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$assert BATCH_TILE >= 1 9#include <assert.h> 10 11#include <xnnpack/math.h> 12#include <xnnpack/vmul.h> 13 14 15$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 16void xnn_${DATATYPE.lower()}_vmul_minmax_${REQUANTIZATION.lower()}_ukernel__scalar_x${BATCH_TILE}( 17 size_t n, 18 const ${XINT8_T}* input_a, 19 const ${XINT8_T}* input_b, 20 ${XINT8_T}* output, 21 const union xnn_${DATATYPE.lower()}_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 22{ 23 const int32_t va_zero_point = params->fp32_scalar.a_zero_point; 24 const int32_t vb_zero_point = params->fp32_scalar.b_zero_point; 25 const float vscale = params->fp32_scalar.scale; 26 const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point; 27 const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point; 28 const float vmagic_bias = params->fp32_scalar.magic_bias; 29 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point; 30 31 $if BATCH_TILE == 1: 32 do { 33 const int32_t va = (int32_t) *input_a++ - va_zero_point; 34 const int32_t vb = (int32_t) *input_b++ - vb_zero_point; 35 const int32_t vacc = va * vb; 36 37 float vfpacc = (float) vacc * vscale; 38 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); 39 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); 40 vfpacc += vmagic_bias; 41 const int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; 42 *output++ = (${XINT8_T}) vout; 43 44 n -= sizeof(${XINT8_T}); 45 } while (n != 0); 46 $else: 47 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 48 $for N in range(BATCH_TILE): 49 const int32_t va${N} = input_a[${N}] - va_zero_point; 50 input_a += ${BATCH_TILE}; 51 52 $for N in range(BATCH_TILE): 53 const int32_t vb${N} = input_b[${N}] - vb_zero_point; 54 input_b += ${BATCH_TILE}; 55 56 $for N in range(BATCH_TILE): 57 const int32_t vacc${N} = va${N} * vb${N}; 58 59 $for N in range(BATCH_TILE): 60 float vfpacc${N} = (float) vacc${N} * vscale; 61 62 $for N in range(BATCH_TILE): 63 vfpacc${N} = math_max_f32(vfpacc${N}, voutput_min_less_zero_point); 64 65 $for N in range(BATCH_TILE): 66 vfpacc${N} = math_min_f32(vfpacc${N}, voutput_max_less_zero_point); 67 68 $for N in range(BATCH_TILE): 69 vfpacc${N} += vmagic_bias; 70 71 $for N in range(BATCH_TILE): 72 const int32_t vout${N} = (int32_t) float_as_uint32(vfpacc${N}) - vmagic_bias_less_output_zero_point; 73 74 $for N in range(BATCH_TILE): 75 output[${N}] = (${XINT8_T}) vout${N}; 76 output += ${BATCH_TILE}; 77 } 78 if XNN_UNLIKELY(n != 0) { 79 $if BATCH_TILE == 2: 80 const int32_t va = (int32_t) *input_a - va_zero_point; 81 const int32_t vb = (int32_t) *input_b - vb_zero_point; 82 const int32_t vacc = va * vb; 83 84 float vfpacc = (float) vacc * vscale; 85 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); 86 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); 87 vfpacc += vmagic_bias; 88 const int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; 89 *output = (${XINT8_T}) vout; 90 $else: 91 do { 92 const int32_t va = (int32_t) *input_a++ - va_zero_point; 93 const int32_t vb = (int32_t) *input_b++ - vb_zero_point; 94 const int32_t vacc = va * vb; 95 96 float vfpacc = (float) vacc * vscale; 97 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); 98 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); 99 vfpacc += vmagic_bias; 100 const int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; 101 *output++ = (${XINT8_T}) vout; 102 103 n -= sizeof(${XINT8_T}); 104 } while (n != 0); 105 } 106} 107