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()}_vadd_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; 23 const int32_t va_multiplier = params->scalar.a_multiplier; 24 const int32_t vb_multiplier = params->scalar.b_multiplier; 25 const uint32_t vshift = params->scalar.shift; 26 const int32_t voutput_min_less_zero_point = params->scalar.output_min_less_zero_point; 27 const int32_t voutput_max_less_zero_point = params->scalar.output_max_less_zero_point; 28 const int32_t voutput_zero_point = params->scalar.output_zero_point; 29 30 $if BATCH_TILE == 1: 31 do { 32 const int32_t va = *input_a++; 33 const int32_t vb = *input_b++; 34 const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier; 35 36 int32_t vout = math_asr_s32(vacc, vshift); 37 vout = math_max_s32(vout, voutput_min_less_zero_point); 38 vout = math_min_s32(vout, voutput_max_less_zero_point); 39 *output++ = (${XINT8_T}) (vout + voutput_zero_point); 40 41 n -= sizeof(${XINT8_T}); 42 } while (n != 0); 43 $else: 44 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 45 $for N in range(BATCH_TILE): 46 const int32_t va${N} = input_a[${N}]; 47 input_a += ${BATCH_TILE}; 48 49 $for N in range(BATCH_TILE): 50 const int32_t vb${N} = input_b[${N}]; 51 int32_t vacc${N} = vbias + va${N} * va_multiplier; 52 input_b += ${BATCH_TILE}; 53 54 $for N in range(BATCH_TILE): 55 vacc${N} += vb${N} * vb_multiplier; 56 57 $for N in range(BATCH_TILE): 58 int32_t vout${N} = math_asr_s32(vacc${N}, vshift); 59 60 $for N in range(BATCH_TILE): 61 vout${N} = math_max_s32(vout${N}, voutput_min_less_zero_point); 62 63 $for N in range(BATCH_TILE): 64 vout${N} = math_min_s32(vout${N}, voutput_max_less_zero_point); 65 66 $for N in range(BATCH_TILE): 67 vout${N} += voutput_zero_point; 68 69 $for N in range(BATCH_TILE): 70 output[${N}] = (${XINT8_T}) vout${N}; 71 output += ${BATCH_TILE}; 72 } 73 if XNN_UNLIKELY(n != 0) { 74 $if BATCH_TILE == 2: 75 const int32_t va = *input_a; 76 const int32_t vb = *input_b; 77 const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier; 78 79 int32_t vout = math_asr_s32(vacc, vshift); 80 vout = math_max_s32(vout, voutput_min_less_zero_point); 81 vout = math_min_s32(vout, voutput_max_less_zero_point); 82 *output++ = (${XINT8_T}) (vout + voutput_zero_point); 83 $else: 84 do { 85 const int32_t va = *input_a++; 86 const int32_t vb = *input_b++; 87 const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier; 88 89 int32_t vout = math_asr_s32(vacc, vshift); 90 vout = math_max_s32(vout, voutput_min_less_zero_point); 91 vout = math_min_s32(vout, voutput_max_less_zero_point); 92 *output++ = (${XINT8_T}) (vout + voutput_zero_point); 93 94 n -= sizeof(${XINT8_T}); 95 } while (n != 0); 96 } 97} 98