Embedded/pikapython
1
0
Fork 0
mirror of https://gitee.com/Lyon1998/pikapython.git synced 2025-01-29 17:22:56 +08:00
Code Issues Packages Projects Releases Wiki Activity
pikapython/package/PikaNN/TinyMaix/src/arch_rv64v.h

375 lines
14 KiB
C
Raw Normal View History

fix valgrind in check CLS add pikann
2022-09-08 22:20:06 +08:00
/* Copyright 2022 Sipeed Technology Co., Ltd. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "stdlib.h"
#include "stdint.h"
#include "math.h"
#include "tm_port.h"
#ifndef __riscv_vector
#error "Vector intrinsics require the vector extension."
#endif
#include <riscv_vector.h>
//RV64 V-extend acceleration
//https://github.com/T-head-Semi/openc906/tree/main/doc
//https://github.com/riscv-non-isa/rvv-intrinsic-doc
//https://occ.t-head.cn/community/download?id=4073475960903634944
//https://github.com/riscv/riscv-v-spec
//Use this toolchain support rvv inntrinsic: Xuantie-900-gcc-elf-newlib-x86_64-V2.2.4-20211227.tar.gz
//-march=rv64gcv0p7_zfh_xtheadc
//more information refer to:
//https://occ-oss-prod.oss-cn-hangzhou.aliyuncs.com/resource//1659318003104/Xuantie+900+Series+RVV-0.7.1+Intrinsic+Manual.pdf
//https://riscv.org/wp-content/uploads/2019/06/17.40-Vector_RISCV-20190611-Vectors.pdf
//Total 32 vector registers
//ELEN: Size of largest element in bits
//VLEN: Number of bits in each vector register; VLEN>=ELEN
//SLEN: Striping distance in bits; VLEN >= SLEN >= ELEN
//SEW: Standard element width, =8,16,32,...,1024
//LMUL: Vector length multiplier, =1,2,4,8 (LMUL registers at the same time)
//EDIV: Vector element divider, =1,2,4,8
//vtype: Vtype sets width of element in each vector register (e.g., 16-bit, 32-bit, …)
//vl: Vector length CSR sets number of elements active in each instruction
//vstart: Resumption element after trap
//fcsr(vxrm/vxsat): Fixed-point rounding mode and saturation flag fields in FP CSR
#define RVV_VLEN 128 //TODO: read csrr_vlenb
/********************************** TM_MDL_FP32 ***********************************************/
#if TM_MDL_TYPE==TM_MDL_FP32
#define PACK_N (RVV_VLEN/32) //fp32
uint32_t tdot = 0;
TM_INLINE void tm_dot_prod(mtype_t* sptr, mtype_t* kptr,uint32_t size, sumtype_t* result)
{
float sumbuf[PACK_N];
float sum = 0.f;
int cnt=size/PACK_N;
//uint64_t t0,t1;
//__ASM volatile("csrr %0, mcycle" : "=r"(t0));
if(cnt>0){
size_t vl = vsetvl_e32m1(PACK_N);
vfloat32m1_t sumv = vfmv_v_f_f32m1(0.f, vl); //set sum=0
vfloat32m1_t v_zero = vfmv_v_f_f32m1(0.f, vl);
for(int i=0; i<cnt; i++){
vfloat32m1_t s = vle32_v_f32m1(sptr, vl);
vfloat32m1_t k = vle32_v_f32m1(kptr, vl);
sumv = vfmacc_vv_f32m1(sumv, s, k, vl);
sptr += PACK_N;
kptr += PACK_N;
}
v_zero = vfredosum_vs_f32m1_f32m1(v_zero, sumv, v_zero, vl);
sum = vfmv_f_s_f32m1_f32(v_zero);
}
for(int i=0; i<size%PACK_N; i++){
sum += *sptr * *kptr;
sptr++;kptr++;
}
*result = sum;
//__ASM volatile("csrr %0, mcycle" : "=r"(t1));
//printf("size=%3d, dt=%ld\r\n", size, t1-t0);
//tdot+=(t1-t0);
return;
}
TM_INLINE void tm_dot_prod_pack2(mtype_t* sptr, mtype_t* kptr, uint32_t size, sumtype_t* result)
{
float sumbuf[PACK_N];
mtype_t* kptr0 = kptr;
mtype_t* kptr1 = kptr+size;
//mtype_t* kptr2 = kptr+size*2;
//mtype_t* kptr3 = kptr+size*3;
float sum0 = 0;
float sum1 = 0;
//float sum2 = 0;
//float sum3 = 0;
int cnt=size/PACK_N;
if(cnt>0){
size_t vl = vsetvl_e32m1(PACK_N);
vfloat32m1_t sumv0 = vfmv_v_f_f32m1(0.f, vl); //set sum=0
vfloat32m1_t sumv1 = vfmv_v_f_f32m1(0.f, vl); //set sum=0
//vfloat32m1_t sumv2 = vfmv_v_f_f32m1(0.f, vl); //set sum=0
//vfloat32m1_t sumv3 = vfmv_v_f_f32m1(0.f, vl); //set sum=0
vfloat32m1_t v_zero0 = vfmv_v_f_f32m1(0.0f, vl);
vfloat32m1_t v_zero1 = vfmv_v_f_f32m1(0.0f, vl);
//vfloat32m1_t v_zero2 = vfmv_v_f_f32m1(0.0f, vl);
//vfloat32m1_t v_zero3 = vfmv_v_f_f32m1(0.0f, vl);
for(int i=0; i<cnt; i++){
vfloat32m1_t s = vle32_v_f32m1(sptr, vl);
vfloat32m1_t k0 = vle32_v_f32m1(kptr0, vl);
vfloat32m1_t k1 = vle32_v_f32m1(kptr1, vl);
//vfloat32m1_t k2 = vle32_v_f32m1(kptr2, vl);
//vfloat32m1_t k3 = vle32_v_f32m1(kptr3, vl);
sumv0 = vfmacc_vv_f32m1(sumv0, s, k0, vl);
sumv1 = vfmacc_vv_f32m1(sumv1, s, k1, vl);
//sumv2 = vfmacc_vv_f32m1(sumv2, s, k2, vl);
//sumv3 = vfmacc_vv_f32m1(sumv3, s, k3, vl);
sptr += PACK_N;
kptr0 += PACK_N;
kptr1 += PACK_N;
//kptr2 += PACK_N;
//kptr3 += PACK_N;
}
v_zero0 = vfredosum_vs_f32m1_f32m1(v_zero0, sumv0, v_zero0, vl);
v_zero1 = vfredosum_vs_f32m1_f32m1(v_zero1, sumv1, v_zero1, vl);
//v_zero2 = vfredosum_vs_f32m1_f32m1(v_zero2, sumv2, v_zero2, vl);
//v_zero3 = vfredosum_vs_f32m1_f32m1(v_zero3, sumv3, v_zero3, vl);
sum0 = vfmv_f_s_f32m1_f32(v_zero0);
sum1 = vfmv_f_s_f32m1_f32(v_zero1);
//sum2 = vfmv_f_s_f32m1_f32(v_zero2);
//sum3 = vfmv_f_s_f32m1_f32(v_zero3);
}
for(int i=0; i<size%PACK_N; i++){
sum0 += *sptr * *kptr0;
sum1 += *sptr * *kptr1;
//sum2 += *sptr * *kptr2;
//sum3 += *sptr * *kptr3;
sptr++;kptr0++;kptr1++;//kptr2++;kptr3++;
}
result[0] = sum0;
result[1] = sum1;
//result[2] = sum2;
//result[3] = sum3;
return;
}
TM_INLINE void tm_dot_prod_gap_3x3x1(mtype_t* sptr, mtype_t* kptr, uint32_t* k_oft, sumtype_t* result)
{
*result = sptr[k_oft[0]]*kptr[0] + sptr[k_oft[1]]*kptr[1] + sptr[k_oft[2]]*kptr[2] + \
sptr[k_oft[3]]*kptr[3] + sptr[k_oft[4]]*kptr[4] + sptr[k_oft[5]]*kptr[5] + \
sptr[k_oft[6]]*kptr[6] + sptr[k_oft[7]]*kptr[7] + sptr[k_oft[8]]*kptr[8] ;
return;
}
TM_INLINE void tm_dot_prod_3x3x1(mtype_t* sptr, mtype_t* kptr, sumtype_t* result)
{
*result = sptr[0]*kptr[0] + sptr[1]*kptr[1] + sptr[2]*kptr[2] + \
sptr[3]*kptr[3] + sptr[4]*kptr[4] + sptr[5]*kptr[5] + \
sptr[6]*kptr[6] + sptr[7]*kptr[7] + sptr[8]*kptr[8] ;
return;
}
/********************************** TM_MDL_FP16 ***********************************************/
#elif TM_MDL_TYPE==TM_MDL_FP16
#define PACK_N (RVV_VLEN/16) //fp16, packn=8
TM_INLINE void tm_dot_prod(mtype_t* sptr, mtype_t* kptr,uint32_t size, sumtype_t* result)
{
float16_t sumbuf[PACK_N];
float16_t sum = 0;
int cnt=size/PACK_N;
//uint64_t t0,t1;
//__ASM volatile("csrr %0, mcycle" : "=r"(t0));
if(cnt>0){
size_t vl = vsetvl_e16m1(PACK_N);
vfloat16m1_t sumv = vfmv_v_f_f16m1(0.f, vl); //set sum=0
vfloat16m1_t v_zero = vfmv_v_f_f16m1(0.0f, vl);
for(int i=0; i<cnt; i++){
vfloat16m1_t s = vle16_v_f16m1(sptr, vl);
vfloat16m1_t k = vle16_v_f16m1(kptr, vl);
sumv = vfmacc_vv_f16m1(sumv, s, k, vl);
sptr += PACK_N;
kptr += PACK_N;
}
v_zero = vfredosum_vs_f16m1_f16m1(v_zero, sumv, v_zero, vl);
sum = vfmv_f_s_f16m1_f16(v_zero);
}
for(int i=0; i<size%PACK_N; i++){
sum += *sptr * *kptr;
sptr++;kptr++;
}
*result = sum;
return;
}
TM_INLINE void tm_dot_prod_pack2(mtype_t* sptr, mtype_t* kptr, uint32_t size, sumtype_t* result)
{
float16_t sumbuf[PACK_N];
mtype_t* kptr0 = kptr;
mtype_t* kptr1 = kptr+size;
//mtype_t* kptr2 = kptr+size*2;
//mtype_t* kptr3 = kptr+size*3;
float16_t sum0 = 0;
float16_t sum1 = 0;
//float16_t sum2 = 0;
//float16_t sum3 = 0;
int cnt=size/PACK_N;
if(cnt>0){
size_t vl = vsetvl_e16m1(PACK_N);
vfloat16m1_t sumv0 = vfmv_v_f_f16m1(0.f, vl); //set sum=0
vfloat16m1_t sumv1 = vfmv_v_f_f16m1(0.f, vl); //set sum=0
//vfloat16m1_t sumv2 = vfmv_v_f_f16m1(0.f, vl); //set sum=0
//vfloat16m1_t sumv3 = vfmv_v_f_f16m1(0.f, vl); //set sum=0
vfloat16m1_t v_zero0 = vfmv_v_f_f16m1(0.0f, vl);
vfloat16m1_t v_zero1 = vfmv_v_f_f16m1(0.0f, vl);
//vfloat16m1_t v_zero2 = vfmv_v_f_f16m1(0.0f, vl);
//vfloat16m1_t v_zero3 = vfmv_v_f_f16m1(0.0f, vl);
for(int i=0; i<cnt; i++){
vfloat16m1_t s = vle16_v_f16m1(sptr, vl);
vfloat16m1_t k0 = vle16_v_f16m1(kptr0, vl);
vfloat16m1_t k1 = vle16_v_f16m1(kptr1, vl);
//vfloat16m1_t k2 = vle16_v_f16m1(kptr2, vl);
//vfloat16m1_t k3 = vle16_v_f16m1(kptr3, vl);
sumv0 = vfmacc_vv_f16m1(sumv0, s, k0, vl);
sumv1 = vfmacc_vv_f16m1(sumv1, s, k1, vl);
//sumv2 = vfmacc_vv_f16m1(sumv2, s, k2, vl);
//sumv3 = vfmacc_vv_f16m1(sumv3, s, k3, vl);
sptr += PACK_N;
kptr0 += PACK_N;
kptr1 += PACK_N;
//kptr2 += PACK_N;
//kptr3 += PACK_N;
}
v_zero0 = vfredosum_vs_f16m1_f16m1(v_zero0, sumv0, v_zero0, vl);
v_zero1 = vfredosum_vs_f16m1_f16m1(v_zero1, sumv1, v_zero1, vl);
//v_zero2 = vfredosum_vs_f16m1_f16m1(v_zero2, sumv2, v_zero2, vl);
//v_zero3 = vfredosum_vs_f16m1_f16m1(v_zero3, sumv3, v_zero3, vl);
sum0 = vfmv_f_s_f16m1_f16(v_zero0);
sum1 = vfmv_f_s_f16m1_f16(v_zero1);
//sum2 = vfmv_f_s_f16m1_f16(v_zero2);
//sum3 = vfmv_f_s_f16m1_f16(v_zero3);
}
for(int i=0; i<size%PACK_N; i++){
sum0 += *sptr * *kptr0;
sum1 += *sptr * *kptr1;
//sum2 += *sptr * *kptr2;
//sum3 += *sptr * *kptr3;
sptr++;kptr0++;kptr1++;//kptr2++;kptr3++;
}
result[0] = sum0;
result[1] = sum1;
//result[2] = sum2;
//result[3] = sum3;
return;
}
TM_INLINE void tm_dot_prod_gap_3x3x1(mtype_t* sptr, mtype_t* kptr, uint32_t* k_oft, sumtype_t* result)
{
*result = sptr[k_oft[0]]*kptr[0] + sptr[k_oft[1]]*kptr[1] + sptr[k_oft[2]]*kptr[2] + \
sptr[k_oft[3]]*kptr[3] + sptr[k_oft[4]]*kptr[4] + sptr[k_oft[5]]*kptr[5] + \
sptr[k_oft[6]]*kptr[6] + sptr[k_oft[7]]*kptr[7] + sptr[k_oft[8]]*kptr[8] ;
return;
}
TM_INLINE void tm_dot_prod_3x3x1(mtype_t* sptr, mtype_t* kptr, sumtype_t* result)
{
*result = sptr[0]*kptr[0] + sptr[1]*kptr[1] + sptr[2]*kptr[2] + \
sptr[3]*kptr[3] + sptr[4]*kptr[4] + sptr[5]*kptr[5] + \
sptr[6]*kptr[6] + sptr[7]*kptr[7] + sptr[8]*kptr[8] ;
return;
}
/********************************** TM_MDL_INT8 ***********************************************/
#elif TM_MDL_TYPE==TM_MDL_INT8
#define PACK_N (RVV_VLEN/8) //int8 16
//i8 to i16 (Widening Floating-Point/Integer Type-Convert Operations)
// vint16m2_t vwcvt_x_x_v_i16m2 (vint8m1_t src, size_t vl);
//Vector Widening Integer Multiply-Add Operations
// vint32m4_t vwmacc_vv_i32m4 (vint32m4_t vd, vint16m2_t vs1, vint16m2_t vs2, size_t vl);
uint32_t tdot = 0;
static uint32_t size0=0;
TM_INLINE void tm_dot_prod(mtype_t* sptr, mtype_t* kptr,uint32_t size, sumtype_t* result)
{
int32_t sumbuf[PACK_N];
int32_t sum = 0;
int cnt=size/PACK_N;
if(cnt>0){
size_t vl = vsetvl_e8m1(PACK_N);
vint32m4_t sumv = vmv_v_x_i32m4(0, vl); //set sum=0
vint32m1_t v_zero = vmv_v_x_i32m1(0, vl);
for(int i=0; i<cnt; i++){
vint8m1_t s8 = vle8_v_i8m1(sptr, vl); //load i8
vint8m1_t k8 = vle8_v_i8m1(kptr, vl);
vint16m2_t s16 = vwcvt_x_x_v_i16m2(s8,vl); //cvt i8 to i16
vint16m2_t k16 = vwcvt_x_x_v_i16m2(k8,vl);
sumv = vwmacc_vv_i32m4(sumv, s16, k16, vl); //mac
sptr += PACK_N;
kptr += PACK_N;
}
v_zero = vredsum_vs_i32m4_i32m1(v_zero, sumv, v_zero, vl);
sum = vmv_x_s_i32m1_i32(v_zero);
}
for(int i=0; i<size%PACK_N; i++){
sum += *sptr * *kptr;
sptr++;kptr++;
}
*result = sum;
return;
}
TM_INLINE void tm_dot_prod_pack2(mtype_t* sptr, mtype_t* kptr, uint32_t size, sumtype_t* result)
{
int32_t sumbuf[PACK_N];
int32_t sum0 = 0;
int32_t sum1 = 0;
mtype_t* kptr0 = kptr;
mtype_t* kptr1 = kptr+size;
int cnt=size/PACK_N;
if(cnt>0){
size_t vl = vsetvl_e8m1(PACK_N);
vint32m4_t sumv0 = vmv_v_x_i32m4(0, vl); //set sum=0
vint32m4_t sumv1 = vmv_v_x_i32m4(0, vl); //set sum=0
vint32m1_t v_zero0 = vmv_v_x_i32m1(0, vl);
vint32m1_t v_zero1 = vmv_v_x_i32m1(0, vl);
for(int i=0; i<cnt; i++){
vint8m1_t s8 = vle8_v_i8m1(sptr, vl); //load i8
vint8m1_t k80 = vle8_v_i8m1(kptr0, vl);
vint8m1_t k81 = vle8_v_i8m1(kptr1, vl);
vint16m2_t s16 = vwcvt_x_x_v_i16m2(s8,vl); //cvt i8 to i16
vint16m2_t k160 = vwcvt_x_x_v_i16m2(k80,vl);
vint16m2_t k161 = vwcvt_x_x_v_i16m2(k81,vl);
sumv0 = vwmacc_vv_i32m4(sumv0, s16, k160, vl); //mac
sumv1 = vwmacc_vv_i32m4(sumv1, s16, k161, vl); //mac
sptr += PACK_N;
kptr0 += PACK_N;
kptr1 += PACK_N;
}
v_zero0 = vredsum_vs_i32m4_i32m1(v_zero0, sumv0, v_zero0, vl);
v_zero1 = vredsum_vs_i32m4_i32m1(v_zero1, sumv1, v_zero1, vl);
sum0 = vmv_x_s_i32m1_i32(v_zero0);
sum1 = vmv_x_s_i32m1_i32(v_zero1);
}
for(int i=0; i<size%PACK_N; i++){
sum0 += *sptr * *kptr0;
sum1 += *sptr * *kptr1;
sptr++;kptr0++;kptr1++;
}
result[0] = sum0;
result[1] = sum1;
return;
}
TM_INLINE void tm_dot_prod_gap_3x3x1(mtype_t* sptr, mtype_t* kptr, uint32_t* k_oft, sumtype_t* result)
{
*result = sptr[k_oft[0]]*kptr[0] + sptr[k_oft[1]]*kptr[1] + sptr[k_oft[2]]*kptr[2] + \
sptr[k_oft[3]]*kptr[3] + sptr[k_oft[4]]*kptr[4] + sptr[k_oft[5]]*kptr[5] + \
sptr[k_oft[6]]*kptr[6] + sptr[k_oft[7]]*kptr[7] + sptr[k_oft[8]]*kptr[8] ;
return;
}
TM_INLINE void tm_dot_prod_3x3x1(mtype_t* sptr, mtype_t* kptr, sumtype_t* result)
{
*result = sptr[0]*kptr[0] + sptr[1]*kptr[1] + sptr[2]*kptr[2] + \
sptr[3]*kptr[3] + sptr[4]*kptr[4] + sptr[5]*kptr[5] + \
sptr[6]*kptr[6] + sptr[7]*kptr[7] + sptr[8]*kptr[8] ;
return;
}
#else
#error "ERR MDL TYPE"
#endif
Reference in New Issue Copy Permalink