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 REQUANTIZATION == "FP32" 8$assert BATCH_TILE >= 1 9#include <assert.h> 10 11#include <fp16.h> 12 13#include <xnnpack/math.h> 14#include <xnnpack/vmul.h> 15 16 17$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 18void xnn_${DATATYPE.lower()}_vmulc_minmax_${REQUANTIZATION.lower()}_ukernel__scalar_x${BATCH_TILE}( 19 size_t n, 20 const ${XINT8_T}* input_a, 21 const ${XINT8_T}* input_b, 22 ${XINT8_T}* output, 23 const union xnn_${DATATYPE.lower()}_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 24{ 25 const int32_t va_zero_point = params->fp32_scalar.a_zero_point; 26 const float vscale = params->fp32_scalar.scale; 27 const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point; 28 const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point; 29 const float vmagic_bias = params->fp32_scalar.magic_bias; 30 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point; 31 32 const int32_t vb = (int32_t) *input_b - params->fp32_scalar.b_zero_point; 33 $if BATCH_TILE == 1: 34 do { 35 const int32_t va = (int32_t) *input_a++ - va_zero_point; 36 const int32_t vacc = va * vb; 37 38 float vfpacc = (float) vacc * vscale; 39 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); 40 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); 41 vfpacc += vmagic_bias; 42 const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point; 43 *output++ = (${XINT8_T}) vout; 44 45 n -= sizeof(${XINT8_T}); 46 } while (n != 0); 47 $else: 48 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 49 $for N in range(BATCH_TILE): 50 const int32_t va${N} = input_a[${N}] - va_zero_point; 51 input_a += ${BATCH_TILE}; 52 53 $for N in range(BATCH_TILE): 54 const int32_t vacc${N} = va${N} * vb; 55 56 $for N in range(BATCH_TILE): 57 float vfpacc${N} = (float) vacc${N} * vscale; 58 59 $for N in range(BATCH_TILE): 60 vfpacc${N} = math_max_f32(vfpacc${N}, voutput_min_less_zero_point); 61 62 $for N in range(BATCH_TILE): 63 vfpacc${N} = math_min_f32(vfpacc${N}, voutput_max_less_zero_point); 64 65 $for N in range(BATCH_TILE): 66 vfpacc${N} += vmagic_bias; 67 68 $for N in range(BATCH_TILE): 69 const int32_t vout${N} = (int32_t) fp32_to_bits(vfpacc${N}) - vmagic_bias_less_output_zero_point; 70 71 $for N in range(BATCH_TILE): 72 output[${N}] = (${XINT8_T}) vout${N}; 73 output += ${BATCH_TILE}; 74 } 75 if XNN_UNLIKELY(n != 0) { 76 $if BATCH_TILE == 2: 77 const int32_t va = (int32_t) *input_a - va_zero_point; 78 const int32_t vacc = va * vb; 79 80 float vfpacc = (float) vacc * vscale; 81 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); 82 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); 83 vfpacc += vmagic_bias; 84 const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point; 85 *output = (${XINT8_T}) vout; 86 $else: 87 do { 88 const int32_t va = (int32_t) *input_a++ - va_zero_point; 89 const int32_t vacc = va * vb; 90 91 float vfpacc = (float) vacc * vscale; 92 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); 93 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); 94 vfpacc += vmagic_bias; 95 const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point; 96 *output++ = (${XINT8_T}) vout; 97 98 n -= sizeof(${XINT8_T}); 99 } while (n != 0); 100 } 101} 102