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

Vortex 2.0 changes:

Browse Source 
+ Microarchitecture optimizations
+ 64-bit support
+ Xilinx FPGA support
+ LLVM-16 support
+ Refactoring and quality control fixes
This commit is contained in:
Blaise Tine
2023-10-19 20:51:22 -07:00
parent d69a64c32c
commit d47cccc157
1300 changed files with 247321 additions and 311189 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
tests/regression/fence/Makefile
View File

@@ -1,78 +1,9 @@
XLEN ?= 32
RISCV_TOOLCHAIN_PATH ?= /opt/riscv-gnu-toolchain
VORTEX_DRV_PATH ?= $(realpath ../../../driver)
VORTEX_RT_PATH ?= $(realpath ../../../runtime)
OPTS ?= -n64
VX_CC = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-gcc
VX_CXX = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-g++
VX_DP = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-objdump
VX_CP = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-objcopy
VX_CFLAGS += -march=rv32imf -mabi=ilp32f -O3 -Wstack-usage=1024 -ffreestanding -nostartfiles -fdata-sections -ffunction-sections
VX_CFLAGS += -I$(VORTEX_RT_PATH)/include -I$(VORTEX_RT_PATH)/../hw
VX_LDFLAGS += -Wl,-Bstatic,-T,$(VORTEX_RT_PATH)/linker/vx_link$(XLEN).ld -Wl,--gc-sections $(VORTEX_RT_PATH)/libvortexrt.a
VX_SRCS = kernel.c
CXXFLAGS += -std=c++11 -Wall -Wextra -pedantic -Wfatal-errors
CXXFLAGS += -I$(VORTEX_DRV_PATH)/include
LDFLAGS += -L$(VORTEX_DRV_PATH)/stub -lvortex
# Debugigng
ifdef DEBUG
CXXFLAGS += -g -O0
else
CXXFLAGS += -O2 -DNDEBUG
endif
PROJECT = fence
SRCS = main.cpp
all: $(PROJECT) kernel.bin kernel.dump
kernel.dump: kernel.elf
$(VX_DP) -D kernel.elf > kernel.dump
VX_SRCS = kernel.cpp
kernel.bin: kernel.elf
$(VX_CP) -O binary kernel.elf kernel.bin
OPTS ?= -n64
kernel.elf: $(VX_SRCS)
$(VX_CC) $(VX_CFLAGS) $(VX_SRCS) $(VX_LDFLAGS) -o kernel.elf
$(PROJECT): $(SRCS)
$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) -o $@
run-simx: $(PROJECT) kernel.bin
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/simx:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-fpga: $(PROJECT) kernel.bin
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/fpga:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-asesim: $(PROJECT) kernel.bin
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/asesim:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-vlsim: $(PROJECT) kernel.bin
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/vlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-rtlsim: $(PROJECT) kernel.bin
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
.depend: $(SRCS)
$(CXX) $(CXXFLAGS) -MM $^ > .depend;
clean:
rm -rf $(PROJECT) *.o .depend
clean-all: clean
rm -rf *.elf *.bin *.dump
ifneq ($(MAKECMDGOALS),clean)
-include .depend
endif
include ../common.mk

6
tests/regression/fence/common.h
View File

@@ -6,9 +6,9 @@
typedef struct {
uint32_t num_tasks;
uint32_t task_size;
uint32_t src0_addr;
uint32_t src1_addr;
uint32_t dst_addr;
uint64_t src0_addr;
uint64_t src1_addr;
uint64_t dst_addr;
} kernel_arg_t;
#endif

8
tests/regression/fence/kernel.c → tests/regression/fence/kernel.cpp
View File

@@ -3,14 +3,13 @@
#include <vx_spawn.h>
#include "common.h"
void kernel_body(int task_id, kernel_arg_t* arg) {
void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
uint32_t count = arg->task_size;
int32_t* src0_ptr = (int32_t*)arg->src0_addr;
int32_t* src1_ptr = (int32_t*)arg->src1_addr;
int32_t* dst_ptr = (int32_t*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
dst_ptr[offset+i] = src0_ptr[offset+i] + src1_ptr[offset+i];
}
@@ -18,7 +17,8 @@ void kernel_body(int task_id, kernel_arg_t* arg) {
vx_fence();
}
void main() {
int main() {
kernel_arg_t* arg = (kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
vx_spawn_tasks(arg->num_tasks, (vx_spawn_tasks_cb)kernel_body, arg);
}
return 0;
}

80
tests/regression/fence/main.cpp
View File

@@ -1,6 +1,7 @@
#include <iostream>
#include <unistd.h>
#include <string.h>
#include <vector>
#include <vortex.h>
#include "common.h"
@@ -20,8 +21,8 @@ const char* kernel_file = "kernel.bin";
uint32_t count = 0;
vx_device_h device = nullptr;
vx_buffer_h staging_buf = nullptr;
kernel_arg_t kernel_arg;
std::vector<uint8_t> staging_buf;
kernel_arg_t kernel_arg = {};
static void show_usage() {
std::cout << "Vortex Test." << std::endl;
@@ -51,9 +52,6 @@ static void parse_args(int argc, char **argv) {
}
void cleanup() {
if (staging_buf) {
vx_buf_free(staging_buf);
}
if (device) {
vx_mem_free(device, kernel_arg.src0_addr);
vx_mem_free(device, kernel_arg.src1_addr);
@@ -71,17 +69,17 @@ int run_test(const kernel_arg_t& kernel_arg,
// wait for completion
std::cout << "wait for completion" << std::endl;
RT_CHECK(vx_ready_wait(device, MAX_TIMEOUT));
RT_CHECK(vx_ready_wait(device, VX_MAX_TIMEOUT));
// download destination buffer
std::cout << "download destination buffer" << std::endl;
RT_CHECK(vx_copy_from_dev(staging_buf, kernel_arg.dst_addr, buf_size, 0));
RT_CHECK(vx_copy_from_dev(device, staging_buf.data(), kernel_arg.dst_addr, buf_size));
// verify result
std::cout << "verify result" << std::endl;
{
int errors = 0;
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
auto buf_ptr = (int32_t*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
int ref = i + i;
int cur = buf_ptr[i];
@@ -101,9 +99,7 @@ int run_test(const kernel_arg_t& kernel_arg,
return 0;
}
int main(int argc, char *argv[]) {
size_t value;
int main(int argc, char *argv[]) {
// parse command arguments
parse_args(argc, argv);
@@ -115,12 +111,12 @@ int main(int argc, char *argv[]) {
std::cout << "open device connection" << std::endl;
RT_CHECK(vx_dev_open(&device));
uint64_t max_cores, max_warps, max_threads;
RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_CORES, &max_cores));
RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_WARPS, &max_warps));
RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_THREADS, &max_threads));
uint64_t num_cores, num_warps, num_threads;
RT_CHECK(vx_dev_caps(device, VX_CAPS_NUM_CORES, &num_cores));
RT_CHECK(vx_dev_caps(device, VX_CAPS_NUM_WARPS, &num_warps));
RT_CHECK(vx_dev_caps(device, VX_CAPS_NUM_THREADS, &num_threads));
uint32_t num_tasks = max_cores * max_warps * max_threads;
uint32_t num_tasks = num_cores * num_warps * num_threads;
uint32_t num_points = count * num_tasks;
uint32_t buf_size = num_points * sizeof(int32_t);
@@ -132,64 +128,60 @@ int main(int argc, char *argv[]) {
RT_CHECK(vx_upload_kernel_file(device, kernel_file));
// allocate device memory
std::cout << "allocate device memory" << std::endl;
RT_CHECK(vx_mem_alloc(device, buf_size, &value));
kernel_arg.src0_addr = value;
RT_CHECK(vx_mem_alloc(device, buf_size, &value));
kernel_arg.src1_addr = value;
RT_CHECK(vx_mem_alloc(device, buf_size, &value));
kernel_arg.dst_addr = value;
std::cout << "allocate device memory" << std::endl;
RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src0_addr));
RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src1_addr));
RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.dst_addr));
kernel_arg.num_tasks = num_tasks;
kernel_arg.task_size = count;
std::cout << "dev_src0=" << std::hex << kernel_arg.src0_addr << std::endl;
std::cout << "dev_src1=" << std::hex << kernel_arg.src1_addr << std::endl;
std::cout << "dev_dst=" << std::hex << kernel_arg.dst_addr << std::endl;
std::cout << "dev_src0=0x" << std::hex << kernel_arg.src0_addr << std::endl;
std::cout << "dev_src1=0x" << std::hex << kernel_arg.src1_addr << std::endl;
std::cout << "dev_dst=0x" << std::hex << kernel_arg.dst_addr << std::endl;
// allocate shared memory
std::cout << "allocate shared memory" << std::endl;
// allocate staging buffer
std::cout << "allocate staging buffer" << std::endl;
uint32_t alloc_size = std::max<uint32_t>(buf_size, sizeof(kernel_arg_t));
RT_CHECK(vx_buf_alloc(device, alloc_size, &staging_buf));
staging_buf.resize(alloc_size);
// upload kernel argument
std::cout << "upload kernel argument" << std::endl;
{
auto buf_ptr = (int*)vx_host_ptr(staging_buf);
auto buf_ptr = (int*)staging_buf.data();
memcpy(buf_ptr, &kernel_arg, sizeof(kernel_arg_t));
RT_CHECK(vx_copy_to_dev(staging_buf, KERNEL_ARG_DEV_MEM_ADDR, sizeof(kernel_arg_t), 0));
RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, staging_buf.data(), sizeof(kernel_arg_t)));
}
// upload source buffer0
{
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
std::cout << "upload source buffer0" << std::endl;
auto buf_ptr = (int32_t*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
buf_ptr[i] = i-1;
}
}
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src0_addr, staging_buf.data(), buf_size));
}
std::cout << "upload source buffer0" << std::endl;
RT_CHECK(vx_copy_to_dev(staging_buf, kernel_arg.src0_addr, buf_size, 0));
// upload source buffer1
{
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
std::cout << "upload source buffer1" << std::endl;
auto buf_ptr = (int32_t*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
buf_ptr[i] = i+1;
}
}
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src1_addr, staging_buf.data(), buf_size));
}
std::cout << "upload source buffer1" << std::endl;
RT_CHECK(vx_copy_to_dev(staging_buf, kernel_arg.src1_addr, buf_size, 0));
// clear destination buffer
{
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
std::cout << "clear destination buffer" << std::endl;
auto buf_ptr = (int32_t*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
buf_ptr[i] = 0xdeadbeef;
}
}
RT_CHECK(vx_copy_to_dev(device, kernel_arg.dst_addr, staging_buf.data(), buf_size));
}
std::cout << "clear destination buffer" << std::endl;
RT_CHECK(vx_copy_to_dev(staging_buf, kernel_arg.dst_addr, buf_size, 0));
// run tests
std::cout << "run tests" << std::endl;