4_MIN}(vmax, vacc${M}x${ABC[N:N+4]}); 138 $elif ACTIVATION == "RELU": 139 const v128_t vzero = wasm_i32x4_const_splat(0); 140 $for N in range(0, NR, 4): 141 $for M in range(MR): 142 vacc${M}x${ABC[N:N+4]} = wasm_i32x4_max(vacc${M}x${ABC[N:N+4]}, vzero); 143 144 if XNN_LIKELY(nc >= ${NR}) { 145 $for M in reversed(range(MR)): 146 wasm_v128_store(c${M}, vacc${M}x${ABC[0:4]}); 147 $for N in range(4, NR, 4): 148 wasm_v128_store(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 149 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 150 151 $for M in reversed(range(MR)): 152 a${M} = (const float*) ((uintptr_t) a${M} - kc); 153 154 nc -= ${NR}; 155 } else { 156 $for LOG2N in reversed(range(NR.bit_length())): 157 $if NR != 1 << LOG2N: 158 if (nc & ${1 << LOG2N}) { 159 $if LOG2N >= 2: 160 $for M in reversed(range(MR)): 161 wasm_v128_store(c${M}, vacc${M}x${ABC[0:4]}); 162 $for N in range(4, 1 << LOG2N, 4): 163 wasm_v128_store(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 164 165 $for M in reversed(range(MR)): 166 $for N in range(0, 1 << (LOG2N - 1), 4): 167 vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; 168 169 $for M in reversed(range(MR)): 170 c${M} += ${1 << LOG2N}; 171 $elif LOG2N == 1: 172 $for M in reversed(range(MR)): 173 *((double*) c${M}) = wasm_f64x2_extract_lane(vacc${M}x${ABC[0:4]}, 0); 174 175 $for M in reversed(range(MR)): 176 vacc${M}x${ABC[0:4]} = wasm_v32x4_shuffle(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}, 2, 3, 2, 3); 177 178 $for M in reversed(range(MR)): 179 c${M} += 2; 180 $elif LOG2N == 0: 181 $for M in reversed(range(MR)): 182 *c${M} = wasm_f32x4_extract_lane(vacc${M}x${ABC[0:4]}, 0); 183 } 184 185 nc = 0; 186 } 187 } while (nc != 0); 188} 189