1// Copyright 2022 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/common.h> 15#include <xnnpack/vcvt.h> 16 17 18$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 19$WASM_X16X8_LOAD8X8 = {"QS8": "wasm_i16x8_load8x8", "QU8": "wasm_u16x8_load8x8"}[DATATYPE] 20$WASM_X16X8_EXTEND_LOW_X8X16 = {"QS8": "wasm_i16x8_extend_low_i8x16", "QU8": "wasm_u16x8_extend_low_u8x16"}[DATATYPE] 21$WASM_X16X8_EXTEND_HIGH_X8X16 = {"QS8": "wasm_i16x8_extend_high_i8x16", "QU8": "wasm_u16x8_extend_high_u8x16"}[DATATYPE] 22$WASM_I16X8_Q15MULR = "__builtin_wasm_relaxed_q15mulr_s_i16x8" if RELAXED else "wasm_i16x8_q15mulr_sat" 23$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE] 24$ISA = "wasmrelaxedsimd" if RELAXED else "wasmsimd" 25void xnn_${DATATYPE.lower()}_vlrelu_ukernel__${ISA}_arm_x${BATCH_TILE}( 26 size_t n, 27 const ${XINT8_T}* x, 28 ${XINT8_T}* y, 29 const union xnn_${DATATYPE.lower()}_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 30{ 31 assert(n != 0); 32 assert(n % sizeof(${XINT8_T}) == 0); 33 assert(x != NULL); 34 assert(y != NULL); 35 36 const v128_t vinput_zero_point = wasm_v128_load64_splat(params->wasmsimd_arm.input_zero_point); 37 const v128_t vpositive_multiplier = wasm_v128_load64_splat(params->wasmsimd_arm.positive_multiplier); 38 const v128_t vnegative_multiplier = wasm_v128_load64_splat(params->wasmsimd_arm.negative_multiplier); 39 const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd_arm.output_zero_point); 40 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 41 v128_t vx${ABC[0]} = wasm_v128_load(x); 42 $for N in range(1, SIMD_TILE): 43 v128_t vx${ABC[N]} = wasm_v128_load(x + ${N * 16}); 44 x += ${BATCH_TILE}; 45 46 $for N in range(SIMD_TILE): 47 v128_t vacc${ABC[2*N]} = wasm_i16x8_sub(vinput_zero_point, ${WASM_X16X8_EXTEND_LOW_X8X16}(vx${ABC[N]})); 48 v128_t vacc${ABC[2*N+1]} = wasm_i16x8_sub(vinput_zero_point, ${WASM_X16X8_EXTEND_HIGH_X8X16}(vx${ABC[N]})); 49 v128_t vmultiplier${ABC[2*N]} = wasm_i16x8_shr(vacc${ABC[2*N]}, 15); 50 v128_t vmultiplier${ABC[2*N+1]} = wasm_i16x8_shr(vacc${ABC[2*N+1]}, 15); 51 52 $for N in range(2*SIMD_TILE): 53 vacc${ABC[N]} = wasm_i16x8_shl(vacc${ABC[N]}, 7); 54 vmultiplier${ABC[N]} = wasm_v128_bitselect(vpositive_multiplier, vnegative_multiplier, vmultiplier${ABC[N]}); 55 56 $for N in range(2*SIMD_TILE): 57 vacc${ABC[N]} = ${WASM_I16X8_Q15MULR}(vacc${ABC[N]}, vmultiplier${ABC[N]}); 58 59 $for N in range(2*SIMD_TILE): 60 vacc${ABC[N]} = wasm_i16x8_add_sat(vacc${ABC[N]}, voutput_zero_point); 61 62 $for N in range(SIMD_TILE): 63 const v128_t vy${ABC[N]} = ${WASM_X8X16_NARROW_I16X8}(vacc${ABC[2*N]}, vacc${ABC[2*N+1]}); 64 65 wasm_v128_store(y, vy${ABC[0]}); 66 $for N in range(1, SIMD_TILE): 67 wasm_v128_store((y + ${N * 16}), vy${ABC[N]}); 68 y += ${BATCH_TILE}; 69 } 70 for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) { 71 const v128_t vx = ${WASM_X16X8_LOAD8X8}(x); 72 v128_t vacc = wasm_i16x8_sub(vinput_zero_point, vx); 73 v128_t vmultiplier = wasm_i16x8_shr(vacc, 15); 74 vacc = wasm_i16x8_shl(vacc, 7); 75 vmultiplier = wasm_v128_bitselect(vpositive_multiplier, vnegative_multiplier, vmultiplier); 76 vacc = ${WASM_I16X8_Q15MULR}(vacc, vmultiplier); 77 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point); 78 x += 8; 79 80 const v128_t vy = ${WASM_X8X16_NARROW_I16X8}(vacc, vacc); 81 wasm_v128_store64_lane(y, vy, 0); 82 y += 8; 83 } 84 if XNN_UNLIKELY(n != 0) { 85 assert(n >= 1 * sizeof(${XINT8_T})); 86 assert(n <= 7 * sizeof(${XINT8_T})); 87 88 const v128_t vx = ${WASM_X16X8_LOAD8X8}(x); 89 v128_t vacc = wasm_i16x8_sub(vinput_zero_point, vx); 90 v128_t vmultiplier = wasm_i16x8_shr(vacc, 15); 91 vacc = wasm_i16x8_shl(vacc, 7); 92 vmultiplier = wasm_v128_bitselect(vpositive_multiplier, vnegative_multiplier, vmultiplier); 93 vacc = ${WASM_I16X8_Q15MULR}(vacc, vmultiplier); 94 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point); 95 96 v128_t vy = ${WASM_X8X16_NARROW_I16X8}(vacc, vacc); 97 if (n & (4 * sizeof(${XINT8_T}))) { 98 wasm_v128_store32_lane(y, vy, 0); 99 vy = wasm_u64x2_shr(vy, 32); 100 y += 4; 101 } 102 if (n & (2 * sizeof(${XINT8_T}))) { 103 wasm_v128_store16_lane(y, vy, 0); 104 vy = wasm_u32x4_shr(vy, 16); 105 y += 2; 106 } 107 if (n & (1 * sizeof(${XINT8_T}))) { 108 wasm_v128_store8_lane(y, vy, 0); 109 } 110 } 111} 112