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$assert DATATYPE in ["QS8", "QU8"] 7$assert BATCH_TILE >= 1 8#include <assert.h> 9 10#include <xnnpack/math.h> 11#include <xnnpack/vadd.h> 12 13 14$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 15void xnn_${DATATYPE.lower()}_vaddc_minmax_ukernel__scalar_x${BATCH_TILE}( 16 size_t n, 17 const ${XINT8_T}* input_a, 18 const ${XINT8_T}* input_b, 19 ${XINT8_T}* output, 20 const union xnn_${DATATYPE.lower()}_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 21{ 22 const int32_t vbias = params->scalar.bias + (int32_t) *input_b * params->scalar.b_multiplier; 23 const int32_t va_multiplier = params->scalar.a_multiplier; 24 const uint32_t vshift = params->scalar.shift; 25 const int32_t voutput_min_less_zero_point = params->scalar.output_min_less_zero_point; 26 const int32_t voutput_max_less_zero_point = params->scalar.output_max_less_zero_point; 27 const int32_t voutput_zero_point = params->scalar.output_zero_point; 28 29 $if BATCH_TILE == 1: 30 do { 31 const int32_t va = *input_a++; 32 const int32_t vacc = vbias + va * va_multiplier; 33 34 int32_t vout = math_asr_s32(vacc, vshift); 35 vout = math_max_s32(vout, voutput_min_less_zero_point); 36 vout = math_min_s32(vout, voutput_max_less_zero_point); 37 *output++ = (${XINT8_T}) (vout + voutput_zero_point); 38 39 n -= sizeof(${XINT8_T}); 40 } while (n != 0); 41 $else: 42 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 43 $for N in range(BATCH_TILE): 44 const int32_t va${N} = input_a[${N}]; 45 input_a += ${BATCH_TILE}; 46 47 $for N in range(BATCH_TILE): 48 const int32_t vacc${N} = vbias + va${N} * va_multiplier; 49 input_b += ${BATCH_TILE}; 50 51 $for N in range(BATCH_TILE): 52 int32_t vout${N} = math_asr_s32(vacc${N}, vshift); 53 54 $for N in range(BATCH_TILE): 55 vout${N} = math_max_s32(vout${N}, voutput_min_less_zero_point); 56 57 $for N in range(BATCH_TILE): 58 vout${N} = math_min_s32(vout${N}, voutput_max_less_zero_point); 59 60 $for N in range(BATCH_TILE): 61 vout${N} += voutput_zero_point; 62 63 $for N in range(BATCH_TILE): 64 output[${N}] = (${XINT8_T}) vout${N}; 65 output += ${BATCH_TILE}; 66 } 67 if XNN_UNLIKELY(n != 0) { 68 $if BATCH_TILE == 2: 69 const int32_t va = *input_a; 70 const int32_t vacc = vbias + va * va_multiplier; 71 72 int32_t vout = math_asr_s32(vacc, vshift); 73 vout = math_max_s32(vout, voutput_min_less_zero_point); 74 vout = math_min_s32(vout, voutput_max_less_zero_point); 75 *output++ = (${XINT8_T}) (vout + voutput_zero_point); 76 $else: 77 do { 78 const int32_t va = *input_a++; 79 const int32_t vacc = vbias + va * va_multiplier; 80 81 int32_t vout = math_asr_s32(vacc, vshift); 82 vout = math_max_s32(vout, voutput_min_less_zero_point); 83 vout = math_min_s32(vout, voutput_max_less_zero_point); 84 *output++ = (${XINT8_T}) (vout + voutput_zero_point); 85 86 n -= sizeof(${XINT8_T}); 87 } while (n != 0); 88 } 89} 90