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$assert DATATYPE in ["QS8", "QU8"] 7$assert BATCH_TILE % 8 == 0 8$assert BATCH_TILE >= 8 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <wasm_simd128.h> 13 14#include <xnnpack/vadd.h> 15 16 17$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 18$WASM_X16X8_LOAD8X8 = {"QS8": "wasm_i16x8_load8x8", "QU8": "wasm_u16x8_load8x8"}[DATATYPE] 19$WASM_X32X4_EXTEND_LOW_X16X8 = {"QS8": "wasm_i32x4_extend_low_i16x8", "QU8": "wasm_u32x4_extend_low_u16x8"}[DATATYPE] 20$WASM_X32X4_EXTEND_HIGH_X16X8 = {"QS8": "wasm_i32x4_extend_high_i16x8", "QU8": "wasm_u32x4_extend_high_u16x8"}[DATATYPE] 21$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE] 22$WASM_X8X16_MIN = {"QS8": "wasm_i8x16_min", "QU8": "wasm_u8x16_min"}[DATATYPE] 23$WASM_X8X16_MAX = {"QS8": "wasm_i8x16_max", "QU8": "wasm_u8x16_max"}[DATATYPE] 24void xnn_${DATATYPE.lower()}_vaddc_minmax_ukernel__wasmsimd_x${BATCH_TILE}( 25 size_t n, 26 const ${XINT8_T}* input_a, 27 const ${XINT8_T}* input_b, 28 ${XINT8_T}* output, 29 const union xnn_${DATATYPE.lower()}_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 30{ 31 const v128_t va_multiplier = wasm_v128_load64_splat(params->wasmsimd.a_multiplier); 32 const uint32_t vshift = params->wasmsimd.shift; 33 const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd.output_zero_point); 34 const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd.output_min); 35 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.output_max); 36 37 v128_t vbias = wasm_i32x4_splat((int32_t) *input_b * params->wasmsimd.b_multiplier[0]); 38 vbias = wasm_i32x4_add(vbias, wasm_v128_load64_splat(params->wasmsimd.bias)); 39 40 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 41 const v128_t va${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_a); 42 $for N in range(8, BATCH_TILE, 8): 43 const v128_t va${ABC[N:N+8]} = ${WASM_X16X8_LOAD8X8}(input_a + ${N}); 44 input_a += ${BATCH_TILE}; 45 46 $for N in range(0, BATCH_TILE, 8): 47 v128_t vacc${ABC[N:N+4]} = wasm_i32x4_add(vbias, wasm_i32x4_mul(${WASM_X32X4_EXTEND_LOW_X16X8}(va${ABC[N:N+8]}), va_multiplier)); 48 v128_t vacc${ABC[N+4:N+8]} = wasm_i32x4_add(vbias, wasm_i32x4_mul(${WASM_X32X4_EXTEND_HIGH_X16X8}(va${ABC[N:N+8]}), va_multiplier)); 49 50 $for N in range(0, BATCH_TILE, 4): 51 vacc${ABC[N:N+4]} = wasm_i32x4_shr(vacc${ABC[N:N+4]}, vshift); 52 53 $for N in range(0, BATCH_TILE, 8): 54 v128_t vout${ABC[N:N+8]} = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point); 55 56 $for N in range(0, BATCH_TILE, 16): 57 $if N + 8 < BATCH_TILE: 58 v128_t vout${ABC[N:N+16]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]}); 59 $else: 60 v128_t vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[N:N+8]}, vout${ABC[N:N+8]}); 61 62 $for N in range(0, BATCH_TILE, 16): 63 $if N + 8 < BATCH_TILE: 64 vout${ABC[N:N+16]} = ${WASM_X8X16_MAX}(vout${ABC[N:N+16]}, voutput_min); 65 $else: 66 vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_MAX}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min); 67 68 $for N in range(0, BATCH_TILE, 16): 69 $if N + 8 < BATCH_TILE: 70 vout${ABC[N:N+16]} = ${WASM_X8X16_MIN}(vout${ABC[N:N+16]}, voutput_max); 71 $else: 72 vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_MIN}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max); 73 74 $if BATCH_TILE >= 16: 75 wasm_v128_store(output, vout${ABC[0:16]}); 76 $else: 77 *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0); 78 $for N in range(16, BATCH_TILE, 16): 79 $if N + 8 < BATCH_TILE: 80 wasm_v128_store(output + ${N}, vout${ABC[N:N+16]}); 81 $else: 82 *((double*) (output + ${N})) = wasm_f64x2_extract_lane(vout${ABC[N:N+8]}${ABC[N:N+8]}, 0); 83 output += ${BATCH_TILE}; 84 } 85 if XNN_UNLIKELY(n != 0) { 86 ${"do " if BATCH_TILE > 8 else ""}{ 87 const v128_t va${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_a); 88 $if BATCH_TILE > 8: 89 input_a += 8; 90 91 v128_t vacc${ABC[0:4]} = wasm_i32x4_add(vbias, wasm_i32x4_mul(${WASM_X32X4_EXTEND_LOW_X16X8}(va${ABC[0:8]}), va_multiplier)); 92 v128_t vacc${ABC[4:8]} = wasm_i32x4_add(vbias, wasm_i32x4_mul(${WASM_X32X4_EXTEND_HIGH_X16X8}(va${ABC[0:8]}), va_multiplier)); 93 94 vacc${ABC[0:4]} = wasm_i32x4_shr(vacc${ABC[0:4]}, vshift); 95 vacc${ABC[4:8]} = wasm_i32x4_shr(vacc${ABC[4:8]}, vshift); 96 97 v128_t vout${ABC[0:8]} = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point); 98 99 v128_t vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[0:8]}, vout${ABC[0:8]}); 100 vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MAX}(vout${ABC[0:8]}${ABC[0:8]}, voutput_min); 101 vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max); 102 103 $if BATCH_TILE > 8: 104 if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) { 105 *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0); 106 output += 8; 107 n -= 8 * sizeof(${XINT8_T}); 108 } else { 109 if (n & (4 * sizeof(${XINT8_T}))) { 110 *((float*) output) = (float) wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0); 111 vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32); 112 output += 4; 113 } 114 uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0); 115 if (n & (2 * sizeof(${XINT8_T}))) { 116 *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]}; 117 vout${ABC[0:4]} >>= 16; 118 output += 2; 119 } 120 if (n & (1 * sizeof(${XINT8_T}))) { 121 *output = (${XINT8_T}) vout${ABC[0:4]}; 122 } 123 n = 0; 124 } 125 $else: 126 if (n & (4 * sizeof(${XINT8_T}))) { 127 *((float*) output) = (float) wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0); 128 vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32); 129 output += 4; 130 } 131 uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0); 132 if (n & (2 * sizeof(${XINT8_T}))) { 133 *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]}; 134 vout${ABC[0:4]} >>= 16; 135 output += 2; 136 } 137 if (n & (1 * sizeof(${XINT8_T}))) { 138 *output = (${XINT8_T}) vout${ABC[0:4]}; 139 } 140 }${" while (n != 0);" if BATCH_TILE > 8 else ""} 141 } 142} 143