yosys/passes/memory/memory_share.cc

553 lines
17 KiB
C++

/*
* yosys -- Yosys Open SYnthesis Suite
*
* Copyright (C) 2012 Claire Xenia Wolf <claire@yosyshq.com>
*
* Permission to use, copy, modify, and/or distribute this software for any
* purpose with or without fee is hereby granted, provided that the above
* copyright notice and this permission notice appear in all copies.
*
* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
*
*/
#include "kernel/yosys.h"
#include "kernel/qcsat.h"
#include "kernel/sigtools.h"
#include "kernel/modtools.h"
#include "kernel/mem.h"
#include "kernel/ffinit.h"
USING_YOSYS_NAMESPACE
PRIVATE_NAMESPACE_BEGIN
struct MemoryShareWorker
{
RTLIL::Design *design;
RTLIL::Module *module;
SigMap sigmap, sigmap_xmux;
ModWalker modwalker;
FfInitVals initvals;
bool flag_widen;
bool flag_sat;
// --------------------------------------------------
// Consolidate read ports that read the same address
// (or close enough to be merged to wide ports)
// --------------------------------------------------
// A simple function to detect ports that couldn't possibly collide
// because of opposite const address bits (simplistic, but enough
// to fix problems with inferring wide ports).
bool rdwr_can_collide(Mem &mem, int ridx, int widx) {
auto &rport = mem.rd_ports[ridx];
auto &wport = mem.wr_ports[widx];
for (int i = std::max(rport.wide_log2, wport.wide_log2); i < GetSize(rport.addr) && i < GetSize(wport.addr); i++) {
if (rport.addr[i] == State::S1 && wport.addr[i] == State::S0)
return false;
if (rport.addr[i] == State::S0 && wport.addr[i] == State::S1)
return false;
}
return true;
}
bool merge_rst_value(Mem &mem, Const &res, int wide_log2, const Const &src1, int sub1, const Const &src2, int sub2) {
res = Const(State::Sx, mem.width << wide_log2);
for (int i = 0; i < GetSize(src1); i++)
res[i + sub1 * mem.width] = src1[i];
for (int i = 0; i < GetSize(src2); i++) {
if (src2[i] == State::Sx)
continue;
auto &dst = res[i + sub2 * mem.width];
if (dst == src2[i])
continue;
if (dst != State::Sx)
return false;
dst = src2[i];
}
return true;
}
bool consolidate_rd_by_addr(Mem &mem)
{
if (GetSize(mem.rd_ports) <= 1)
return false;
log("Consolidating read ports of memory %s.%s by address:\n", log_id(module), log_id(mem.memid));
bool changed = false;
for (int i = 0; i < GetSize(mem.rd_ports); i++)
{
auto &port1 = mem.rd_ports[i];
if (port1.removed)
continue;
for (int j = i + 1; j < GetSize(mem.rd_ports); j++)
{
auto &port2 = mem.rd_ports[j];
if (port2.removed)
continue;
if (port1.clk_enable != port2.clk_enable)
continue;
if (port1.clk_enable) {
if (port1.clk != port2.clk)
continue;
if (port1.clk_polarity != port2.clk_polarity)
continue;
}
if (port1.en != port2.en)
continue;
if (port1.arst != port2.arst)
continue;
if (port1.srst != port2.srst)
continue;
if (port1.ce_over_srst != port2.ce_over_srst)
continue;
// If the width of the ports doesn't match, they can still be
// merged by widening the narrow one. Check if the conditions
// hold for that.
int wide_log2 = std::max(port1.wide_log2, port2.wide_log2);
SigSpec addr1 = sigmap_xmux(port1.addr);
SigSpec addr2 = sigmap_xmux(port2.addr);
if (GetSize(addr1) <= wide_log2)
continue;
if (GetSize(addr2) <= wide_log2)
continue;
if (!addr1.extract(0, wide_log2).is_fully_const())
continue;
if (!addr2.extract(0, wide_log2).is_fully_const())
continue;
if (addr1.extract_end(wide_log2) != addr2.extract_end(wide_log2)) {
// Incompatible addresses after widening. Last chance — widen both
// ports by one more bit to merge them.
if (!flag_widen)
continue;
wide_log2++;
if (addr1.extract_end(wide_log2) != addr2.extract_end(wide_log2))
continue;
if (!addr1.extract(0, wide_log2).is_fully_const())
continue;
if (!addr2.extract(0, wide_log2).is_fully_const())
continue;
}
// Combine init/reset values.
SigSpec sub1_c = port1.addr.extract(0, wide_log2);
log_assert(sub1_c.is_fully_const());
int sub1 = sub1_c.as_int();
SigSpec sub2_c = port2.addr.extract(0, wide_log2);
log_assert(sub2_c.is_fully_const());
int sub2 = sub2_c.as_int();
Const init_value, arst_value, srst_value;
if (!merge_rst_value(mem, init_value, wide_log2, port1.init_value, sub1, port2.init_value, sub2))
continue;
if (!merge_rst_value(mem, arst_value, wide_log2, port1.arst_value, sub1, port2.arst_value, sub2))
continue;
if (!merge_rst_value(mem, srst_value, wide_log2, port1.srst_value, sub1, port2.srst_value, sub2))
continue;
// At this point we are committed to the merge.
{
log(" Merging ports %d, %d (address %s).\n", i, j, log_signal(port1.addr));
port1.addr = addr1;
port2.addr = addr2;
mem.prepare_rd_merge(i, j, &initvals);
mem.widen_prep(wide_log2);
SigSpec new_data = module->addWire(NEW_ID, mem.width << wide_log2);
module->connect(port1.data, new_data.extract(sub1 * mem.width, mem.width << port1.wide_log2));
module->connect(port2.data, new_data.extract(sub2 * mem.width, mem.width << port2.wide_log2));
for (int k = 0; k < wide_log2; k++)
port1.addr[k] = State::S0;
port1.init_value = init_value;
port1.arst_value = arst_value;
port1.srst_value = srst_value;
port1.wide_log2 = wide_log2;
port1.data = new_data;
port2.removed = true;
changed = true;
}
}
}
if (changed)
mem.emit();
return changed;
}
// ------------------------------------------------------
// Consolidate write ports that write to the same address
// (or close enough to be merged to wide ports)
// ------------------------------------------------------
bool consolidate_wr_by_addr(Mem &mem)
{
if (GetSize(mem.wr_ports) <= 1)
return false;
log("Consolidating write ports of memory %s.%s by address:\n", log_id(module), log_id(mem.memid));
bool changed = false;
for (int i = 0; i < GetSize(mem.wr_ports); i++)
{
auto &port1 = mem.wr_ports[i];
if (port1.removed)
continue;
if (!port1.clk_enable)
continue;
for (int j = i + 1; j < GetSize(mem.wr_ports); j++)
{
auto &port2 = mem.wr_ports[j];
if (port2.removed)
continue;
if (!port2.clk_enable)
continue;
if (port1.clk != port2.clk)
continue;
if (port1.clk_polarity != port2.clk_polarity)
continue;
// If the width of the ports doesn't match, they can still be
// merged by widening the narrow one. Check if the conditions
// hold for that.
int wide_log2 = std::max(port1.wide_log2, port2.wide_log2);
SigSpec addr1 = sigmap_xmux(port1.addr);
SigSpec addr2 = sigmap_xmux(port2.addr);
if (GetSize(addr1) <= wide_log2)
continue;
if (GetSize(addr2) <= wide_log2)
continue;
if (!addr1.extract(0, wide_log2).is_fully_const())
continue;
if (!addr2.extract(0, wide_log2).is_fully_const())
continue;
if (addr1.extract_end(wide_log2) != addr2.extract_end(wide_log2)) {
// Incompatible addresses after widening. Last chance — widen both
// ports by one more bit to merge them.
if (!flag_widen)
continue;
wide_log2++;
if (addr1.extract_end(wide_log2) != addr2.extract_end(wide_log2))
continue;
if (!addr1.extract(0, wide_log2).is_fully_const())
continue;
if (!addr2.extract(0, wide_log2).is_fully_const())
continue;
}
log(" Merging ports %d, %d (address %s).\n", i, j, log_signal(addr1));
port1.addr = addr1;
port2.addr = addr2;
mem.prepare_wr_merge(i, j, &initvals);
mem.widen_wr_port(i, wide_log2);
mem.widen_wr_port(j, wide_log2);
int pos = 0;
while (pos < GetSize(port1.data)) {
int epos = pos;
while (epos < GetSize(port1.data) && port1.en[epos] == port1.en[pos] && port2.en[epos] == port2.en[pos])
epos++;
int width = epos - pos;
SigBit new_en;
if (port2.en[pos] == State::S0) {
new_en = port1.en[pos];
} else if (port1.en[pos] == State::S0) {
port1.data.replace(pos, port2.data.extract(pos, width));
new_en = port2.en[pos];
} else {
port1.data.replace(pos, module->Mux(NEW_ID, port1.data.extract(pos, width), port2.data.extract(pos, width), port2.en[pos]));
new_en = module->Or(NEW_ID, port1.en[pos], port2.en[pos]);
}
for (int k = pos; k < epos; k++)
port1.en[k] = new_en;
pos = epos;
}
changed = true;
port2.removed = true;
}
}
if (changed)
mem.emit();
return changed;
}
// --------------------------------------------------------
// Consolidate write ports using sat-based resource sharing
// --------------------------------------------------------
void consolidate_wr_using_sat(Mem &mem)
{
if (GetSize(mem.wr_ports) <= 1)
return;
// Get a list of ports that have any chance of being mergeable.
pool<int> eligible_ports;
for (int i = 0; i < GetSize(mem.wr_ports); i++) {
auto &port = mem.wr_ports[i];
std::vector<RTLIL::SigBit> bits = modwalker.sigmap(port.en);
for (auto bit : bits)
if (bit == RTLIL::State::S1)
goto port_is_always_active;
eligible_ports.insert(i);
port_is_always_active:;
}
if (eligible_ports.size() <= 1)
return;
log("Consolidating write ports of memory %s.%s using sat-based resource sharing:\n", log_id(module), log_id(mem.memid));
// Group eligible ports by clock domain and width.
pool<int> checked_ports;
std::vector<std::vector<int>> groups;
for (int i = 0; i < GetSize(mem.wr_ports); i++)
{
auto &port1 = mem.wr_ports[i];
if (!eligible_ports.count(i))
continue;
if (checked_ports.count(i))
continue;
std::vector<int> group;
group.push_back(i);
for (int j = i + 1; j < GetSize(mem.wr_ports); j++)
{
auto &port2 = mem.wr_ports[j];
if (!eligible_ports.count(j))
continue;
if (checked_ports.count(j))
continue;
if (port1.clk_enable != port2.clk_enable)
continue;
if (port1.clk_enable) {
if (port1.clk != port2.clk)
continue;
if (port1.clk_polarity != port2.clk_polarity)
continue;
}
if (port1.wide_log2 != port2.wide_log2)
continue;
group.push_back(j);
}
for (auto j : group)
checked_ports.insert(j);
if (group.size() <= 1)
continue;
groups.push_back(group);
}
bool changed = false;
for (auto &group : groups) {
auto &some_port = mem.wr_ports[group[0]];
string ports;
for (auto idx : group) {
if (idx != group[0])
ports += ", ";
ports += std::to_string(idx);
}
if (!some_port.clk_enable) {
log(" Checking unclocked group, width %d: ports %s.\n", mem.width << some_port.wide_log2, ports.c_str());
} else {
log(" Checking group clocked with %sedge %s, width %d: ports %s.\n", some_port.clk_polarity ? "pos" : "neg", log_signal(some_port.clk), mem.width << some_port.wide_log2, ports.c_str());
}
// Okay, time to actually run the SAT solver.
QuickConeSat qcsat(modwalker);
// create SAT representation of common input cone of all considered EN signals
dict<int, int> port_to_sat_variable;
for (auto idx : group)
port_to_sat_variable[idx] = qcsat.ez->expression(qcsat.ez->OpOr, qcsat.importSig(mem.wr_ports[idx].en));
qcsat.prepare();
log(" Common input cone for all EN signals: %d cells.\n", GetSize(qcsat.imported_cells));
log(" Size of unconstrained SAT problem: %d variables, %d clauses\n", qcsat.ez->numCnfVariables(), qcsat.ez->numCnfClauses());
// now try merging the ports.
for (int ii = 0; ii < GetSize(group); ii++) {
int idx1 = group[ii];
auto &port1 = mem.wr_ports[idx1];
if (port1.removed)
continue;
for (int jj = ii + 1; jj < GetSize(group); jj++) {
int idx2 = group[jj];
auto &port2 = mem.wr_ports[idx2];
if (port2.removed)
continue;
if (qcsat.ez->solve(port_to_sat_variable.at(idx1), port_to_sat_variable.at(idx2))) {
log(" According to SAT solver sharing of port %d with port %d is not possible.\n", idx1, idx2);
continue;
}
log(" Merging port %d into port %d.\n", idx2, idx1);
mem.prepare_wr_merge(idx1, idx2, &initvals);
port_to_sat_variable.at(idx1) = qcsat.ez->OR(port_to_sat_variable.at(idx1), port_to_sat_variable.at(idx2));
RTLIL::SigSpec last_addr = port1.addr;
RTLIL::SigSpec last_data = port1.data;
std::vector<RTLIL::SigBit> last_en = modwalker.sigmap(port1.en);
RTLIL::SigSpec this_addr = port2.addr;
RTLIL::SigSpec this_data = port2.data;
std::vector<RTLIL::SigBit> this_en = modwalker.sigmap(port2.en);
RTLIL::SigBit this_en_active = module->ReduceOr(NEW_ID, this_en);
if (GetSize(last_addr) < GetSize(this_addr))
last_addr.extend_u0(GetSize(this_addr));
else
this_addr.extend_u0(GetSize(last_addr));
SigSpec new_addr = module->Mux(NEW_ID, last_addr.extract_end(port1.wide_log2), this_addr.extract_end(port1.wide_log2), this_en_active);
port1.addr = SigSpec({new_addr, port1.addr.extract(0, port1.wide_log2)});
port1.data = module->Mux(NEW_ID, last_data, this_data, this_en_active);
std::map<std::pair<RTLIL::SigBit, RTLIL::SigBit>, int> groups_en;
RTLIL::SigSpec grouped_last_en, grouped_this_en, en;
RTLIL::Wire *grouped_en = module->addWire(NEW_ID, 0);
for (int j = 0; j < int(this_en.size()); j++) {
std::pair<RTLIL::SigBit, RTLIL::SigBit> key(last_en[j], this_en[j]);
if (!groups_en.count(key)) {
grouped_last_en.append(last_en[j]);
grouped_this_en.append(this_en[j]);
groups_en[key] = grouped_en->width;
grouped_en->width++;
}
en.append(RTLIL::SigSpec(grouped_en, groups_en[key]));
}
module->addMux(NEW_ID, grouped_last_en, grouped_this_en, this_en_active, grouped_en);
port1.en = en;
port2.removed = true;
changed = true;
}
}
}
if (changed)
mem.emit();
}
// -------------
// Setup and run
// -------------
MemoryShareWorker(RTLIL::Design *design, bool flag_widen, bool flag_sat) : design(design), modwalker(design), flag_widen(flag_widen), flag_sat(flag_sat) {}
void operator()(RTLIL::Module* module)
{
std::vector<Mem> memories = Mem::get_selected_memories(module);
this->module = module;
sigmap.set(module);
initvals.set(&sigmap, module);
sigmap_xmux = sigmap;
for (auto cell : module->cells())
{
if (cell->type == ID($mux))
{
RTLIL::SigSpec sig_a = sigmap_xmux(cell->getPort(ID::A));
RTLIL::SigSpec sig_b = sigmap_xmux(cell->getPort(ID::B));
if (sig_a.is_fully_undef())
sigmap_xmux.add(cell->getPort(ID::Y), sig_b);
else if (sig_b.is_fully_undef())
sigmap_xmux.add(cell->getPort(ID::Y), sig_a);
}
}
for (auto &mem : memories) {
while (consolidate_rd_by_addr(mem));
while (consolidate_wr_by_addr(mem));
}
if (!flag_sat)
return;
modwalker.setup(module);
for (auto &mem : memories)
consolidate_wr_using_sat(mem);
}
};
struct MemorySharePass : public Pass {
MemorySharePass() : Pass("memory_share", "consolidate memory ports") { }
void help() override
{
// |---v---|---v---|---v---|---v---|---v---|---v---|---v---|---v---|---v---|---v---|
log("\n");
log(" memory_share [-nosat] [-nowiden] [selection]\n");
log("\n");
log("This pass merges share-able memory ports into single memory ports.\n");
log("\n");
log("The following methods are used to consolidate the number of memory ports:\n");
log("\n");
log(" - When multiple write ports access the same address then this is converted\n");
log(" to a single write port with a more complex data and/or enable logic path.\n");
log("\n");
log(" - When multiple read or write ports access adjacent aligned addresses, they are\n");
log(" merged to a single wide read or write port. This transformation can be\n");
log(" disabled with the \"-nowiden\" option.\n");
log("\n");
log(" - When multiple write ports are never accessed at the same time (a SAT\n");
log(" solver is used to determine this), then the ports are merged into a single\n");
log(" write port. This transformation can be disabled with the \"-nosat\" option.\n");
log("\n");
log("Note that in addition to the algorithms implemented in this pass, the $memrd\n");
log("and $memwr cells are also subject to generic resource sharing passes (and other\n");
log("optimizations) such as \"share\" and \"opt_merge\".\n");
log("\n");
}
void execute(std::vector<std::string> args, RTLIL::Design *design) override {
bool flag_widen = true;
bool flag_sat = true;
log_header(design, "Executing MEMORY_SHARE pass (consolidating $memrd/$memwr cells).\n");
size_t argidx;
for (argidx = 1; argidx < args.size(); argidx++)
{
if (args[argidx] == "-nosat")
{
flag_sat = false;
continue;
}
if (args[argidx] == "-nowiden")
{
flag_widen = false;
continue;
}
break;
}
extra_args(args, 1, design);
MemoryShareWorker msw(design, flag_widen, flag_sat);
for (auto module : design->selected_modules())
msw(module);
}
} MemorySharePass;
PRIVATE_NAMESPACE_END