1// Copyright 2021 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$assert BATCH_TILE >= 16 7$assert BATCH_TILE % 16 == 0 8$SIMD_TILE = BATCH_TILE // 16 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <wasm_simd128.h> 13 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/lut.h> 16#include <xnnpack/common.h> 17 18 19void xnn_x8_lut_ukernel__wasmsimd_x${BATCH_TILE}( 20 size_t n, 21 const uint8_t* x, 22 uint8_t* y, 23 const uint8_t t[restrict XNN_MIN_ELEMENTS(256)]) 24{ 25 assert(n != 0); 26 assert(x != NULL); 27 assert(y != NULL); 28 29 const v128_t vtable0 = wasm_v128_load(t); 30 $for T in range(1, 16): 31 const v128_t vtable${T} = wasm_v128_load(t + ${T * 16}); 32 const v128_t voffset = wasm_i8x16_const_splat(16); 33 $if BATCH_TILE > 16: 34 for (; n >= ${BATCH_TILE} * sizeof(uint8_t); n -= ${BATCH_TILE} * sizeof(uint8_t)) { 35 v128_t vx0 = wasm_v128_load(x); 36 $for N in range(1, SIMD_TILE): 37 v128_t vx${N} = wasm_v128_load(x + ${N * 16}); 38 x += ${BATCH_TILE}; 39 40 $for N in range(SIMD_TILE): 41 v128_t vy${N} = wasm_i8x16_swizzle(vtable0, vx${N}); 42 43 $for T in range(1, 16): 44 $for N in range(SIMD_TILE): 45 vx${N} = wasm_i8x16_sub(vx${N}, voffset); 46 vy${N} = wasm_v128_or(vy${N}, wasm_i8x16_swizzle(vtable${T}, vx${N})); 47 48 wasm_v128_store(y, vy0); 49 $for N in range(1, SIMD_TILE): 50 wasm_v128_store(y + ${N * 16}, vy${N}); 51 y += ${BATCH_TILE}; 52 } 53 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) { 54 v128_t vx = wasm_v128_load(x); 55 x += 16; 56 57 v128_t vy = wasm_i8x16_swizzle(vtable0, vx); 58 59 $for T in range(1, 16): 60 vx = wasm_i8x16_sub(vx, voffset); 61 vy = wasm_v128_or(vy, wasm_i8x16_swizzle(vtable${T}, vx)); 62 63 wasm_v128_store(y, vy); 64 y += 16; 65 } 66 if XNN_UNLIKELY(n != 0) { 67 v128_t vx = wasm_v128_load(x); 68 69 v128_t vy = wasm_i8x16_swizzle(vtable0, vx); 70 71 $for T in range(1, 16): 72 vx = wasm_i8x16_sub(vx, voffset); 73 vy = wasm_v128_or(vy, wasm_i8x16_swizzle(vtable${T}, vx)); 74 75 if (n & (8 * sizeof(uint8_t))) { 76 *((double*) y) = wasm_f64x2_extract_lane(vy, 0); 77 vy = wasm_v64x2_shuffle(vy, vy, 1, 1); 78 y += 8; 79 } 80 if (n & (4 * sizeof(uint8_t))) { 81 *((float*) y) = wasm_f32x4_extract_lane(vy, 0); 82 vy = wasm_u64x2_shr(vy, 32); 83 y += 4; 84 } 85 uint32_t vy_lo = wasm_i32x4_extract_lane(vy, 0); 86 if (n & (2 * sizeof(uint8_t))) { 87 *((uint16_t*) y) = (uint16_t) vy_lo; 88 vy_lo >>= 16; 89 y += 2; 90 } 91 if (n & (1 * sizeof(uint8_t))) { 92 *y = (uint8_t) vy_lo; 93 } 94 } 95} 96