Vortex 2.0 changes:

+ 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

View File

@@ -1,71 +1,7 @@
XLEN ?= 32
LLVM_PREFIX ?= /opt/llvm-riscv
RISCV_TOOLCHAIN_PATH ?= /opt/riscv-gnu-toolchain
SYSROOT ?= $(RISCV_TOOLCHAIN_PATH)/riscv32-unknown-elf
POCL_CC_PATH ?= /opt/pocl/compiler
POCL_RT_PATH ?= /opt/pocl/runtime
VORTEX_DRV_PATH ?= $(realpath ../../../driver)
VORTEX_RT_PATH ?= $(realpath ../../../runtime)
OPTS ?= -n1024
K_LLCFLAGS += "-O3 -march=riscv32 -target-abi=ilp32f -mcpu=generic-rv32 -mattr=+m,+f -mattr=+vortex -float-abi=hard -code-model=small"
K_CFLAGS += "-v -O3 --sysroot=$(SYSROOT) --gcc-toolchain=$(RISCV_TOOLCHAIN_PATH) -march=rv32imf -mabi=ilp32f -Xclang -target-feature -Xclang +vortex -I$(VORTEX_RT_PATH)/include -fno-rtti -fno-exceptions -ffreestanding -nostartfiles -fdata-sections -ffunction-sections"
K_LDFLAGS += "-Wl,-Bstatic,-T$(VORTEX_RT_PATH)/linker/vx_link$(XLEN).ld -Wl,--gc-sections $(VORTEX_RT_PATH)/libvortexrt.a -lm"
CXXFLAGS += -std=c++11 -Wall -Wextra -Wfatal-errors
CXXFLAGS += -Wno-deprecated-declarations -Wno-unused-parameter -Wno-narrowing
CXXFLAGS += -I$(POCL_RT_PATH)/include
LDFLAGS += -L$(POCL_RT_PATH)/lib -L$(VORTEX_DRV_PATH)/stub -lOpenCL -lvortex
# Debugigng
ifdef DEBUG
CXXFLAGS += -g -O0
else
CXXFLAGS += -O2 -DNDEBUG
endif
PROJECT = saxpy
SRCS = main.cc
all: $(PROJECT) kernel.pocl
OPTS ?= -n1024
kernel.pocl: kernel.cl
LLVM_PREFIX=$(LLVM_PREFIX) POCL_DEBUG=all LD_LIBRARY_PATH=$(LLVM_PREFIX)/lib:$(POCL_CC_PATH)/lib $(POCL_CC_PATH)/bin/poclcc -LLCFLAGS $(K_LLCFLAGS) -CFLAGS $(K_CFLAGS) -LDFLAGS $(K_LDFLAGS) -o kernel.pocl kernel.cl
$(PROJECT): $(SRCS)
$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) -o $@
run-fpga: $(PROJECT) kernel.pocl
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/fpga:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-asesim: $(PROJECT) kernel.pocl
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/asesim:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-vlsim: $(PROJECT) kernel.pocl
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/vlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-simx: $(PROJECT) kernel.pocl
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/simx:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
run-rtlsim: $(PROJECT) kernel.pocl
LD_LIBRARY_PATH=$(POCL_RT_PATH)/lib:$(VORTEX_DRV_PATH)/rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) $(OPTS)
.depend: $(SRCS)
$(CXX) $(CXXFLAGS) -MM $^ > .depend;
clean:
rm -rf $(PROJECT) *.o .depend
clean-all: clean
rm -rf *.pocl *.dump
ifneq ($(MAKECMDGOALS),clean)
-include .depend
endif
include ../common.mk

Binary file not shown.

View File

@@ -31,6 +31,7 @@
#include <string.h>
#include <unistd.h>
#include <chrono>
#include <vector>
#define CL_CHECK(_expr) \
do { \
@@ -78,6 +79,14 @@ static int read_kernel_file(const char* filename, uint8_t** data, size_t* size)
return 0;
}
static bool almost_equal(float a, float b, int ulp = 4) {
union fi_t { int i; float f; };
fi_t fa, fb;
fa.f = a;
fb.f = b;
return std::abs(fa.i - fb.i) <= ulp;
}
uint8_t *kernel_bin = NULL;
///
@@ -142,7 +151,11 @@ int main(int argc, char **argv) {
cl_platform_id platform_id;
cl_device_id device_id;
cl_mem input_buffer;
cl_mem output_buffer;
size_t kernel_size;
cl_context context;
cl_command_queue queue;
cl_int binary_status = 0;
// read kernel binary from file
@@ -153,10 +166,7 @@ int main(int argc, char **argv) {
CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
CL_CHECK(clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_DEFAULT, 1, &device_id, NULL));
cl_context context;
context = CL_CHECK_ERR(clCreateContext(NULL, 1, &device_id, &pfn_notify, NULL, &_err));
cl_command_queue queue;
context = CL_CHECK_ERR(clCreateContext(NULL, 1, &device_id, &pfn_notify, NULL, &_err));
queue = CL_CHECK_ERR(clCreateCommandQueue(context, device_id, 0, &_err));
cl_kernel kernel = 0;
@@ -165,7 +175,7 @@ int main(int argc, char **argv) {
// Create OpenCL program - first attempt to load cached binary.
// If that is not available, then create the program from source
// and store the binary for future use.
std::cout << "Attempting to create program from binary..." << std::endl;
printf("create program from binary...\n");
cl_program program = CL_CHECK_ERR(clCreateProgramWithBinary(
context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err));
if (program == NULL) {
@@ -173,7 +183,7 @@ int main(int argc, char **argv) {
Cleanup(device_id, context, queue, program, kernel, memObjects);
return 1;
} else {
std::cout << "Read program from binary." << std::endl;
printf("Read program from binary.\n");
}
// Build program
@@ -181,22 +191,18 @@ int main(int argc, char **argv) {
size_t nbytes = sizeof(float) * size;
printf("attempting to create input buffer\n");
cl_mem input_buffer;
input_buffer = CL_CHECK_ERR(clCreateBuffer(
context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
printf("create input buffer\n");
input_buffer = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
printf("attempting to create output buffer\n");
cl_mem output_buffer;
output_buffer = CL_CHECK_ERR(clCreateBuffer(
context, CL_MEM_WRITE_ONLY, nbytes, NULL, &_err));
printf("create output buffer\n");
output_buffer = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, nbytes, NULL, &_err));
memObjects[0] = input_buffer;
memObjects[1] = output_buffer;
float factor = ((float)rand() / (float)(RAND_MAX)) * 100.0;
printf("attempting to create kernel\n");
printf("create kernel\n");
kernel = CL_CHECK_ERR(clCreateKernel(program, "saxpy", &_err));
printf("setting up kernel args\n");
@@ -204,36 +210,65 @@ int main(int argc, char **argv) {
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(output_buffer), &output_buffer));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(factor), &factor));
printf("attempting to enqueue write buffer\n");
float* h_src = (float*)malloc(nbytes);
for (int i = 0; i < size; i++) {
h_src[i] = ((float)rand() / (float)(RAND_MAX)) * 100.0;
}
CL_CHECK(clEnqueueWriteBuffer(queue, input_buffer, CL_TRUE, 0, nbytes, h_src, 0, NULL, NULL));
free(h_src);
size_t global_offset[1] = {0};
size_t global_work_size[1] = {size};
size_t local_work_size[1] = {1};
size_t global_work_size[] = {size/2, size/2};
printf("attempting to enqueue kernel\n");
printf("initialize buffers\n");
std::vector<float> ref_vec(size, 0.0f);
{
std::vector<float> dst_vec(size, 0.0f);
std::vector<float> src_vec(size);
for (int i = 0; i < size; i++) {
src_vec[i] = ((float)rand() / (float)(RAND_MAX)) * 100.0;
}
CL_CHECK(clEnqueueWriteBuffer(queue, input_buffer, CL_TRUE, 0, nbytes, src_vec.data(), 0, NULL, NULL));
CL_CHECK(clEnqueueWriteBuffer(queue, output_buffer, CL_TRUE, 0, nbytes, dst_vec.data(), 0, NULL, NULL));
size_t num_groups_x = global_work_size[0] / local_work_size[0];
for (size_t workgroup_id_x = 0; workgroup_id_x < num_groups_x; ++workgroup_id_x) {
for (size_t local_id_x = 0; local_id_x < local_work_size[0]; ++local_id_x) {
// Calculate global ID for the work-item
int global_id_x = global_offset[0] + local_work_size[0] * workgroup_id_x + local_id_x;
// kernel operation
int i = global_id_x;
ref_vec[i] += src_vec[i] * factor;
}
}
}
printf("enqueue kernel\n");
auto time_start = std::chrono::high_resolution_clock::now();
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size,
NULL, 0, NULL, NULL));
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, global_offset, global_work_size, local_work_size, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
auto time_end = std::chrono::high_resolution_clock::now();
double elapsed = std::chrono::duration_cast<std::chrono::milliseconds>(time_end - time_start).count();
printf("Elapsed time: %lg ms\n", elapsed);
printf("Download destination buffer\n");
float* h_dst = (float*)malloc(nbytes);
CL_CHECK(clEnqueueReadBuffer(queue, output_buffer, CL_TRUE, 0, nbytes, h_dst, 0, NULL, NULL));
printf("Verify result\n");
int errors = 0;
{
std::vector<float> dst_vec(size);
CL_CHECK(clEnqueueReadBuffer(queue, output_buffer, CL_TRUE, 0, nbytes, dst_vec.data(), 0, NULL, NULL));
/*printf("Result:");
for (int i = 0; i < size; i++) {
float data = h_dst[i];
printf(" %f", data);
}*/
free(h_dst);
for (int i = 0; i < size; ++i) {
if (!almost_equal(dst_vec[i], ref_vec[i])) {
if (errors < 100)
printf("*** error: [%d] expected=%f, actual=%f\n", i, ref_vec[i], dst_vec[i]);
++errors;
}
}
if (0 == errors) {
printf("PASSED!\n");
} else {
printf("FAILED! - %d errors\n", errors);
}
}
Cleanup(device_id, context, queue, program, kernel, memObjects);
return 0;
return errors;
}

File diff suppressed because it is too large Load Diff