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
OPTS ?= -n32
VORTEX_DRV_PATH ?= $(realpath ../../../driver)
VORTEX_RT_PATH ?= $(realpath ../../../runtime)
K_LLCFLAGS += "-O3 -march=riscv32 -target-abi=ilp32f -mcpu=generic-rv32 -mattr=+m,+f -mattr=+vortex -float-abi=hard -code-model=small"
K_CFLAGS += "-v -O3 --sysroot=$(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 = sgemm
SRCS = main.cc
all: $(PROJECT) kernel.pocl
OPTS ?= -n32
kernel.pocl: kernel.cl
LLVM_PREFIX=$(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

View File

@@ -0,0 +1,12 @@
#ifndef COMMON_H
#define COMMON_H
#define USE_FLOAT
#ifdef USE_FLOAT
#define TYPE float
#else
#define TYPE int
#endif
#endif // COMMON_H

View File

@@ -1,6 +1,8 @@
__kernel void sgemm (__global const float *A,
__global const float *B,
__global float *C,
#include "common.h"
__kernel void sgemm (__global const TYPE *A,
__global const TYPE *B,
__global TYPE *C,
int N)
{
// Thread identifiers
@@ -8,7 +10,7 @@ __kernel void sgemm (__global const float *A,
const int c = get_global_id(1); // Col ID
// Compute a single element (loop a K)
float acc = 0.0f;
TYPE acc = 0;
for (int k = 0; k < N; k++) {
acc += A[k * N + r] * B[c * N + k];
}

Binary file not shown.

View File

@@ -6,6 +6,8 @@
#include <time.h>
#include <unistd.h>
#include <chrono>
#include <vector>
#include "common.h"
#define KERNEL_NAME "sgemm"
@@ -52,25 +54,31 @@ static int read_kernel_file(const char* filename, uint8_t** data, size_t* size)
return 0;
}
static void matmul(float *C, const float* A, const float *B, int M, int N, int K) {
/*static void matmul(TYPE *C, const TYPE* A, const TYPE *B, int M, int N, int K) {
for (int m = 0; m < M; ++m) {
for (int n = 0; n < N; ++n) {
float acc = 0.0f;
TYPE acc = 0;
for (int k = 0; k < K; ++k) {
acc += A[k * M + m] * B[n * K + k];
}
C[n * M + m] = acc;
}
}
}
}*/
static bool almost_equal(float a, float b, int ulp = 21) {
#ifdef USE_FLOAT
static bool compare_equal(float a, float b, int ulp = 21) {
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;
}
#else
static bool compare_equal(int a, int b, int ulp = 21) {
return (a == b);
}
#endif
cl_device_id device_id = NULL;
cl_context context = NULL;
@@ -80,9 +88,9 @@ cl_kernel kernel = NULL;
cl_mem a_memobj = NULL;
cl_mem b_memobj = NULL;
cl_mem c_memobj = NULL;
float *h_a = NULL;
float *h_b = NULL;
float *h_c = NULL;
TYPE *h_a = NULL;
TYPE *h_b = NULL;
TYPE *h_c = NULL;
uint8_t *kernel_bin = NULL;
static void cleanup() {
@@ -155,7 +163,7 @@ int main (int argc, char **argv) {
context = CL_CHECK2(clCreateContext(NULL, 1, &device_id, NULL, NULL, &_err));
// Allocate device buffers
size_t nbytes = size * size * sizeof(float);
size_t nbytes = size * size * sizeof(TYPE);
a_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
b_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
c_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_WRITE_ONLY, nbytes, NULL, &_err));
@@ -182,16 +190,54 @@ int main (int argc, char **argv) {
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(width), (void*)&width));
// Allocate memories for input arrays and output arrays.
h_a = (float*)malloc(nbytes);
h_b = (float*)malloc(nbytes);
h_c = (float*)malloc(nbytes);
h_a = (TYPE*)malloc(nbytes);
h_b = (TYPE*)malloc(nbytes);
h_c = (TYPE*)malloc(nbytes);
// Initialize values for array members.
for (int i = 0; i < (size * size); ++i) {
#ifdef USE_FLOAT
h_a[i] = (float)rand() / (float)RAND_MAX;
h_b[i] = (float)rand() / (float)RAND_MAX;
#else
h_a[i] = rand();
h_b[i] = rand();
#endif
h_c[i] = 0xdeadbeef;
//printf("*** [%d]: h_a=%f, h_b=%f\n", i, h_a[i], h_b[i]);
}
size_t global_offset[2] = {0, 0};
size_t global_work_size[2] = {size, size};
size_t local_work_size[2] = {1, 1};
std::vector<float> ref_vec(size * size);
// reference generation
size_t num_groups_y = global_work_size[1] / local_work_size[1];
size_t num_groups_x = global_work_size[0] / local_work_size[0];
for (size_t workgroup_id_y = 0; workgroup_id_y < num_groups_y; ++workgroup_id_y) {
for (size_t workgroup_id_x = 0; workgroup_id_x < num_groups_x; ++workgroup_id_x) {
for (size_t local_id_y = 0; local_id_y < local_work_size[1]; ++local_id_y) {
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;
int global_id_y = global_offset[1] + local_work_size[1] * workgroup_id_y + local_id_y;
// kernel operation
int r = global_id_x;
int c = global_id_y;
TYPE acc = 0;
for (int k = 0; k < width; k++) {
acc += h_a[k * width + r] * h_b[c * width + k];
}
/*#ifdef USE_FLOAT
printf("*** r=%d, c=%d, v=%f\n", r, c, acc);
#else
printf("*** r=%d, c=%d, v=%d\n", r, c, acc);
#endif*/
ref_vec[c * width + r] = acc;
}
}
}
}
// Creating command queue
@@ -201,11 +247,9 @@ int main (int argc, char **argv) {
CL_CHECK(clEnqueueWriteBuffer(commandQueue, a_memobj, CL_TRUE, 0, nbytes, h_a, 0, NULL, NULL));
CL_CHECK(clEnqueueWriteBuffer(commandQueue, b_memobj, CL_TRUE, 0, nbytes, h_b, 0, NULL, NULL));
printf("Execute the kernel\n");
size_t global_work_size[2] = {size, size};
size_t local_work_size[2] = {1, 1};
printf("Execute the kernel\n");
auto time_start = std::chrono::high_resolution_clock::now();
CL_CHECK(clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL, global_work_size, local_work_size, 0, NULL, NULL));
CL_CHECK(clEnqueueNDRangeKernel(commandQueue, kernel, 2, global_offset, global_work_size, local_work_size, 0, NULL, NULL));
CL_CHECK(clFinish(commandQueue));
auto time_end = std::chrono::high_resolution_clock::now();
double elapsed = std::chrono::duration_cast<std::chrono::milliseconds>(time_end - time_start).count();
@@ -216,16 +260,17 @@ int main (int argc, char **argv) {
printf("Verify result\n");
int errors = 0;
float* h_ref = (float*)malloc(nbytes);
matmul(h_ref, h_a, h_b, size, size, size);
for (int i = 0; i < (size * size); i++) {
if (!almost_equal(h_c[i], h_ref[i])) {
if (!compare_equal(h_c[i], ref_vec[i])) {
if (errors < 100)
printf("*** error: [%d] expected=%f, actual=%f\n", i, h_ref[i], h_c[i]);
#ifdef USE_FLOAT
printf("*** error: [%d] expected=%f, actual=%f\n", i, ref_vec[i], h_c[i]);
#else
printf("*** error: [%d] expected=%d, actual=%d\n", i, ref_vec[i], h_c[i]);
#endif
++errors;
}
}
free(h_ref);
}
if (errors != 0) {
printf("FAILED! - %d errors\n", errors);
} else {

File diff suppressed because it is too large Load Diff