wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fixed lmp_mult parameters, ram init filepath

Browse Source
This commit is contained in:
Blaise Tine
2020-09-04 07:51:46 -07:00
parent dccea80b68
commit 42e3b6c45d
36 changed files with 738 additions and 495 deletions
Download Patch File Download Diff File Expand all files Collapse all files

185
hw/rtl/fp_cores/svdpi/float_dpi.cpp
View File

@@ -8,21 +8,19 @@
#include "VX_config.h"
extern "C" {
void dpi_fadd(int inst, bool enable, bool valid, int a, int b, int* result);
void dpi_fsub(int inst, bool enable, bool valid, int a, int b, int* result);
void dpi_fmul(int inst, bool enable, bool valid, int a, int b, int* result);
void dpi_fmadd(int inst, bool enable, bool valid, int a, int b, int c, int* result);
void dpi_fmsub(int inst, bool enable, bool valid, int a, int b, int c, int* result);
void dpi_fdiv(int inst, bool enable, bool valid, int a, int b, int* result);
void dpi_fsqrt(int inst, bool enable, bool valid, int a, int* result);
void dpi_ftoi(int inst, bool enable, bool valid, int a, int* result);
void dpi_ftou(int inst, bool enable, bool valid, int a, int* result);
void dpi_itof(int inst, bool enable, bool valid, int a, int* result);
void dpi_utof(int inst, bool enable, bool valid, int a, int* result);
void dpi_fadd(int inst, bool enable, int a, int b, int* result);
void dpi_fsub(int inst, bool enable, int a, int b, int* result);
void dpi_fmul(int inst, bool enable, int a, int b, int* result);
void dpi_fmadd(int inst, bool enable, int a, int b, int c, int* result);
void dpi_fmsub(int inst, bool enable, int a, int b, int c, int* result);
void dpi_fdiv(int inst, bool enable, int a, int b, int* result);
void dpi_fsqrt(int inst, bool enable, int a, int* result);
void dpi_ftoi(int inst, bool enable, int a, int* result);
void dpi_ftou(int inst, bool enable, int a, int* result);
void dpi_itof(int inst, bool enable, int a, int* result);
void dpi_utof(int inst, bool enable, int a, int* result);
}
extern double sc_time_stamp();
class ShiftRegister {
public:
ShiftRegister() : init_(false), depth_(0) {}
@@ -35,37 +33,36 @@ public:
}
}
void push(int value, bool enable, bool valid) {
void push(int value, bool enable) {
if (!enable)
return;
for (unsigned i = 0; i < depth_-1; ++i) {
buffer_[i] = buffer_[i+1];
}
buffer_[depth_-1].value = value;
buffer_[depth_-1].valid = valid;
buffer_[depth_-1] = value;
}
int top() const {
return buffer_[0].value;
}
bool valid() const {
return buffer_[0].valid;
return buffer_[0];
}
private:
struct entry_t {
int value;
bool valid;
};
std::vector<entry_t> buffer_;
int top_;
unsigned depth_;
std::vector<int> buffer_;
bool init_;
unsigned depth_;
};
union Float_t {
float f;
int i;
struct {
uint32_t man : 23;
uint32_t exp : 8;
uint32_t sign : 1;
} parts;
};
class Instances {
public:
ShiftRegister& get(int inst) {
@@ -82,130 +79,152 @@ private:
Instances instances;
void dpi_fadd(int inst, bool enable, bool valid, int a, int b, int* result) {
void dpi_fadd(int inst, bool enable, int a, int b, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fb = *(float*)&b;
float fr = fa + fb;
Float_t fa, fb, fr;
sr.ensure_init(LATENCY_FMADD);
sr.push(*(int*)&fr, enable, valid);
fa.i = a;
fb.i = b;
fr.f = fa.f + fb.f;
sr.ensure_init(LATENCY_FADDMUL);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_fsub(int inst, bool enable, bool valid, int a, int b, int* result) {
void dpi_fsub(int inst, bool enable, int a, int b, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fb = *(float*)&b;
float fr = fa - fb;
Float_t fa, fb, fr;
sr.ensure_init(LATENCY_FMADD);
sr.push(*(int*)&fr, enable, valid);
fa.i = a;
fb.i = b;
fr.f = fa.f - fb.f;
sr.ensure_init(LATENCY_FADDMUL);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_fmul(int inst, bool enable, bool valid, int a, int b, int* result) {
void dpi_fmul(int inst, bool enable, int a, int b, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fb = *(float*)&b;
float fr = fa * fb;
Float_t fa, fb, fr;
sr.ensure_init(LATENCY_FMADD);
sr.push(*(int*)&fr, enable, valid);
fa.i = a;
fb.i = b;
fr.f = fa.f * fb.f;
sr.ensure_init(LATENCY_FADDMUL);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_fmadd(int inst, bool enable, bool valid, int a, int b, int c, int* result) {
void dpi_fmadd(int inst, bool enable, int a, int b, int c, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fb = *(float*)&b;
float fc = *(float*)&c;
float fr = fa * fb + fc;
Float_t fa, fb, fc, fr;
fa.i = a;
fb.i = b;
fc.i = c;
fr.f = fa.f * fb.f + fc.f;
sr.ensure_init(LATENCY_FMADD);
sr.push(*(int*)&fr, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_fmsub(int inst, bool enable, bool valid, int a, int b, int c, int* result) {
void dpi_fmsub(int inst, bool enable, int a, int b, int c, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fb = *(float*)&b;
float fc = *(float*)&c;
float fr = fa * fb - fc;
Float_t fa, fb, fc, fr;
fa.i = a;
fb.i = b;
fc.i = c;
fr.f = fa.f * fb.f - fc.f;
sr.ensure_init(LATENCY_FMADD);
sr.push(*(int*)&fr, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_fdiv(int inst, bool enable, bool valid, int a, int b, int* result) {
void dpi_fdiv(int inst, bool enable, int a, int b, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fb = *(float*)&b;
float fr = fa / fb;
Float_t fa, fb, fr;
fa.i = a;
fb.i = b;
fr.f = fa.f / fb.f;
sr.ensure_init(LATENCY_FDIV);
sr.push(*(int*)&fr, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_fsqrt(int inst, bool enable, bool valid, int a, int* result) {
void dpi_fsqrt(int inst, bool enable, int a, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
float fr = sqrtf(fa);
Float_t fa, fr;
fa.i = a;
fr.f = sqrtf(fa.f);
sr.ensure_init(LATENCY_FSQRT);
sr.push(*(int*)&fr, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_ftoi(int inst, bool enable, bool valid, int a, int* result) {
void dpi_ftoi(int inst, bool enable, int a, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
int ir = int(fa);
Float_t fa, fr;
fa.i = a;
fr.i = int(fa.f);
sr.ensure_init(LATENCY_FTOI);
sr.push(ir, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_ftou(int inst, bool enable, bool valid, int a, int* result) {
void dpi_ftou(int inst, bool enable, int a, int* result) {
ShiftRegister& sr = instances.get(inst);
float fa = *(float*)&a;
unsigned ir = unsigned(fa);
Float_t fa, fr;
fa.i = a;
fr.i = unsigned(fa.f);
sr.ensure_init(LATENCY_FTOI);
sr.push(ir, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_itof(int inst, bool enable, bool valid, int a, int* result) {
void dpi_itof(int inst, bool enable, int a, int* result) {
ShiftRegister& sr = instances.get(inst);
float fr = (float)a;
Float_t fa, fr;
fr.f = (float)a;
sr.ensure_init(LATENCY_ITOF);
sr.push(*(int*)&fr, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}
void dpi_utof(int inst, bool enable, bool valid, int a, int* result) {
void dpi_utof(int inst, bool enable, int a, int* result) {
ShiftRegister& sr = instances.get(inst);
unsigned ua = *(unsigned*)&a;
float fr = (float)ua;
Float_t fa, fr;
unsigned ua = a;
fr.f = (float)ua;
sr.ensure_init(LATENCY_ITOF);
sr.push(*(int*)&fr, enable, valid);
sr.push(fr.i, enable);
*result = sr.top();
}

22
hw/rtl/fp_cores/svdpi/float_dpi.vh
View File

@@ -1,16 +1,16 @@
`ifndef FLOAT_DPI
`define FLOAT_DPI
import "DPI-C" context function void dpi_fadd(int inst, input logic enable, input logic valid, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fsub(int inst, input logic enable, input logic valid, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fmul(int inst, input logic enable, input logic valid, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fmadd(int inst, input logic enable, input logic valid, input int a, input int b, input int c, output int result);
import "DPI-C" context function void dpi_fmsub(int inst, input logic enable, input logic valid, input int a, input int b, input int c, output int result);
import "DPI-C" context function void dpi_fdiv(int inst, input logic enable, input logic valid, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fsqrt(int inst, input logic enable, input logic valid, input int a, output int result);
import "DPI-C" context function void dpi_ftoi(int inst, input logic enable, input logic valid, input int a, output int result);
import "DPI-C" context function void dpi_ftou(int inst, input logic enable, input logic valid, input int a, output int result);
import "DPI-C" context function void dpi_itof(int inst, input logic enable, input logic valid, input int a, output int result);
import "DPI-C" context function void dpi_utof(int inst, input logic enable, input logic valid, input int a, output int result);
import "DPI-C" context function void dpi_fadd(int inst, input logic enable, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fsub(int inst, input logic enable, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fmul(int inst, input logic enable, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fmadd(int inst, input logic enable, input int a, input int b, input int c, output int result);
import "DPI-C" context function void dpi_fmsub(int inst, input logic enable, input int a, input int b, input int c, output int result);
import "DPI-C" context function void dpi_fdiv(int inst, input logic enable, input int a, input int b, output int result);
import "DPI-C" context function void dpi_fsqrt(int inst, input logic enable, input int a, output int result);
import "DPI-C" context function void dpi_ftoi(int inst, input logic enable, input int a, output int result);
import "DPI-C" context function void dpi_ftou(int inst, input logic enable, input int a, output int result);
import "DPI-C" context function void dpi_itof(int inst, input logic enable, input int a, output int result);
import "DPI-C" context function void dpi_utof(int inst, input logic enable, input int a, output int result);
`endif