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

L2 and L1 using different block size support, RTLsim fixes, dram_rsp_ready optimization

Browse Source
This commit is contained in:
Blaise Tine
2020-11-21 09:47:56 -08:00
parent a7da36c007
commit 1795980a52
50 changed files with 972 additions and 952 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/opae/Makefile
View File

@@ -1,7 +1,7 @@
OPAE_HOME ?= /tools/opae/1.4.0
CXXFLAGS += -std=c++11 -O2 -DNDEBUG -Wall -Wextra -pedantic -Wfatal-errors
#CXXFLAGS += -std=c++11 -O0 -g -Wall -Wextra -pedantic -Wfatal-errors
#CXXFLAGS += -std=c++11 -O2 -DNDEBUG -Wall -Wextra -pedantic -Wfatal-errors
CXXFLAGS += -std=c++11 -O0 -g -Wall -Wextra -pedantic -Wfatal-errors
CXXFLAGS += -I../include -I$(OPAE_HOME)/include -I../../hw
@@ -60,7 +60,7 @@ json: ../../hw/opae/vortex_afu.json
afu_json_mgr json-info --afu-json=$^ --c-hdr=$@
fpga: $(SRCS)
$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) $(FPGA_LIBS) -o $(PROJECT)
$(CXX) $(CXXFLAGS) -DUSE_FPGA $^ $(LDFLAGS) $(FPGA_LIBS) -o $(PROJECT)
asesim: $(SRCS) $(ASE_DIR)
$(CXX) $(CXXFLAGS) -DUSE_ASE $(SRCS) $(LDFLAGS) $(ASE_LIBS) -o $(PROJECT_ASE)

8
driver/opae/vlsim/Makefile
View File

@@ -1,5 +1,5 @@
CFLAGS += -std=c++11 -O2 -DNDEBUG -Wall -Wextra -Wfatal-errors
#CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -Wfatal-errors
#CFLAGS += -std=c++11 -O2 -DNDEBUG -Wall -Wextra -Wfatal-errors
CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -Wfatal-errors
CFLAGS += -I../../../../hw
@@ -30,7 +30,7 @@ CONFIGS ?= -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0
CFLAGS += -fPIC
CFLAGS += -DUSE_RTLSIM $(CONFIGS)
CFLAGS += -DUSE_VLSIM $(CONFIGS)
CFLAGS += -DDUMP_PERF_STATS
@@ -79,7 +79,7 @@ VL_FLAGS += -DNOPAE
CFLAGS += -DNOPAE
# use DPI FPU
#VL_FLAGS += -DFPU_FAST
VL_FLAGS += -DFPU_FAST
RTL_INCLUDE += -I../../../hw/opae -I../../../hw/opae/ccip

21
driver/opae/vlsim/opae_sim.cpp
View File

@@ -206,11 +206,10 @@ void opae_sim::sRxPort_bus() {
vortex_afu_->vcp2af_sRxPort_c0_rspValid = 1;
memcpy(vortex_afu_->vcp2af_sRxPort_c0_data, cci_rd_it->block.data(), CACHE_BLOCK_SIZE);
vortex_afu_->vcp2af_sRxPort_c0_hdr_mdata = cci_rd_it->mdata;
/*printf("*** [vlsim] read-rsp: addr=%ld, mdata=%d, data=", cci_rd_it->addr, cci_rd_it->mdata);
/*printf("%0ld: [sim] CCI Rd Rsp: addr=%ld, mdata=%d, data=", timestamp, cci_rd_it->addr, cci_rd_it->mdata);
for (int i = 0; i < CACHE_BLOCK_SIZE; ++i)
printf("%02x", cci_rd_it->block[CACHE_BLOCK_SIZE-1-i]);
printf("\n");*/
fflush(stdout);
cci_reads_.erase(cci_rd_it);
}
}
@@ -225,8 +224,7 @@ void opae_sim::sTxPort_bus() {
cci_req.mdata = vortex_afu_->af2cp_sTxPort_c0_hdr_mdata;
auto host_ptr = (uint64_t*)(vortex_afu_->af2cp_sTxPort_c0_hdr_address * CACHE_BLOCK_SIZE);
memcpy(cci_req.block.data(), host_ptr, CACHE_BLOCK_SIZE);
//printf("*** [vlsim] read-req: addr=%ld, mdata=%d\n", vortex_afu_->af2cp_sTxPort_c0_hdr_address, cci_req.mdata);
fflush(stdout);
//printf("%0ld: [sim] CCI Rd Req: addr=%ld, mdata=%d\n", timestamp, vortex_afu_->af2cp_sTxPort_c0_hdr_address, cci_req.mdata);
cci_reads_.emplace_back(cci_req);
}
@@ -265,12 +263,12 @@ void opae_sim::avs_bus() {
memcpy(vortex_afu_->avs_readdata, dram_rd_it->block.data(), CACHE_BLOCK_SIZE);
uint32_t tag = dram_rd_it->tag;
dram_reads_.erase(dram_rd_it);
/*printf("%0ld: VLSIM: DRAM rsp: addr=%x, pending={", timestamp, tag);
/*printf("%0ld: [sim] DRAM Rd Rsp: addr=%x, pending={", timestamp, tag);
for (auto& req : dram_reads_) {
if (req.cycles_left != 0)
printf(" !%0x", req.tag);
printf(" !%0x", req.tag);
else
printf(" %0x", req.tag);
printf(" %0x", req.tag);
}
printf("}\n");*/
}
@@ -288,7 +286,8 @@ void opae_sim::avs_bus() {
// process DRAM requests
if (!dram_stalled) {
if (vortex_afu_->avs_write) {
assert(!vortex_afu_->avs_read || !vortex_afu_->avs_write);
if (vortex_afu_->avs_write) {
assert(0 == vortex_afu_->mem_bank_select);
uint64_t byteen = vortex_afu_->avs_byteenable;
unsigned base_addr = (vortex_afu_->avs_address * CACHE_BLOCK_SIZE);
@@ -307,12 +306,12 @@ void opae_sim::avs_bus() {
ram_.read(base_addr, CACHE_BLOCK_SIZE, dram_req.block.data());
dram_req.tag = base_addr;
dram_reads_.emplace_back(dram_req);
/*printf("%0ld: VLSIM: DRAM req: addr=%x, pending={", timestamp, base_addr);
/*printf("%0ld: [sim] DRAM Rd Req: addr=%x, pending={", timestamp, base_addr);
for (auto& req : dram_reads_) {
if (req.cycles_left != 0)
printf(" !%0x", req.tag);
printf(" !%0x", req.tag);
else
printf(" %0x", req.tag);
printf(" %0x", req.tag);
}
printf("}\n");*/
}

6
driver/opae/vortex.cpp
View File

@@ -7,11 +7,11 @@
#include <assert.h>
#include <cmath>
#ifdef USE_VLSIM
#include "vlsim/fpga.h"
#else
#if defined(USE_FPGA) || defined(USE_ASE)
#include <opae/fpga.h>
#include <uuid/uuid.h>
#elif defined(USE_VLSIM)
#include "vlsim/fpga.h"
#endif
#include <vortex.h>

6
driver/rtlsim/Makefile
View File

@@ -1,5 +1,5 @@
CFLAGS += -std=c++11 -O2 -DNDEBUG -Wall -Wextra -Wfatal-errors
#CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -Wfatal-errors
#CFLAGS += -std=c++11 -O2 -DNDEBUG -Wall -Wextra -Wfatal-errors
CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -Wfatal-errors
CFLAGS += -I../../include -I../../../hw/simulate -I../../../hw
@@ -65,7 +65,7 @@ else
endif
# use DPI FPU
#VL_FLAGS += -DFPU_FAST
VL_FLAGS += -DFPU_FAST
PROJECT = libvortex.so
# PROJECT = libvortex.dylib

2
driver/tests/demo/demo.cpp
View File

@@ -92,7 +92,7 @@ int run_test(const kernel_arg_t& kernel_arg,
}
}
if (errors != 0) {
std::cout << "Found " << errors << " errors!" << std::endl;
std::cout << "Found " << std::dec << errors << " errors!" << std::endl;
std::cout << "FAILED!" << std::endl;
return 1;
}

2
driver/tests/dogfood/dogfood.cpp
View File

@@ -260,7 +260,7 @@ int main(int argc, char *argv[]) {
(void*)vx_host_ptr(src1_buf),
(void*)vx_host_ptr(src2_buf));
if (errors != 0) {
std::cout << "found " << errors << " errors!" << std::endl;
std::cout << "found " << std::dec << errors << " errors!" << std::endl;
std::cout << "Test" << t << "-" << name << " FAILED!" << std::endl << std::flush;
if (stop_on_error) {
cleanup();

21
driver/tests/dogfood/testcases.h
View File

@@ -14,29 +14,36 @@ union Float_t {
} parts;
};
inline float fround(float x, int32_t precision = 4) {
inline float fround(float x, int32_t precision = 8) {
auto power_of_10 = std::pow(10, precision);
return std::round(x * power_of_10) / power_of_10;
}
inline bool almost_equal_eps(float a, float b, float eps = std::numeric_limits<float>::epsilon()) {
auto tolerance = std::min(fabs(a), fabs(b)) * eps;
return fabs(a - b) <= tolerance;
inline bool almost_equal_eps(float a, float b, int ulp = 128) {
auto eps = std::numeric_limits<float>::epsilon() * (std::max(fabs(a), fabs(b)) * ulp);
auto d = fabs(a - b);
if (d > eps) {
std::cout << "*** almost_equal_eps: d=" << d << ", eps=" << eps << std::endl;
return false;
}
return true;
}
inline bool almost_equal_ulp(float a, float b, int32_t ulp = 4) {
inline bool almost_equal_ulp(float a, float b, int32_t ulp = 6) {
Float_t fa{a}, fb{b};
auto d = std::abs(fa.i - fb.i);
if (d > ulp) {
std::cout << "*** float compare: a=" << a << ", b=" << b << ", ulp=" << d << ", ia=" << std::hex << fa.i << ", ib=" << fb.i << std::endl;
std::cout << "*** almost_equal_ulp: a=" << a << ", b=" << b << ", ulp=" << d << ", ia=" << std::hex << fa.i << ", ib=" << fb.i << std::endl;
return false;
}
return true;
}
inline bool almost_equal(float a, float b) {
if (almost_equal_eps(a, b))
if (a == b)
return true;
/*if (almost_equal_eps(a, b))
return true;*/
return almost_equal_ulp(a, b);
}