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

77
tests/regression/dogfood/Makefile
View File

@@ -1,80 +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 $(VORTEX_RT_PATH)/libvortexrt.a
VX_LDFLAGS += -lm
VX_SRCS = kernel.c
CXXFLAGS += -std=c++11 -Wall -Wextra -pedantic -Wfatal-errors
CXXFLAGS += -I$(VORTEX_DRV_PATH)/include -I$(VORTEX_RT_PATH)/../hw
LDFLAGS += -L$(VORTEX_DRV_PATH)/stub -lvortex
# Debugigng
ifdef DEBUG
CXXFLAGS += -g -O0
else
CXXFLAGS += -O2 -DNDEBUG
endif
PROJECT = dogfood
SRCS = main.cpp
all: $(PROJECT) kernel.bin kernel.dump
kernel.dump: kernel.elf
$(VX_DP) -d -r -t kernel.elf > kernel.dump
VX_SRCS = kernel.cpp
kernel.bin: kernel.elf
$(VX_CP) -O binary kernel.elf kernel.bin
OPTS ?= -n64 -x19 -x20
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/dogfood/common.h
View File

@@ -7,9 +7,9 @@ typedef struct {
uint32_t testid;
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

334
tests/regression/dogfood/kernel.c
View File

@@ -1,334 +0,0 @@
#include <stdint.h>
#include <math.h>
#include <vx_intrinsics.h>
#include <vx_spawn.h>
#include "common.h"
typedef void (*PFN_Kernel)(int task_id, kernel_arg_t* arg);
inline float __ieee754_sqrtf (float x) {
asm ("fsqrt.s %0, %1" : "=f" (x) : "f" (x));
return x;
}
void kernel_iadd(int task_id, kernel_arg_t* 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) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
int32_t c = a + b;
dst_ptr[offset+i] = c;
}
}
void kernel_imul(int task_id, kernel_arg_t* 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) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
int32_t c = a * b;
dst_ptr[offset+i] = c;
}
}
void kernel_idiv(int task_id, kernel_arg_t* 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) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
int32_t c = a / b;
dst_ptr[offset+i] = c;
}
}
void kernel_idiv_mul(int task_id, kernel_arg_t* 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) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
int32_t c = a / b;
int32_t d = a * b;
int32_t e = c + d;
dst_ptr[offset+i] = e;
}
}
void kernel_fadd(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a + b;
dst_ptr[offset+i] = c;
}
}
void kernel_fsub(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a - b;
dst_ptr[offset+i] = c;
}
}
void kernel_fmul(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a * b;
dst_ptr[offset+i] = c;
}
}
void kernel_fmadd(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a * b + b;
dst_ptr[offset+i] = c;
}
}
void kernel_fmsub(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a * b - b;
dst_ptr[offset+i] = c;
}
}
void kernel_fnmadd(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c =-a * b - b;
dst_ptr[offset+i] = c;
}
}
void kernel_fnmsub(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c =-a * b + b;
dst_ptr[offset+i] = c;
}
}
void kernel_fnmadd_madd(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c =-a * b - b;
float d = a * b + b;
float e = c + d;
dst_ptr[offset+i] = e;
}
}
void kernel_fdiv(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a / b;
dst_ptr[offset+i] = c;
}
}
void kernel_fdiv2(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a / b;
float d = b / a;
float e = c + d;
dst_ptr[offset+i] = e;
}
}
void kernel_fsqrt(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = __ieee754_sqrtf(a * b);
dst_ptr[offset+i] = c;
}
}
void kernel_ftoi(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)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) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a + b;
int32_t d = (int32_t)c;
dst_ptr[offset+i] = d;
}
}
void kernel_ftou(int task_id, kernel_arg_t* arg) {
uint32_t count = arg->task_size;
float* src0_ptr = (float*)arg->src0_addr;
float* src1_ptr = (float*)arg->src1_addr;
uint32_t* dst_ptr = (uint32_t*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a + b;
uint32_t d = (uint32_t)c;
dst_ptr[offset+i] = d;
}
}
void kernel_itof(int task_id, kernel_arg_t* 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;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
int32_t c = a + b;
float d = (float)c;
dst_ptr[offset+i] = d;
}
}
void kernel_utof(int task_id, kernel_arg_t* 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;
float* dst_ptr = (float*)arg->dst_addr;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
uint32_t c = a + b;
float d = (float)c;
dst_ptr[offset+i] = d;
}
}
static const PFN_Kernel sc_tests[] = {
kernel_iadd,
kernel_imul,
kernel_idiv,
kernel_idiv_mul,
kernel_fadd,
kernel_fsub,
kernel_fmul,
kernel_fmadd,
kernel_fmsub,
kernel_fnmadd,
kernel_fnmsub,
kernel_fnmadd_madd,
kernel_fdiv,
kernel_fdiv2,
kernel_fsqrt,
kernel_ftoi,
kernel_ftou,
kernel_itof,
kernel_utof,
};
void main() {
kernel_arg_t* arg = (kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
vx_spawn_tasks(arg->num_tasks, (vx_spawn_tasks_cb)sc_tests[arg->testid], arg);
}

396
tests/regression/dogfood/kernel.cpp Normal file
View File

@@ -0,0 +1,396 @@
#include <stdint.h>
#include <math.h>
#include <vx_intrinsics.h>
#include <vx_spawn.h>
#include "common.h"
typedef void (*PFN_Kernel)(int task_id, kernel_arg_t* __UNIFORM__ arg);
inline float __ieee754_sqrtf (float x) {
asm ("fsqrt.s %0, %1" : "=f" (x) : "f" (x));
return x;
}
void kernel_iadd(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (int32_t*)arg->src0_addr;
auto src1_ptr = (int32_t*)arg->src1_addr;
auto dst_ptr = (int32_t*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
int32_t a = src0_ptr[offset+i];
int32_t b = src1_ptr[offset+i];
int32_t c = a + b;
dst_ptr[offset+i] = c;
}
}
void kernel_imul(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (int32_t*)arg->src0_addr;
auto src1_ptr = (int32_t*)arg->src1_addr;
auto dst_ptr = (int32_t*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a * b;
dst_ptr[offset+i] = c;
}
}
void kernel_idiv(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (int32_t*)arg->src0_addr;
auto src1_ptr = (int32_t*)arg->src1_addr;
auto dst_ptr = (int32_t*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a / b;
dst_ptr[offset+i] = c;
}
}
void kernel_idiv_mul(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (int32_t*)arg->src0_addr;
auto src1_ptr = (int32_t*)arg->src1_addr;
auto dst_ptr = (int32_t*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a / b;
auto d = a * b;
auto e = c + d;
dst_ptr[offset+i] = e;
}
}
void kernel_fadd(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
float a = src0_ptr[offset+i];
float b = src1_ptr[offset+i];
float c = a + b;
dst_ptr[offset+i] = c;
}
}
void kernel_fsub(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a - b;
dst_ptr[offset+i] = c;
}
}
void kernel_fmul(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a * b;
dst_ptr[offset+i] = c;
}
}
void kernel_fmadd(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a * b + b;
dst_ptr[offset+i] = c;
}
}
void kernel_fmsub(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a * b - b;
dst_ptr[offset+i] = c;
}
}
void kernel_fnmadd(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c =-a * b - b;
dst_ptr[offset+i] = c;
}
}
void kernel_fnmsub(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c =-a * b + b;
dst_ptr[offset+i] = c;
}
}
void kernel_fnmadd_madd(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c =-a * b - b;
auto d = a * b + b;
auto e = c + d;
dst_ptr[offset+i] = e;
}
}
void kernel_fdiv(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a / b;
dst_ptr[offset+i] = c;
}
}
void kernel_fdiv2(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a / b;
auto d = b / a;
auto e = c + d;
dst_ptr[offset+i] = e;
}
}
void kernel_fsqrt(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = __ieee754_sqrtf(a * b);
dst_ptr[offset+i] = c;
}
}
void kernel_ftoi(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (int32_t*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a + b;
auto d = (int32_t)c;
dst_ptr[offset+i] = d;
}
}
void kernel_ftou(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (float*)arg->src0_addr;
auto src1_ptr = (float*)arg->src1_addr;
auto dst_ptr = (uint32_t*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a + b;
auto d = (uint32_t)c;
dst_ptr[offset+i] = d;
}
}
void kernel_itof(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (int32_t*)arg->src0_addr;
auto src1_ptr = (int32_t*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a + b;
auto d = (float)c;
dst_ptr[offset+i] = d;
}
}
void kernel_utof(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto count = arg->task_size;
auto src0_ptr = (int32_t*)arg->src0_addr;
auto src1_ptr = (int32_t*)arg->src1_addr;
auto dst_ptr = (float*)arg->dst_addr;
auto offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {
auto a = src0_ptr[offset+i];
auto b = src1_ptr[offset+i];
auto c = a + b;
auto d = (float)c;
dst_ptr[offset+i] = d;
}
}
void kernel_bar(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto num_warps = vx_num_warps();
auto num_threads = vx_num_threads();
auto cid = vx_core_id();
auto wid = vx_warp_id();
auto tid = vx_thread_id();
auto src0_ptr = (uint32_t*)arg->src0_addr;
auto dst_ptr = (uint32_t*)arg->dst_addr;
// per warp delay
uint32_t barrier_stall = 0;
for (int i = 0; i <= wid; ++i) {
barrier_stall += src0_ptr[0] * src0_ptr[i];
}
// memory fence
vx_fence();
// local barrier
vx_barrier(0, num_warps);
// update destination
auto src_idx = (cid * num_warps + (num_warps - 1 - wid)) * num_threads + tid;
dst_ptr[task_id] = src0_ptr[src_idx] + barrier_stall;
}
void kernel_gbar(int task_id, kernel_arg_t* __UNIFORM__ arg) {
auto num_cores = vx_num_cores();
auto num_warps = vx_num_warps();
auto num_threads = vx_num_threads();
auto cid = vx_core_id();
auto wid = vx_warp_id();
auto tid = vx_thread_id();
auto src0_ptr = (uint32_t*)arg->src0_addr;
auto dst_ptr = (uint32_t*)arg->dst_addr;
// per core delay
uint32_t barrier_stall = 0;
for (int i = 0; i <= cid; ++i) {
for (int j = 0; j <= wid; ++j) {
barrier_stall += src0_ptr[0] * src0_ptr[i + j];
}
}
// memory fence
vx_fence();
// global barrier
vx_barrier(0x80000000, num_cores);
// update destination
auto src_idx = ((num_cores - 1 - cid) * num_warps + (num_warps - 1 - wid)) * num_threads + tid;
dst_ptr[task_id] = src0_ptr[src_idx] + barrier_stall;
}
static const PFN_Kernel sc_tests[] = {
kernel_iadd,
kernel_imul,
kernel_idiv,
kernel_idiv_mul,
kernel_fadd,
kernel_fsub,
kernel_fmul,
kernel_fmadd,
kernel_fmsub,
kernel_fnmadd,
kernel_fnmsub,
kernel_fnmadd_madd,
kernel_fdiv,
kernel_fdiv2,
kernel_fsqrt,
kernel_ftoi,
kernel_ftou,
kernel_itof,
kernel_utof,
kernel_bar,
kernel_gbar
};
int main() {
auto arg = (kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
vx_spawn_tasks(arg->num_tasks, (vx_spawn_tasks_cb)sc_tests[arg->testid], arg);
return 0;
}

197
tests/regression/dogfood/main.cpp
View File

@@ -1,109 +1,49 @@
#include <iostream>
#include <vector>
#include <unordered_set>
#include <unistd.h>
#include <string.h>
#include <vector>
#include <vortex.h>
#include <VX_config.h>
#include "testcases.h"
#include "common.h"
#define RT_CHECK(_expr) \
do { \
int _ret = _expr; \
if (0 == _ret) \
break; \
printf("Error: '%s' returned %d!\n", #_expr, (int)_ret); \
cleanup(); \
exit(-1); \
} while (false)
///////////////////////////////////////////////////////////////////////////////
class TestMngr {
public:
TestMngr() {
this->add_test("iadd", new Test_IADD());
this->add_test("imul", new Test_IMUL());
this->add_test("idiv", new Test_IDIV());
this->add_test("idiv-mul", new Test_IDIV_MUL());
#ifdef EXT_F_ENABLE
this->add_test("fadd", new Test_FADD());
this->add_test("fsub", new Test_FSUB());
this->add_test("fmul", new Test_FMUL());
this->add_test("fmadd", new Test_FMADD());
this->add_test("fmsub", new Test_FMSUB());
this->add_test("fnmadd", new Test_FNMADD());
this->add_test("fnmsub", new Test_FNMSUB());
this->add_test("fnmadd-madd", new Test_FNMADD_MADD());
this->add_test("fdiv", new Test_FDIV());
this->add_test("fdiv2", new Test_FDIV2());
this->add_test("fsqrt", new Test_FSQRT());
this->add_test("ftoi", new Test_FTOI());
this->add_test("ftou", new Test_FTOU());
this->add_test("itof", new Test_ITOF());
this->add_test("utof", new Test_UTOF());
#endif
}
~TestMngr() {
for (size_t i = 0; i < _tests.size(); ++i) {
delete _tests[i];
}
}
const std::string& get_name(int testid) const {
return _names.at(testid);
}
ITestCase* get_test(int testid) const {
return _tests.at(testid);
}
void add_test(const char* name, ITestCase* test) {
_names.push_back(name);
_tests.push_back(test);
}
size_t size() const {
return _tests.size();
}
private:
std::vector<std::string> _names;
std::vector<ITestCase*> _tests;
};
///////////////////////////////////////////////////////////////////////////////
TestMngr testMngr;
TestSuite* testSuite = nullptr;
const char* kernel_file = "kernel.bin";
int count = 0;
int count = 0;
std::unordered_set<int> included;
std::unordered_set<int> excluded;
int testid_s = 0;
int testid_e = (testMngr.size() - 1);
int testid_e = 0;
bool stop_on_error = true;
vx_device_h device = nullptr;
vx_buffer_h arg_buf = nullptr;
vx_buffer_h src1_buf = nullptr;
vx_buffer_h src2_buf = nullptr;
vx_buffer_h dst_buf = nullptr;
kernel_arg_t kernel_arg;
vx_device_h device = nullptr;
std::vector<uint8_t> arg_buf;
std::vector<uint8_t> src1_buf;
std::vector<uint8_t> src2_buf;
std::vector<uint8_t> dst_buf;
kernel_arg_t kernel_arg = {};
static void show_usage() {
std::cout << "Vortex Test." << std::endl;
std::cout << "Usage: [-t:testid] [-s:testid] [-e:testid] [-k: kernel] [-n words] [-c] [-h: help]" << std::endl;
std::cout << "Usage: [-t<testid>: selected test] [-s<testid>: start test] [-e<testid>: end test] [-x<testid>: excluded tests]" << std::endl;
std::cout << " [-k<kernel>] [-n<words>] [-c] [-h: help]" << std::endl;
}
static void parse_args(int argc, char **argv) {
int c;
while ((c = getopt(argc, argv, "n:t:s:e:k:ch?")) != -1) {
while ((c = getopt(argc, argv, "n:t:x:s:e:k:ch?")) != -1) {
switch (c) {
case 'n':
count = atoi(optarg);
break;
case 't':
testid_s = atoi(optarg);
testid_e = atoi(optarg);
included.insert(atoi(optarg));
break;
case 'x':
excluded.insert(atoi(optarg));
break;
case 's':
testid_s = atoi(optarg);
@@ -130,17 +70,8 @@ static void parse_args(int argc, char **argv) {
}
void cleanup() {
if (arg_buf) {
vx_buf_free(arg_buf);
}
if (src1_buf) {
vx_buf_free(src1_buf);
}
if (src2_buf) {
vx_buf_free(src2_buf);
}
if (dst_buf) {
vx_buf_free(dst_buf);
if (testSuite) {
delete testSuite;
}
if (device) {
vx_mem_free(device, kernel_arg.src0_addr);
@@ -152,7 +83,6 @@ void cleanup() {
int main(int argc, char *argv[]) {
int exitcode = 0;
size_t value;
// parse command arguments
parse_args(argc, argv);
@@ -171,12 +101,12 @@ 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));
int num_tasks = max_cores * max_warps * max_threads;
int num_tasks = num_cores * num_warps * num_threads;
int num_points = count * num_tasks;
size_t buf_size = num_points * sizeof(uint32_t);
@@ -188,59 +118,69 @@ 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::dec << std::endl;
std::cout << "dev_src1=" << std::hex << kernel_arg.src1_addr << std::dec << std::endl;
std::cout << "dev_dst=" << std::hex << kernel_arg.dst_addr << std::dec << std::endl;
std::cout << "dev_src0=0x" << std::hex << kernel_arg.src0_addr << std::dec << std::endl;
std::cout << "dev_src1=0x" << std::hex << kernel_arg.src1_addr << std::dec << std::endl;
std::cout << "dev_dst=0x" << std::hex << kernel_arg.dst_addr << std::dec << std::endl;
// allocate shared memory
std::cout << "allocate shared memory" << std::endl;
RT_CHECK(vx_buf_alloc(device, sizeof(kernel_arg_t), &arg_buf));
RT_CHECK(vx_buf_alloc(device, buf_size, &src1_buf));
RT_CHECK(vx_buf_alloc(device, buf_size, &src2_buf));
RT_CHECK(vx_buf_alloc(device, buf_size, &dst_buf));
// allocate staging buffer
std::cout << "allocate staging buffer" << std::endl;
arg_buf.resize(sizeof(kernel_arg_t));
src1_buf.resize(buf_size);
src2_buf.resize(buf_size);
dst_buf.resize(buf_size);
// allocate test suite
testSuite = new TestSuite(device);
if (testid_e == 0) {
testid_e = (testSuite->size() - 1);
}
// execute tests
for (int t = testid_s; t <= testid_e; ++t) {
auto name = testMngr.get_name(t);
auto test = testMngr.get_test(t);
if (!included.empty()) {
if (included.count(t) == 0)
continue;
}
if (!excluded.empty()) {
if (excluded.count(t) != 0)
continue;
}
auto test = testSuite->get_test(t);
auto name = test->name();
std::cout << "Test" << t << ": " << name << std::endl;
// upload kernel argument
std::cout << "upload kernel argument" << std::endl;
kernel_arg.testid = t;
memcpy((void*)vx_host_ptr(arg_buf), &kernel_arg, sizeof(kernel_arg_t));
RT_CHECK(vx_copy_to_dev(arg_buf, KERNEL_ARG_DEV_MEM_ADDR, sizeof(kernel_arg_t), 0));
memcpy(arg_buf.data(), &kernel_arg, sizeof(kernel_arg_t));
RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, arg_buf.data(), sizeof(kernel_arg_t)));
// get test arguments
std::cout << "get test arguments" << std::endl;
test->setup(num_points, (void*)vx_host_ptr(src1_buf), (void*)vx_host_ptr(src2_buf));
RT_CHECK(test->setup(num_points, (void*)src1_buf.data(), (void*)src2_buf.data()));
// upload source buffer0
std::cout << "upload source buffer0" << std::endl;
RT_CHECK(vx_copy_to_dev(src1_buf, kernel_arg.src0_addr, buf_size, 0));
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src0_addr, src1_buf.data(), buf_size));
// upload source buffer1
std::cout << "upload source buffer1" << std::endl;
RT_CHECK(vx_copy_to_dev(src2_buf, kernel_arg.src1_addr, buf_size, 0));
std::cout << "upload source buffer1" << std::endl;
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src1_addr, src2_buf.data(), buf_size));
// clear destination buffer
std::cout << "clear destination buffer" << std::endl;
for (int i = 0; i < num_points; ++i) {
((uint32_t*)vx_host_ptr(dst_buf))[i] = 0xdeadbeef;
((uint32_t*)dst_buf.data())[i] = 0xdeadbeef;
}
RT_CHECK(vx_copy_to_dev(dst_buf, kernel_arg.dst_addr, buf_size, 0));
RT_CHECK(vx_copy_to_dev(device, kernel_arg.dst_addr, dst_buf.data(), buf_size));
// start device
std::cout << "start device" << std::endl;
@@ -248,18 +188,15 @@ int main(int argc, char *argv[]) {
// 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(dst_buf, kernel_arg.dst_addr, buf_size, 0));
RT_CHECK(vx_copy_from_dev(device, dst_buf.data(), kernel_arg.dst_addr, buf_size));
// verify destination
std::cout << "verify test result" << std::endl;
int errors = test->verify(num_points,
(void*)vx_host_ptr(dst_buf),
(void*)vx_host_ptr(src1_buf),
(void*)vx_host_ptr(src2_buf));
int errors = test->verify(num_points, dst_buf.data(), src1_buf.data(), src2_buf.data());
if (errors != 0) {
std::cout << "found " << std::dec << errors << " errors!" << std::endl;
std::cout << "Test" << t << "-" << name << " FAILED!" << std::endl << std::flush;

389
tests/regression/dogfood/testcases.h
View File

@@ -3,6 +3,19 @@
#include <iostream>
#include <math.h>
#include <limits>
#include <assert.h>
void cleanup();
#define RT_CHECK(_expr) \
do { \
int _ret = _expr; \
if (0 == _ret) \
break; \
printf("Error: '%s' returned %d!\n", #_expr, (int)_ret); \
cleanup(); \
exit(-1); \
} while (false)
union Float_t {
float f;
@@ -47,33 +60,72 @@ inline bool almost_equal(float a, float b) {
return almost_equal_ulp(a, b);
}
class ITestCase;
class TestSuite {
public:
TestSuite(vx_device_h device);
~TestSuite();
ITestCase* get_test(int testid) const;
void add_test(ITestCase* test);
size_t size() const;
vx_device_h device() const;
private:
std::vector<ITestCase*> _tests;
vx_device_h device_;
};
class ITestCase {
public:
ITestCase() {}
ITestCase(TestSuite* suite, const char* name)
: suite_(suite)
, name_(name)
{}
virtual ~ITestCase() {}
virtual void setup(int n, void* src1, void* src2) = 0;
virtual int verify(int n, void* dst, const void* src1, const void* src2) = 0;
TestSuite* suite() const {
return suite_;
}
const char* name() const {
return name_;
}
virtual int setup(uint32_t n, void* src1, void* src2) = 0;
virtual int verify(uint32_t n, void* dst, const void* src1, const void* src2) = 0;
protected:
TestSuite* suite_;
const char* const name_;
};
class Test_IADD : public ITestCase {
public:
Test_IADD(TestSuite* suite) : ITestCase(suite, "iadd") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = n/2 - i;
b[i] = n/2 + i;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
auto c = (int32_t*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] + b[i];
if (c[i] != ref) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -86,22 +138,24 @@ public:
class Test_IMUL : public ITestCase {
public:
Test_IMUL(TestSuite* suite) : ITestCase(suite, "imul") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = n/2 - i;
b[i] = n/2 + i;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
auto c = (int32_t*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] * b[i];
if (c[i] != ref) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -114,22 +168,24 @@ public:
class Test_IDIV : public ITestCase {
public:
Test_IDIV(TestSuite* suite) : ITestCase(suite, "idiv") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = n/2 - i;
b[i] = n/2 + i;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
auto c = (int32_t*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] / b[i];
if (c[i] != ref) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -142,22 +198,24 @@ public:
class Test_IDIV_MUL : public ITestCase {
public:
Test_IDIV_MUL(TestSuite* suite) : ITestCase(suite, "idiv-mul") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = n/2 - i;
b[i] = n/2 + i;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
auto c = (int32_t*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = a[i] / b[i];
auto y = a[i] * b[i];
auto ref = x + y;
@@ -172,22 +230,24 @@ public:
class Test_FADD : public ITestCase {
public:
Test_FADD(TestSuite* suite) : ITestCase(suite, "fadd") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] + b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -200,22 +260,24 @@ public:
class Test_FSUB : public ITestCase {
public:
Test_FSUB(TestSuite* suite) : ITestCase(suite, "fsub") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] - b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -228,22 +290,24 @@ public:
class Test_FMUL : public ITestCase {
public:
Test_FMUL(TestSuite* suite) : ITestCase(suite, "fmul") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] * b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -256,22 +320,24 @@ public:
class Test_FMADD : public ITestCase {
public:
Test_FMADD(TestSuite* suite) : ITestCase(suite, "fmadd") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] * b[i] + b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -284,22 +350,24 @@ public:
class Test_FMSUB : public ITestCase {
public:
Test_FMSUB(TestSuite* suite) : ITestCase(suite, "fmsub") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] * b[i] - b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -312,22 +380,24 @@ public:
class Test_FNMADD : public ITestCase {
public:
Test_FNMADD(TestSuite* suite) : ITestCase(suite, "fnmadd") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = -a[i] * b[i] - b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -340,22 +410,24 @@ public:
class Test_FNMSUB : public ITestCase {
public:
Test_FNMSUB(TestSuite* suite) : ITestCase(suite, "fnmsub") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = -a[i] * b[i] + b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -368,22 +440,24 @@ public:
class Test_FNMADD_MADD : public ITestCase {
public:
Test_FNMADD_MADD(TestSuite* suite) : ITestCase(suite, "fnmadd-madd") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = -a[i] * b[i] - b[i];
auto y = a[i] * b[i] + b[i];
auto ref = x + y;
@@ -398,22 +472,24 @@ public:
class Test_FDIV : public ITestCase {
public:
Test_FDIV(TestSuite* suite) : ITestCase(suite, "fdiv") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = a[i] / b[i];
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -426,22 +502,24 @@ public:
class Test_FDIV2 : public ITestCase {
public:
Test_FDIV2(TestSuite* suite) : ITestCase(suite, "fdiv2") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = fround((n - i) * (1.0f/n));
b[i] = fround((n + i) * (1.0f/n));
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = a[i] / b[i];
auto y = b[i] / a[i];
auto ref = x + y;
@@ -456,23 +534,25 @@ public:
class Test_FSQRT : public ITestCase {
public:
Test_FSQRT(TestSuite* suite) : ITestCase(suite, "fsqrt") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
float q = 1.0f + (i % 64);
a[i] = q;
b[i] = q;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto ref = sqrt(a[i] * b[i]);
if (!almost_equal(c[i], ref)) {
std::cout << "error at result #" << i << ": expected=" << ref << ", actual=" << c[i] << ", a=" << a[i] << ", b=" << b[i] << std::endl;
@@ -485,22 +565,25 @@ public:
class Test_FTOI : public ITestCase {
public:
Test_FTOI(TestSuite* suite) : ITestCase(suite, "ftoi") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
a[i] = fround((n/2 - i) + (float(i)/n));
b[i] = fround((n/2 - i) + (float(i)/n));
for (uint32_t i = 0; i < n; ++i) {
float q = fround(float(n/2) - i + (float(i) / n));
a[i] = q;
b[i] = q;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (int32_t*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = a[i] + b[i];
auto ref = (int32_t)x;
if (c[i] != ref) {
@@ -514,22 +597,25 @@ public:
class Test_FTOU : public ITestCase {
public:
Test_FTOU(TestSuite* suite) : ITestCase(suite, "ftou") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (float*)src1;
auto b = (float*)src2;
for (int i = 0; i < n; ++i) {
a[i] = fround(i + (float(i)/n));
b[i] = fround(i + (float(i)/n));
for (uint32_t i = 0; i < n; ++i) {
float q = fround(i + (float(i) / n));
a[i] = q;
b[i] = q;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (float*)src1;
auto b = (float*)src2;
auto c = (uint32_t*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = a[i] + b[i];
auto ref = (uint32_t)x;
if (c[i] != ref) {
@@ -543,22 +629,24 @@ public:
class Test_ITOF : public ITestCase {
public:
Test_ITOF(TestSuite* suite) : ITestCase(suite, "itof") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = n/2 - i;
b[i] = n/2 - i;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (int32_t*)src1;
auto b = (int32_t*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = a[i] + b[i];
auto ref = (float)x;
if (!almost_equal(c[i], ref)) {
@@ -572,22 +660,24 @@ public:
class Test_UTOF : public ITestCase {
public:
Test_UTOF(TestSuite* suite) : ITestCase(suite, "utof") {}
void setup(int n, void* src1, void* src2) override {
int setup(uint32_t n, void* src1, void* src2) override {
auto a = (uint32_t*)src1;
auto b = (uint32_t*)src2;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
a[i] = i;
b[i] = i;
}
return 0;
}
int verify(int n, void* dst, const void* src1, const void* src2) override {
int verify(uint32_t n, void* dst, const void* src1, const void* src2) override {
int errors = 0;
auto a = (uint32_t*)src1;
auto b = (uint32_t*)src2;
auto c = (float*)dst;
for (int i = 0; i < n; ++i) {
for (uint32_t i = 0; i < n; ++i) {
auto x = a[i] + b[i];
auto ref = (float)x;
if (!almost_equal(c[i], ref)) {
@@ -597,4 +687,135 @@ public:
}
return errors;
}
};
};
class Test_BAR : public ITestCase {
public:
Test_BAR(TestSuite* suite) : ITestCase(suite, "bar") {}
int setup(uint32_t n, void* src1, void* /*src2*/) override {
RT_CHECK(vx_dev_caps(suite_->device(), VX_CAPS_NUM_WARPS, &num_warps_));
if (num_warps_ == 1) {
std::cout << "Error: multiple warps configuration required!" << std::endl;
return -1;
}
RT_CHECK(vx_dev_caps(suite_->device(), VX_CAPS_NUM_THREADS, &num_threads_));
auto a = (uint32_t*)src1;
for (uint32_t i = 0; i < n; ++i) {
a[i] = i;
}
return 0;
}
int verify(uint32_t n, void* dst, const void* src1, const void* /*src2*/) override {
int errors = 0;
auto a = (uint32_t*)src1;
auto c = (uint32_t*)dst;
for (uint32_t i = 0; i < n; ++i) {
auto tid = i % num_threads_;
auto wid = (i / num_threads_) % num_warps_;
auto cid = i / (num_warps_ * num_threads_);
auto src_idx = (cid * num_warps_ + (num_warps_ - 1 - wid)) * num_threads_ + tid;
uint32_t ref = a[src_idx];
if (c[i] != ref) {
std::cout << "error at result #" << i << ": expected=" << std::hex << ref << ", actual=" << c[i] << std::endl;
++errors;
}
}
return errors;
}
uint64_t num_warps_;
uint64_t num_threads_;
};
class Test_GBAR : public ITestCase {
public:
Test_GBAR(TestSuite* suite) : ITestCase(suite, "gbar") {}
int setup(uint32_t n, void* src1, void* /*src2*/) override {
RT_CHECK(vx_dev_caps(suite_->device(), VX_CAPS_NUM_CORES, &num_cores_));
if (num_cores_ == 1) {
std::cout << "Error: multiple cores configuration required!" << std::endl;
return -1;
}
RT_CHECK(vx_dev_caps(suite_->device(), VX_CAPS_NUM_WARPS, &num_warps_));
RT_CHECK(vx_dev_caps(suite_->device(), VX_CAPS_NUM_THREADS, &num_threads_));
auto a = (uint32_t*)src1;
for (uint32_t i = 0; i < n; ++i) {
a[i] = i;
}
return 0;
}
int verify(uint32_t n, void* dst, const void* src1, const void* /*src2*/) override {
int errors = 0;
auto a = (uint32_t*)src1;
auto c = (uint32_t*)dst;
for (uint32_t i = 0; i < n; ++i) {
auto tid = i % num_threads_;
auto wid = (i / num_threads_) % num_warps_;
auto cid = i / (num_warps_ * num_threads_);
auto src_idx = ((num_cores_ - 1 - cid) * num_warps_ + (num_warps_ - 1 - wid)) * num_threads_ + tid;
uint32_t ref = a[src_idx];
if (c[i] != ref) {
std::cout << "error at result #" << i << ": expected=" << std::hex << ref << ", actual=" << c[i] << std::endl;
++errors;
}
}
return errors;
}
uint64_t num_cores_;
uint64_t num_warps_;
uint64_t num_threads_;
};
///////////////////////////////////////////////////////////////////////////////
TestSuite::TestSuite(vx_device_h device)
: device_(device) {
this->add_test(new Test_IADD(this));
this->add_test(new Test_IMUL(this));
this->add_test(new Test_IDIV(this));
this->add_test(new Test_IDIV_MUL(this));
this->add_test(new Test_FADD(this));
this->add_test(new Test_FSUB(this));
this->add_test(new Test_FMUL(this));
this->add_test(new Test_FMADD(this));
this->add_test(new Test_FMSUB(this));
this->add_test(new Test_FNMADD(this));
this->add_test(new Test_FNMSUB(this));
this->add_test(new Test_FNMADD_MADD(this));
this->add_test(new Test_FDIV(this));
this->add_test(new Test_FDIV2(this));
this->add_test(new Test_FSQRT(this));
this->add_test(new Test_FTOI(this));
this->add_test(new Test_FTOU(this));
this->add_test(new Test_ITOF(this));
this->add_test(new Test_UTOF(this));
this->add_test(new Test_BAR(this));
this->add_test(new Test_GBAR(this));
}
TestSuite::~TestSuite() {
for (size_t i = 0; i < _tests.size(); ++i) {
delete _tests[i];
}
}
ITestCase* TestSuite::get_test(int testid) const {
return _tests.at(testid);
}
void TestSuite::add_test(ITestCase* test) {
_tests.push_back(test);
}
size_t TestSuite::size() const {
return _tests.size();
}
vx_device_h TestSuite::device() const {
return device_;
}