speeding up simulation using dedicated full dpi-based FPU core

hw/dpi/float_dpi.cpp

@@ -0,0 +1,264 @@
+#include <stdio.h>
+#include <math.h>
+#include <unordered_map>
+#include <vector>
+#include <mutex>
+#include <iostream>
+#include "svdpi.h"
+#include "verilated_vpi.h"
+#include "VX_config.h"
+
+extern "C" {
+  void dpi_fadd(int a, int b, int frm, int* result, int* fflags);
+  void dpi_fsub(int a, int b, int frm, int* result, int* fflags);
+  void dpi_fmul(int a, int b, int frm, int* result, int* fflags);
+  void dpi_fmadd(int a, int b, int c, int frm, int* result, int* fflags);
+  void dpi_fmsub(int a, int b, int c, int frm, int* result, int* fflags);
+  void dpi_fnmadd(int a, int b, int c, int frm, int* result, int* fflags);
+  void dpi_fnmsub(int a, int b, int c, int frm, int* result, int* fflags);
+
+  void dpi_fdiv(int a, int b, int frm, int* result, int* fflags);
+  void dpi_fsqrt(int a, int frm, int* result, int* fflags);
+  
+  void dpi_ftoi(int a, int frm, int* result, int* fflags);
+  void dpi_ftou(int a, int frm, int* result, int* fflags);
+  void dpi_itof(int a, int frm, int* result, int* fflags);
+  void dpi_utof(int a, int frm, int* result, int* fflags);
+
+  void dpi_fclss(int a, int* result);
+  void dpi_fsgnj(int a, int* result);
+  void dpi_fsgnjn(int a, int* result);
+  void dpi_fsgnjx(int a, int* result);
+
+  void dpi_flt(int a, int b, int* result, int* fflags);
+  void dpi_fle(int a, int b, int* result, int* fflags);
+  void dpi_feq(int a, int b, int* result, int* fflags);
+  void dpi_fmin(int a, int b, int* result, int* fflags);
+  void dpi_fmax(int a, int b, int* result, int* fflags);
+}
+
+union Float_t {    
+    float f;
+    int   i;
+    struct {
+        uint32_t man  : 23;
+        uint32_t exp  : 8;
+        uint32_t sign : 1;
+    } parts;
+};
+
+void dpi_fadd(int a, int b, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f + fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fsub(int a, int b, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f - fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fmul(int a, int b, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f * fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fmadd(int a, int b, int c, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fc, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fc.i = c;
+  fr.f = fa.f * fb.f + fc.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fmsub(int a, int b, int c, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fc, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fc.i = c;
+  fr.f = fa.f * fb.f - fc.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fnmadd(int a, int b, int c, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fc, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fc.i = c;
+  fr.f = -(fa.f * fb.f + fc.f);
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fnmsub(int a, int b, int c, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fc, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fc.i = c;
+  fr.f = -(fa.f * fb.f - fc.f);
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fdiv(int a, int b, int frm, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f / fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fsqrt(int a, int frm, int* result, int* fflags) {
+  Float_t fa, fr;
+
+  fa.i = a;
+  fr.f = sqrtf(fa.f);
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_ftoi(int a, int frm, int* result, int* fflags) {
+  Float_t fa, fr;
+
+  fa.i = a;
+  fr.i = int(fa.f);   
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_ftou(int a, int frm, int* result, int* fflags) {
+  Float_t fa, fr;
+
+  fa.i = a;
+  fr.i = unsigned(fa.f);   
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_itof(int a, int frm, int* result, int* fflags) {
+  Float_t fa, fr;
+
+  fr.f = (float)a;   
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_utof(int a, int frm, int* result, int* fflags) {
+  Float_t fa, fr;
+
+  unsigned ua = a;
+  fr.f = (float)ua;   
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_flt(int a, int b, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f < fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fle(int a, int b, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f <= fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_feq(int a, int b, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = fa.f == fb.f;
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fmin(int a, int b, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = std::min<float>(fa.f, fb.f);
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fmax(int a, int b, int* result, int* fflags) {
+  Float_t fa, fb, fr;
+
+  fa.i = a;
+  fb.i = b;
+  fr.f = std::max<float>(fa.f, fb.f);
+
+  *result = fr.i;
+  *fflags = 0;
+}
+
+void dpi_fclss(int a, int* result) {
+  // TODO
+  *result = 0;
+}
+
+void dpi_fsgnj(int a, int* result) {
+  // TODO
+  *result = 0;
+}
+
+void dpi_fsgnjn(int a, int* result) {
+  // TODO
+  *result = 0;
+}
+
+void dpi_fsgnjx(int a, int* result) {
+  // TODO
+  *result = 0;
+}

hw/dpi/float_dpi.vh

@@ -0,0 +1,31 @@
+`ifndef FLOAT_DPI
+`define FLOAT_DPI
+
+import "DPI-C" context function void dpi_fadd(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fsub(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fmul(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fmadd(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fmsub(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fnmadd(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fnmsub(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+
+import "DPI-C" context function void dpi_fdiv(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fsqrt(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+
+import "DPI-C" context function void dpi_ftoi(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_ftou(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_itof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_utof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
+
+import "DPI-C" context function void dpi_fclss(input int a, output int result);
+import "DPI-C" context function void dpi_fsgnj(input int a, output int result);
+import "DPI-C" context function void dpi_fsgnjn(input int a, output int result);
+import "DPI-C" context function void dpi_fsgnjx(input int a, output int result);
+
+import "DPI-C" context function void dpi_flt(input int a, input int b, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fle(input int a, input int b, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_feq(input int a, input int b, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fmin(input int a, input int b, output int result, output bit[4:0] fflags);
+import "DPI-C" context function void dpi_fmax(input int a, input int b, output int result, output bit[4:0] fflags);
+
+`endif

hw/dpi/util_dpi.cpp

@@ -0,0 +1,84 @@
+#include <stdio.h>
+#include <math.h>
+#include <unordered_map>
+#include <vector>
+#include <mutex>
+#include <iostream>
+#include "svdpi.h"
+#include "verilated_vpi.h"
+#include "VX_config.h"
+
+extern "C" {
+  int dpi_register();
+  void dpi_assert(int inst, bool cond, int delay);
+}
+
+class ShiftRegister {
+public:
+  ShiftRegister() : init_(false), depth_(0) {}
+
+  void ensure_init(int depth) {
+    if (!init_) {
+      buffer_.resize(depth);
+      init_  = true;
+      depth_ = depth;
+    }
+  }
+
+  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;
+  }
+
+  int top() const {
+    return buffer_[0];
+  }
+
+private:
+
+  std::vector<int> buffer_;
+  bool init_;
+  unsigned depth_;  
+};
+
+class Instances {
+public:
+  ShiftRegister& get(int inst) {
+    return instances_.at(inst);
+  }
+
+  int allocate() {
+    mutex_.lock();   
+    int inst = instances_.size();
+    instances_.resize(inst + 1); 
+    mutex_.unlock();
+    return inst;
+  }
+
+private:
+  std::vector<ShiftRegister> instances_;
+  std::mutex mutex_;
+};
+
+Instances instances;
+
+int dpi_register() {
+  return instances.allocate();
+}
+
+void dpi_assert(int inst, bool cond, int delay) {
+  ShiftRegister& sr = instances.get(inst);
+
+  sr.ensure_init(delay);
+  sr.push(!cond, 1);
+
+  auto status = sr.top();
+  if (status) {
+    printf("delayed assertion at %s!\n", svGetNameFromScope(svGetScope()));
+    std::abort();
+  }
+}

hw/dpi/util_dpi.vh

@@ -0,0 +1,7 @@
+`ifndef UTIL_DPI
+`define UTIL_DPI
+
+import "DPI-C" context function int dpi_register();
+import "DPI-C" context function void dpi_assert(int inst, input logic cond, input int delay);
+
+`endif