gpu:intel:sycl: use only l0 queries for l0 devices

Author: mgouicem

src/gpu/intel/sycl/device_info.cpp

@@ -19,6 +19,7 @@
 #include "gpu/intel/sycl/compat.hpp"
 #include "gpu/intel/sycl/device_info.hpp"
 #include "gpu/intel/sycl/engine.hpp"
+#include "gpu/intel/sycl/l0/utils.hpp"
 #include "gpu/intel/sycl/utils.hpp"
 
 #include "gpu/intel/ocl/ocl_gpu_hw_info.hpp"
@@ -34,6 +35,7 @@ status_t device_info_t::init_arch(impl::engine_t *engine) {
     auto *sycl_engine
             = utils::downcast<const gpu::intel::sycl::engine_t *>(engine);
     auto &device = sycl_engine->device();
+    auto &ctx = sycl_engine->context();
 
     // skip cpu engines
     if (!device.is_gpu()) return status::success;
@@ -43,34 +45,23 @@ status_t device_info_t::init_arch(impl::engine_t *engine) {
 
     auto be = xpu::sycl::get_backend(device);
     if (be == xpu::sycl::backend_t::opencl) {
-        cl_int err = CL_SUCCESS;
-
         auto ocl_dev = xpu::sycl::compat::get_native<cl_device_id>(device);
         auto ocl_dev_wrapper = xpu::ocl::make_wrapper(ocl_dev);
 
-        auto ocl_ctx_wrapper = xpu::ocl::make_wrapper(
-                clCreateContext(nullptr, 1, &ocl_dev, nullptr, nullptr, &err));
-        OCL_CHECK(err);
+        auto ocl_ctx = xpu::sycl::compat::get_native<cl_context>(ctx);
+        auto ocl_ctx_wrapper = xpu::ocl::make_wrapper(ocl_ctx);
 
         gpu::intel::ocl::init_gpu_hw_info(engine, ocl_dev_wrapper,
                 ocl_ctx_wrapper, ip_version_, gpu_arch_, gpu_product_family_,
                 stepping_id_, native_extensions_, mayiuse_systolic_,
                 mayiuse_ngen_kernels_);
     } else if (be == xpu::sycl::backend_t::level0) {
-        // TODO: add support for L0 binary ngen check
-        // XXX: query from ocl_engine for now
-        std::unique_ptr<gpu::intel::ocl::ocl_gpu_engine_t, engine_deleter_t>
-                ocl_engine;
-        CHECK(gpu::intel::sycl::create_ocl_engine(&ocl_engine, sycl_engine));
-
-        auto *dev_info = ocl_engine->device_info();
-        ip_version_ = dev_info->ip_version();
-        gpu_arch_ = dev_info->gpu_arch();
-        gpu_product_family_ = dev_info->gpu_product_family();
-        stepping_id_ = dev_info->stepping_id();
-        native_extensions_ = dev_info->native_extensions();
-        mayiuse_systolic_ = dev_info->mayiuse_systolic();
-        mayiuse_ngen_kernels_ = dev_info->mayiuse_ngen_kernels();
+        auto ze_dev = xpu::sycl::compat::get_native<ze_device_handle_t>(device);
+        auto ze_ctx = xpu::sycl::compat::get_native<ze_context_handle_t>(ctx);
+
+        gpu::intel::sycl::init_gpu_hw_info(engine, ze_dev, ze_ctx, ip_version_,
+                gpu_arch_, gpu_product_family_, stepping_id_,
+                native_extensions_, mayiuse_systolic_, mayiuse_ngen_kernels_);
     } else {
         assert(!"not_expected");
     }

src/gpu/intel/sycl/l0/utils.cpp

@@ -17,6 +17,10 @@
 #include "gpu/intel/sycl/l0/utils.hpp"
 #include "oneapi/dnnl/dnnl_config.h"
 
+#include "gpu/intel/jit/binary_format.hpp"
+#include "gpu/intel/jit/ngen/ngen_level_zero.hpp"
+#include "gpu/intel/jit/utils/ngen_type_bridge.hpp"
+
 #if defined(__linux__)
 #include <dlfcn.h>
 #elif defined(_WIN32)
@@ -26,6 +30,7 @@
 #endif
 
 #include "gpu/intel/sycl/l0/level_zero/ze_api.h"
+#include "gpu/intel/sycl/l0/ze_intel_gpu.h"
 
 #if !defined(__SYCL_COMPILER_VERSION)
 #error "Unsupported compiler"
@@ -173,6 +178,16 @@ status_t func_zeDeviceGetProperties(
     return status::success;
 }
 
+status_t func_zeDeviceGetModuleProperties(ze_device_handle_t hDevice,
+        ze_device_module_properties_t *pDeviceProperties) {
+    static auto f = find_ze_symbol<decltype(&zeDeviceGetModuleProperties)>(
+            "zeDeviceGetModuleProperties");
+
+    if (!f) return status::runtime_error;
+    ZE_CHECK(f(hDevice, pDeviceProperties));
+    return status::success;
+}
+
 } // namespace
 
 // This function is called from compatibility layer that ensures compatibility
@@ -272,6 +287,115 @@ bool compare_ze_devices(const ::sycl::device &lhs, const ::sycl::device &rhs) {
     return lhs_ze_handle == rhs_ze_handle;
 }
 
+status_t get_device_ip(ze_device_handle_t device, uint32_t &ip_version) {
+    ze_device_properties_t deviceProps = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
+    ze_device_ip_version_ext_t devicePropsIP
+            = {ZE_STRUCTURE_TYPE_DEVICE_IP_VERSION_EXT};
+    deviceProps.pNext = &devicePropsIP;
+    CHECK(func_zeDeviceGetProperties(device, &deviceProps));
+    ip_version = devicePropsIP.ipVersion;
+    return status::success;
+}
+
+status_t get_l0_device_enabled_systolic_intel(
+        ze_device_handle_t device, bool &mayiuse_systolic) {
+    ze_device_module_properties_t deviceModProps
+            = {ZE_STRUCTURE_TYPE_DEVICE_MODULE_PROPERTIES};
+    // Note: supported by Intel Driver 24.05 and onwards
+    ze_intel_device_module_dp_exp_properties_t deviceModPropsExt
+            = {ZE_STRUCTURE_INTEL_DEVICE_MODULE_DP_EXP_PROPERTIES};
+    deviceModProps.pNext = &deviceModPropsExt;
+
+    CHECK(func_zeDeviceGetModuleProperties(device, &deviceModProps));
+    mayiuse_systolic
+            = deviceModPropsExt.flags & ZE_INTEL_DEVICE_MODULE_EXP_FLAG_DPAS;
+    return status::success;
+}
+
+status_t get_l0_device_enabled_native_float_atomics(
+        ze_device_handle_t device, uint64_t native_extensions) {
+    using namespace gpu::intel::compute;
+
+    ze_device_properties_t deviceProps = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
+    ze_float_atomic_ext_properties_t fltAtom
+            = {ZE_STRUCTURE_TYPE_FLOAT_ATOMIC_EXT_PROPERTIES};
+    deviceProps.pNext = &fltAtom;
+    CHECK(func_zeDeviceGetProperties(device, &deviceProps));
+
+    ze_device_fp_atomic_ext_flags_t atomic_load_store
+            = ZE_DEVICE_FP_ATOMIC_EXT_FLAG_GLOBAL_LOAD_STORE
+            | ZE_DEVICE_FP_ATOMIC_EXT_FLAG_LOCAL_LOAD_STORE;
+    ze_device_fp_atomic_ext_flags_t atomic_add
+            = ZE_DEVICE_FP_ATOMIC_EXT_FLAG_GLOBAL_ADD
+            | ZE_DEVICE_FP_ATOMIC_EXT_FLAG_LOCAL_ADD;
+    ze_device_fp_atomic_ext_flags_t atomic_min_max
+            = ZE_DEVICE_FP_ATOMIC_EXT_FLAG_GLOBAL_MIN_MAX
+            | ZE_DEVICE_FP_ATOMIC_EXT_FLAG_LOCAL_MIN_MAX;
+
+    if ((fltAtom.fp16Flags & atomic_load_store) == atomic_load_store)
+        native_extensions |= (uint64_t)native_ext_t::fp16_atomic_load_store;
+    if ((fltAtom.fp16Flags & atomic_add) == atomic_add)
+        native_extensions |= (uint64_t)native_ext_t::fp16_atomic_add;
+    if ((fltAtom.fp16Flags & atomic_add) == atomic_min_max)
+        native_extensions |= (uint64_t)native_ext_t::fp16_atomic_min_max;
+
+    if ((fltAtom.fp32Flags & atomic_load_store) == atomic_load_store)
+        native_extensions |= (uint64_t)native_ext_t::fp32_atomic_load_store;
+    if ((fltAtom.fp32Flags & atomic_add) == atomic_add)
+        native_extensions |= (uint64_t)native_ext_t::fp32_atomic_add;
+    if ((fltAtom.fp32Flags & atomic_add) == atomic_min_max)
+        native_extensions |= (uint64_t)native_ext_t::fp32_atomic_min_max;
+
+    if ((fltAtom.fp64Flags & atomic_load_store) == atomic_load_store)
+        native_extensions |= (uint64_t)native_ext_t::fp64_atomic_load_store;
+    if ((fltAtom.fp64Flags & atomic_add) == atomic_add)
+        native_extensions |= (uint64_t)native_ext_t::fp64_atomic_add;
+    if ((fltAtom.fp64Flags & atomic_add) == atomic_min_max)
+        native_extensions |= (uint64_t)native_ext_t::fp64_atomic_min_max;
+
+    return status::success;
+}
+
+status_t get_l0_device_eu_count(ze_device_handle_t device, int &eucount) {
+    ze_device_properties_t deviceProps = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
+    ze_eu_count_ext_t eucnt = ze_eu_count_ext_t();
+    deviceProps.pNext = &eucnt;
+
+    CHECK(func_zeDeviceGetProperties(device, &deviceProps));
+    eucount = eucnt.numTotalEUs;
+    return status::success;
+}
+
+void init_gpu_hw_info(impl::engine_t *engine, ze_device_handle_t device,
+        ze_context_handle_t context, uint32_t &ip_version,
+        compute::gpu_arch_t &gpu_arch, int &gpu_product_family,
+        int &stepping_id, uint64_t &native_extensions, bool &mayiuse_systolic,
+        bool &mayiuse_ngen_kernels) {
+    using namespace ngen;
+    HW hw = HW::Unknown;
+    Product product = {ProductFamily::Unknown, 0};
+    LevelZeroCodeGenerator<HW::Unknown>::detectHWInfo(
+            context, device, hw, product);
+
+    gpu_arch = jit::convert_ngen_arch_to_dnnl(hw);
+    gpu_product_family = static_cast<int>(product.family);
+    stepping_id = product.stepping;
+
+    mayiuse_systolic = false;
+    status_t ret
+            = get_l0_device_enabled_systolic_intel(device, mayiuse_systolic);
+    // TODO: xelpg has no f64 support. check that the query properly handle that
+    ret = get_l0_device_enabled_native_float_atomics(device, native_extensions);
+    MAYBE_UNUSED(ret);
+
+    auto status
+            = jit::gpu_supports_binary_format(&mayiuse_ngen_kernels, engine);
+    if (status != status::success) mayiuse_ngen_kernels = false;
+
+    ip_version = 0;
+    get_device_ip(device, ip_version);
+}
+
 } // namespace sycl
 } // namespace intel
 } // namespace gpu

src/gpu/intel/sycl/l0/utils.hpp

@@ -45,6 +45,12 @@ bool compare_ze_devices(const ::sycl::device &lhs, const ::sycl::device &rhs);
 status_t func_zeModuleGetNativeBinary(ze_module_handle_t hModule, size_t *pSize,
         uint8_t *pModuleNativeBinary);
 
+void init_gpu_hw_info(impl::engine_t *engine, ze_device_handle_t device,
+        ze_context_handle_t context, uint32_t &ip_version,
+        compute::gpu_arch_t &gpu_arch, int &gpu_product_family,
+        int &stepping_id, uint64_t &native_extensions, bool &mayiuse_systolic,
+        bool &mayiuse_ngen_kernels);
+
 } // namespace sycl
 } // namespace intel
 } // namespace gpu