mirror of
https://github.com/fumiama/gozel.git
synced 2026-06-27 06:10:27 +08:00
feat(ze): add event support & vadd demo & refactor (#5)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
This commit is contained in:
fumiama
GitHub
231
examples/vadd_event/main.go
Normal file
231
examples/vadd_event/main.go
Normal file
@@ -0,0 +1,231 @@
|
||||
// Package main demonstrates vector addition using the gozel Level Zero bindings.
|
||||
package main
|
||||
|
||||
import (
|
||||
_ "embed"
|
||||
"fmt"
|
||||
"math"
|
||||
"math/rand"
|
||||
"os"
|
||||
"strconv"
|
||||
"strings"
|
||||
"time"
|
||||
"unsafe"
|
||||
|
||||
"github.com/fumiama/gozel/gozel"
|
||||
"github.com/fumiama/gozel/ze"
|
||||
)
|
||||
|
||||
//go:generate clang++ -fsycl -fsycl-device-only -fsycl-targets=spirv64 -Xclang -emit-llvm-bc main.cpp -o device_kern.bc
|
||||
//go:generate sycl-post-link -symbols -split=auto -o device_kern.table device_kern.bc
|
||||
//go:generate clang++ -target spirv64-unknown-unknown -S -emit-llvm -x ir device_kern_0.bc -o device_kern.ll
|
||||
//go:generate llvm-spirv -o main.spv device_kern.bc
|
||||
//go:generate clang++ -target spirv64-unknown-unknown -S -emit-llvm -x ir device_kern.bc -o main.ll
|
||||
|
||||
//go:embed main.spv
|
||||
var kernelspv []byte
|
||||
|
||||
func main() {
|
||||
gpus, err := ze.InitGPUDrivers()
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
if len(gpus) == 0 {
|
||||
panic("no gpu available")
|
||||
}
|
||||
gpu := gpus[0]
|
||||
|
||||
ctx, err := gpu.ContextCreate()
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
devs, err := gpu.DeviceGet()
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
if len(devs) == 0 {
|
||||
panic("no device available")
|
||||
}
|
||||
dev := devs[0]
|
||||
|
||||
prop, err := dev.DeviceGetProperties()
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
fmt.Println("=============== Device Basic Properties ===============")
|
||||
name, _, _ := strings.Cut(string(prop.Name[:]), "\x00")
|
||||
fmt.Println(
|
||||
"Running on device: ID =", prop.Deviceid, ", Name =", name,
|
||||
"@", strconv.FormatFloat(float64(prop.Coreclockrate)/1024/1024/1024, 'f', 2, 64), "GHz.",
|
||||
)
|
||||
|
||||
cprop, err := dev.DeviceGetComputeProperties()
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
fmt.Println("=============== Device Compute Properties ===============")
|
||||
fmt.Printf("%-28s (%d, %d, %d)\n", "Max Group Size (X, Y, Z):", cprop.Maxgroupsizex, cprop.Maxgroupsizey, cprop.Maxgroupsizez)
|
||||
fmt.Printf("%-28s (%d, %d, %d)\n", "Max Group Count (X, Y, Z):", cprop.Maxgroupcountx, cprop.Maxgroupcounty, cprop.Maxgroupcountz)
|
||||
fmt.Printf("%-28s %d\n", "Max Total Group Size:", cprop.Maxtotalgroupsize)
|
||||
fmt.Printf("%-28s %d\n", "Max Shared Local Memory:", cprop.Maxsharedlocalmemory)
|
||||
fmt.Printf("%-28s %d\n", "Num Subgroup Sizes:", cprop.Numsubgroupsizes)
|
||||
fmt.Printf("%-28s %v\n", "Subgroup Sizes:", cprop.Subgroupsizes[:])
|
||||
|
||||
var (
|
||||
X, Y, Z = uintptr(cprop.Maxgroupsizex), uintptr(1), uintptr(1)
|
||||
groupCount = uintptr(65536)
|
||||
N = X * groupCount
|
||||
bufsz = N * unsafe.Sizeof(float32(0))
|
||||
)
|
||||
fmt.Println("=============== Computation Configuration ===============")
|
||||
fmt.Printf("%-28s (%d, %d, %d)\n", "Group Size (X, Y, Z):", X, Y, Z)
|
||||
fmt.Printf("%-28s %d\n", "Group Count:", groupCount)
|
||||
fmt.Printf("%-28s %d\n", "Total Elements (N):", N)
|
||||
fmt.Printf("%-28s %d MiB\n", "Buffer Size:", bufsz/1024/1024)
|
||||
|
||||
hbufV1, err := ctx.MemAllocHost(bufsz, 1)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer ctx.MemFree(hbufV1)
|
||||
|
||||
hbufV2, err := ctx.MemAllocHost(bufsz, 1)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer ctx.MemFree(hbufV2)
|
||||
|
||||
dbufV1, err := ctx.MemAllocDevice(dev, bufsz, 1)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer ctx.MemFree(dbufV1)
|
||||
|
||||
dbufV2, err := ctx.MemAllocDevice(dev, bufsz, 1)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer ctx.MemFree(dbufV2)
|
||||
|
||||
floatbuf := make([]float32, 2*N)
|
||||
for i := range floatbuf {
|
||||
floatbuf[i] = rand.Float32()
|
||||
}
|
||||
|
||||
zev1, zev2 := unsafe.Slice((*float32)(hbufV1), N), unsafe.Slice((*float32)(hbufV2), N)
|
||||
copy(zev1, floatbuf[:N])
|
||||
copy(zev2, floatbuf[N:])
|
||||
|
||||
mod, err := ctx.ModuleCreate(dev, kernelspv)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer mod.Destroy()
|
||||
|
||||
krn, err := mod.KernelCreate("__sycl_kernel_vector_add")
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer krn.Destroy()
|
||||
|
||||
err = krn.SetArgumentValue(0, &dbufV1)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
err = krn.SetArgumentValue(1, &dbufV2)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
err = krn.SetGroupSize(uint32(X), uint32(Y), uint32(Z))
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
evp, err := ctx.EventPoolCreate(3, dev)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer evp.Destroy()
|
||||
|
||||
evcph2dv1, err := evp.EventCreate(0, gozel.ZE_EVENT_SCOPE_FLAG_HOST, 0)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer evcph2dv1.Destroy()
|
||||
evcph2dv2, err := evp.EventCreate(1, gozel.ZE_EVENT_SCOPE_FLAG_HOST, 0)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer evcph2dv2.Destroy()
|
||||
|
||||
start := time.Now()
|
||||
|
||||
lst, err := ctx.CommandListCreateImmediate(dev, gozel.ZE_COMMAND_QUEUE_MODE_DEFAULT)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
err = lst.AppendMemoryCopy(dbufV1, hbufV1, bufsz, evcph2dv1)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
err = lst.AppendMemoryCopy(dbufV2, hbufV2, bufsz, evcph2dv2)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
evk, err := evp.EventCreate(2, gozel.ZE_EVENT_SCOPE_FLAG_HOST, 0)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
defer evk.Destroy()
|
||||
|
||||
err = lst.AppendLaunchKernel(krn, &gozel.ZeGroupCount{
|
||||
Groupcountx: uint32(groupCount), Groupcounty: 1, Groupcountz: 1,
|
||||
}, evk, evcph2dv1, evcph2dv2)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
err = lst.AppendMemoryCopy(hbufV1, dbufV1, bufsz, 0, evk)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
err = lst.HostSynchronize(math.MaxUint64)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
elapsed := time.Since(start)
|
||||
|
||||
fmt.Println("=============== Calculation Results ===============")
|
||||
fmt.Printf("%-28s %.6f ms\n", "GPU Execution Time:", elapsed.Seconds()*1000)
|
||||
fmt.Printf("%-28s %.2f GiB/s\n", "GPU Throughput:", float64(bufsz)/elapsed.Seconds()/1e9)
|
||||
|
||||
tmpbuf := make([]float32, N)
|
||||
start = time.Now()
|
||||
for i := range N {
|
||||
tmpbuf[i] = floatbuf[i] + floatbuf[N+i]
|
||||
}
|
||||
elapsed = time.Since(start)
|
||||
|
||||
fmt.Println("=============== Validation Results ===============")
|
||||
fmt.Printf("%-28s %.6f ms\n", "CPU Execution Time:", elapsed.Seconds()*1000)
|
||||
fmt.Printf("%-28s %.2f GiB/s\n", "CPU Throughput:", float64(bufsz)/elapsed.Seconds()/1e9)
|
||||
|
||||
fail := false
|
||||
for i := range N {
|
||||
expect := floatbuf[i] + floatbuf[N+i]
|
||||
if zev1[i] != expect {
|
||||
fail = true
|
||||
fmt.Printf("[%05d] expect %f = %f + %f, got %f.\n", i, expect, floatbuf[i], floatbuf[N+i], zev1[i])
|
||||
}
|
||||
}
|
||||
|
||||
if fail {
|
||||
os.Exit(1)
|
||||
}
|
||||
|
||||
fmt.Println("Test Passed!!!")
|
||||
}
|
||||
Reference in New Issue
Block a user