[Beignet] [PATCH v2] Enable OpenCL 2.0 only where supported
Bruno Pagani
bruno.n.pagani at gmail.com
Mon Jan 23 09:21:11 UTC 2017
Yes indeed, I think Rebecca patch is not ideal in that direction.
Removing the lspci thing is right (by the way, I think that GetGenID.sh
should be removed too at this point), but regarding whether OpenCL 2.0
is compiled or not I would rather think of something like that:
– if llvm39 and libdrm2466 are found, compile with OpenCL 2.0 by default.
– if they’re not, compile without it. Issue an error message if
-DENABLE_OPENCL_20=1 was passed.
Bruno
Le 23/01/2017 à 03:44, Pan, Xiuli a écrit :
> I think if we want to enable OpenCL 2.0 as default we must check if there is a LLVM 3.9 first, or there will have errors to ask people to update llvm. We may turn the OCL 20 off when no llvm39 is found.
>
> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of Rebecca N. Palmer
> Sent: Sunday, January 22, 2017 9:16 PM
> To: Bruno Pagani <bruno.n.pagani at gmail.com>; beignet at lists.freedesktop.org; ignatenko at redhat.com
> Subject: [Beignet] [PATCH v2] Enable OpenCL 2.0 only where supported
>
> This allows a single beignet binary to both offer 2.0 where available, and still work on older hardware.
> (2.0 uses 64 bit pointers, which assert-fail on Ivybridge:
> backend/src/backend/gen_context.cpp:2259).
>
> V2:
> Default to 1.2 when -cl-std is not set (required by the OpenCL spec, and also likely to be faster). Update documentation.
> Unconditionally build 2.0-only tests, but skip them on non-2.0 hardware. Add -cl-std=2.0 to tests that need it.
>
> Signed-off-by: Rebecca N. Palmer <rebecca_palmer at zoho.com>
>
> --- a/CMakeLists.txt
> +++ b/CMakeLists.txt
> @@ -231,19 +231,8 @@ IF (EXPERIMENTAL_DOUBLE)
> ADD_DEFINITIONS(-DENABLE_FP64)
> ENDIF(EXPERIMENTAL_DOUBLE)
>
> -OPTION(ENABLE_OPENCL_20 "Enable opencl 2.0 support" OFF)
> +OPTION(ENABLE_OPENCL_20 "Enable opencl 2.0 support" ON)
> IF (ENABLE_OPENCL_20)
> - Find_Program(LSPCI lspci)
> - IF (NOT LSPCI)
> - MESSAGE(FATAL_ERROR "Looking for lspci - not found")
> - ENDIF (NOT LSPCI)
> - EXECUTE_PROCESS(COMMAND "${CMAKE_CURRENT_SOURCE_DIR}/GetGenID.sh"
> - RESULT_VARIABLE SUPPORT_OCL20_DEVICE
> - OUTPUT_VARIABLE PCI_ID_NOT_USED)
> -
> - IF (NOT SUPPORT_OCL20_DEVICE EQUAL 1)
> - MESSAGE(FATAL_ERROR "Only SKL and newer devices support OpenCL 2.0 now, your device don't support.")
> - ENDIF (NOT SUPPORT_OCL20_DEVICE EQUAL 1)
>
> IF (NOT HAVE_DRM_INTEL_BO_SET_SOFTPIN)
> MESSAGE(FATAL_ERROR "Please update libdrm to version 2.4.66 or later to enable OpenCL 2.0.")
> --- a/backend/src/backend/program.cpp
> +++ b/backend/src/backend/program.cpp
> @@ -31,6 +31,7 @@
> #include "ir/value.hpp"
> #include "ir/unit.hpp"
> #include "ir/printf.hpp"
> +#include "../src/cl_device_data.h"
>
> #ifdef GBE_COMPILER_AVAILABLE
> #include "llvm/llvm_to_gen.hpp"
> @@ -855,6 +856,7 @@ namespace gbe {
> size_t *errSize,
> uint32_t &oclVersion)
> {
> + uint32_t maxoclVersion = oclVersion;
> std::string pchFileName;
> bool findPCH = false;
> #if defined(__ANDROID__)
> @@ -1022,15 +1024,9 @@ EXTEND_QUOTE:
> }
>
> if (useDefaultCLCVersion) {
> -#ifdef ENABLE_OPENCL_20
> - clOpt.push_back("-D__OPENCL_C_VERSION__=200");
> - clOpt.push_back("-cl-std=CL2.0");
> - oclVersion = 200;
> -#else
> clOpt.push_back("-D__OPENCL_C_VERSION__=120");
> clOpt.push_back("-cl-std=CL1.2");
> oclVersion = 120;
> -#endif
> }
> //for clCompilerProgram usage.
> if(temp_header_path){
> @@ -1061,7 +1057,12 @@ EXTEND_QUOTE:
> clOpt.push_back("-include-pch");
> clOpt.push_back(pchFileName);
> }
> -
> + if (oclVersion > maxoclVersion){
> + if (err && stringSize > 0 && errSize) {
> + *errSize = snprintf(err, stringSize, "Requested OpenCL version %lf is higher than maximum supported version %lf\n", (float)oclVersion/100.0,(float)maxoclVersion/100.0);
> + }
> + return false;
> + }
> return true;
> }
>
> @@ -1076,7 +1077,7 @@ EXTEND_QUOTE:
> std::vector<std::string> clOpt;
> std::string dumpLLVMFileName, dumpASMFileName;
> std::string dumpSPIRBinaryName;
> - uint32_t oclVersion = 0;
> + uint32_t oclVersion = MAX_OCLVERSION(deviceID);
> if (!processSourceAndOption(source, options, NULL, clOpt,
> dumpLLVMFileName, dumpASMFileName, dumpSPIRBinaryName,
> optLevel, @@ -1139,7 +1140,7 @@ EXTEND_QUOTE:
> std::vector<std::string> clOpt;
> std::string dumpLLVMFileName, dumpASMFileName;
> std::string dumpSPIRBinaryName;
> - uint32_t oclVersion = 0;
> + uint32_t oclVersion = MAX_OCLVERSION(deviceID);
> if (!processSourceAndOption(source, options, temp_header_path, clOpt,
> dumpLLVMFileName, dumpASMFileName, dumpSPIRBinaryName,
> optLevel, stringSize, err, errSize, oclVersion))
> --- a/src/cl_device_data.h
> +++ b/src/cl_device_data.h
> @@ -363,5 +363,7 @@
>
> #define IS_GEN9(devid) (IS_SKYLAKE(devid) || IS_BROXTON(devid) || IS_KABYLAKE(devid))
>
> +#define MAX_OCLVERSION(devid) (IS_GEN9(devid) ? 200 : 120)
> +
> #endif /* __CL_DEVICE_DATA_H__ */
>
> --- a/src/cl_gen9_device.h
> +++ b/src/cl_gen9_device.h
> @@ -27,5 +27,7 @@
> .max_mem_alloc_size = 4 * 1024 * 1024 * 1024ul, .global_mem_size = 4 * 1024 * 1024 * 1024ul,
>
> +#define GEN9_DEVICE 1
> #include "cl_gt_device.h"
> +#undef GEN9_DEVICE
>
> --- a/src/cl_gt_device.h
> +++ b/src/cl_gt_device.h
> @@ -16,7 +16,13 @@
> *
> * Author: Benjamin Segovia <benjamin.segovia at intel.com>
> */
> -
> +#ifdef GEN9_DEVICE
> +#define LIBCL_VERSION_STRING GEN9_LIBCL_VERSION_STRING #define
> +LIBCL_C_VERSION_STRING GEN9_LIBCL_C_VERSION_STRING #else #define
> +LIBCL_VERSION_STRING NONGEN9_LIBCL_VERSION_STRING #define
> +LIBCL_C_VERSION_STRING NONGEN9_LIBCL_C_VERSION_STRING #endif
> /* Common fields for both all GT devices (IVB / SNB) */ .device_type = CL_DEVICE_TYPE_GPU,
> .device_id=0,/* == device_id (set when requested) */ @@ -39,7 +45,7 @@ .native_vector_width_float = 4, .native_vector_width_double = 2, .native_vector_width_half = 8, -#ifdef ENABLE_OPENCL_20
> +#if defined(ENABLE_OPENCL_20) && defined (GEN9_DEVICE)
> .address_bits = 64,
> #else
> .address_bits = 32,
> --- a/src/cl_platform_id.c
> +++ b/src/cl_platform_id.c
> @@ -32,7 +32,7 @@
>
> static struct _cl_platform_id intel_platform_data = {
> DECL_INFO_STRING(profile, "FULL_PROFILE")
> - DECL_INFO_STRING(version, LIBCL_VERSION_STRING)
> + DECL_INFO_STRING(version, GEN9_LIBCL_VERSION_STRING)
> DECL_INFO_STRING(name, "Intel Gen OCL Driver")
> DECL_INFO_STRING(vendor, "Intel")
> DECL_INFO_STRING(icd_suffix_khr, "Intel")
> --- a/src/cl_platform_id.h
> +++ b/src/cl_platform_id.h
> @@ -72,8 +72,10 @@ extern cl_int cl_get_platform_ids(cl_uin #else #define LIBCL_DRIVER_VERSION_STRING _JOINT(LIBCL_DRIVER_VERSION_MAJOR, LIBCL_DRIVER_VERSION_MINOR) #endif -#define LIBCL_VERSION_STRING "OpenCL " _JOINT(LIBCL_C_VERSION_MAJOR, LIBCL_C_VERSION_MINOR) " beignet " LIBCL_DRIVER_VERSION_STRING BEIGNET_GIT_SHA1_STRING -#define LIBCL_C_VERSION_STRING "OpenCL C " _JOINT(LIBCL_C_VERSION_MAJOR, LIBCL_C_VERSION_MINOR) " beignet " LIBCL_DRIVER_VERSION_STRING BEIGNET_GIT_SHA1_STRING
> +#define GEN9_LIBCL_VERSION_STRING "OpenCL "
> +_JOINT(LIBCL_C_VERSION_MAJOR, LIBCL_C_VERSION_MINOR) " beignet "
> +LIBCL_DRIVER_VERSION_STRING BEIGNET_GIT_SHA1_STRING #define
> +GEN9_LIBCL_C_VERSION_STRING "OpenCL C " _JOINT(LIBCL_C_VERSION_MAJOR,
> +LIBCL_C_VERSION_MINOR) " beignet " LIBCL_DRIVER_VERSION_STRING
> +BEIGNET_GIT_SHA1_STRING #define NONGEN9_LIBCL_VERSION_STRING "OpenCL
> +1.2 beignet " LIBCL_DRIVER_VERSION_STRING BEIGNET_GIT_SHA1_STRING
> +#define NONGEN9_LIBCL_C_VERSION_STRING "OpenCL C 1.2 beignet "
> +LIBCL_DRIVER_VERSION_STRING BEIGNET_GIT_SHA1_STRING
>
> #endif /* __CL_PLATFORM_ID_H__ */
>
> --- a/utests/CMakeLists.txt
> +++ b/utests/CMakeLists.txt
> @@ -297,7 +297,6 @@ if (LLVM_VERSION_NODOT VERSION_GREATER 34)
> compiler_overflow.cpp)
> endif (LLVM_VERSION_NODOT VERSION_GREATER 34)
>
> -if (ENABLE_OPENCL_20)
> SET(utests_sources
> ${utests_sources}
> compiler_program_global.cpp
> @@ -308,7 +307,6 @@ if (ENABLE_OPENCL_20)
> runtime_pipe_query.cpp
> compiler_pipe_builtin.cpp
> compiler_device_enqueue.cpp)
> -endif (ENABLE_OPENCL_20)
>
> if (NOT_BUILD_STAND_ALONE_UTEST)
> if (X11_FOUND)
> --- a/utests/compiler_atomic_functions_20.cpp
> +++ b/utests/compiler_atomic_functions_20.cpp
> @@ -64,8 +64,9 @@ static void compiler_atomic_functions(co
> globals[0] = n;
> locals[0] = LOCAL_SIZE;
>
> + if(!cl_check_ocl20(false)){return;}
> // Setup kernel and buffers
> - OCL_CREATE_KERNEL_FROM_FILE("compiler_atomic_functions_20", kernel_name);
> + OCL_CALL(cl_kernel_init, "compiler_atomic_functions_20.cl",
> + kernel_name, SOURCE, "-cl-std=CL2.0");
> OCL_CREATE_BUFFER(buf[0], 0, 24 * sizeof(int), NULL);
> OCL_CREATE_BUFFER(buf[1], 0, locals[0] * sizeof(int), NULL);
> OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
> --- a/utests/compiler_device_enqueue.cpp
> +++ b/utests/compiler_device_enqueue.cpp
> @@ -6,8 +6,9 @@ void compiler_device_enqueue(void)
> const uint32_t global_sz = 3;
> uint32_t result = 0;
>
> + if(!cl_check_ocl20(false)){return;}
> // Setup kernel and buffers
> - OCL_CREATE_KERNEL("compiler_device_enqueue");
> + OCL_CALL(cl_kernel_init, "compiler_device_enqueue.cl",
> + "compiler_device_enqueue", SOURCE, "-cl-std=CL2.0");
> OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(uint32_t), NULL);
> OCL_SET_ARG(0, sizeof(uint32_t), &global_sz);
> OCL_SET_ARG(1, sizeof(cl_mem), &buf[0]);
> --- a/utests/compiler_generic_atomic.cpp
> +++ b/utests/compiler_generic_atomic.cpp
> @@ -6,6 +6,7 @@ void test_atomic(const char* kernelName)
> const int n = 16;
> T cpu_src[16];
>
> + if(!cl_check_ocl20(false)){return;}
> // Setup kernel and buffers
> OCL_CALL(cl_kernel_init, "compiler_generic_atomic.cl", kernelName, SOURCE, "-cl-std=CL2.0");
> OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(T), NULL);
> --- a/utests/compiler_generic_pointer.cpp
> +++ b/utests/compiler_generic_pointer.cpp
> @@ -6,6 +6,7 @@ void test(const char* kernelName)
> const int n = 16;
> T cpu_src[16];
>
> + if(!cl_check_ocl20(false)){return;}
> // Setup kernel and buffers
> OCL_CALL(cl_kernel_init, "compiler_generic_pointer.cl", kernelName, SOURCE, "-cl-std=CL2.0");
> OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(T), NULL);
> --- a/utests/compiler_pipe_builtin.cpp
> +++ b/utests/compiler_pipe_builtin.cpp
> @@ -11,8 +11,9 @@ static void compiler_pipe_##GROUP##_##TYPE(void) \
> const size_t w = 16; \
> uint32_t ans_host = 0; \
> uint32_t ans_device = 0; \
> + if(!cl_check_ocl20(false)){return;}\
> /* pipe write kernel*/ \
> - OCL_CREATE_KERNEL_FROM_FILE("compiler_pipe_builtin", "compiler_pipe_"#GROUP"_write_"#TYPE); \
> + OCL_CALL(cl_kernel_init, "compiler_pipe_builtin.cl",
> + "compiler_pipe_"#GROUP"_write_"#TYPE, SOURCE, "-cl-std=CL2.0");\
> OCL_CALL2(clCreatePipe, buf[0], ctx, 0, sizeof(TYPE), w, NULL);\
> OCL_CREATE_BUFFER(buf[1], CL_MEM_READ_WRITE, w * sizeof(TYPE), NULL);\
> OCL_MAP_BUFFER(1);\
> @@ -26,7 +27,7 @@ static void compiler_pipe_##GROUP##_##TYPE(void) \
> OCL_NDRANGE(1);\
> OCL_CALL(clReleaseKernel, kernel);\
> /* pipe read kernel */\
> - OCL_CREATE_KERNEL_FROM_FILE("compiler_pipe_builtin", "compiler_pipe_"#GROUP"_read_"#TYPE);\
> + OCL_CALL(cl_kernel_init, "compiler_pipe_builtin.cl",
> + "compiler_pipe_"#GROUP"_read_"#TYPE, SOURCE, "-cl-std=CL2.0");\
> OCL_CREATE_BUFFER(buf[2], CL_MEM_READ_WRITE, w * sizeof(TYPE), NULL);\
> OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);\
> OCL_SET_ARG(1, sizeof(cl_mem), &buf[2]);\ @@ -52,8 +53,9 @@ PIPE_BUILTIN(mystruct, workgroup) static void compiler_pipe_query(void) {
> const size_t w = 32;
> const size_t sz = 16;
> + if(!cl_check_ocl20(false)){return;}
> /* pipe write kernel */
> - OCL_CREATE_KERNEL_FROM_FILE("compiler_pipe_builtin", "compiler_pipe_query");
> + OCL_CALL(cl_kernel_init, "compiler_pipe_builtin.cl",
> + "compiler_pipe_query", SOURCE, "-cl-std=CL2.0");
> OCL_CALL2(clCreatePipe, buf[0], ctx, 0, sizeof(uint32_t), w, NULL);
> OCL_CREATE_BUFFER(buf[1], CL_MEM_READ_WRITE, sz * sizeof(uint32_t), NULL);
> OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
> --- a/utests/compiler_program_global.cpp
> +++ b/utests/compiler_program_global.cpp
> @@ -25,6 +25,7 @@ void compiler_program_global()
> int cpu_src[16];
> cl_int err;
>
> + if(!cl_check_ocl20(false)){return;}
> // Setup kernel and buffers
> cl_program program;
> init_program("compiler_program_global.cl", ctx, &program);
> --- a/utests/runtime_pipe_query.cpp
> +++ b/utests/runtime_pipe_query.cpp
> @@ -4,6 +4,7 @@ static void runtime_pipe_query(void) {
> const size_t w = 16;
> const size_t sz = 8;
> cl_uint retnum, retsz;
> + if(!cl_check_ocl20(false)){return;}
> /* pipe write kernel */
> OCL_CALL2(clCreatePipe, buf[0], ctx, 0, sz, w, NULL);
> OCL_CALL(clGetPipeInfo, buf[0], CL_PIPE_MAX_PACKETS, sizeof(retnum), &retnum, NULL);
> --- a/utests/utest_helper.hpp
> +++ b/utests/utest_helper.hpp
> @@ -310,8 +310,8 @@ extern clGetKernelSubGroupInfoKHR_cb* ut
> /* Check if cl_intel_motion_estimation enabled. */ extern int cl_check_motion_estimation(void);
>
> -/* Check is cl version 2.0. */
> -extern int cl_check_ocl20(void);
> +/* Check is cl version 2.0. (if or_beignet is set, also accept any
> +beignet device, whatever its CL version)*/ extern int
> +cl_check_ocl20(bool or_beignet=true);
>
> /* Check is FP16 enabled. */
> extern int cl_check_half(void);
> --- a/utests/utest_helper.cpp
> +++ b/utests/utest_helper.cpp
> @@ -939,13 +939,14 @@ int cl_check_subgroups_short(void)
> return 1;
> }
>
> -int cl_check_ocl20(void)
> +int cl_check_ocl20(bool or_beignet)
> {
> size_t param_value_size;
> size_t ret_sz;
> OCL_CALL(clGetDeviceInfo, device, CL_DEVICE_OPENCL_C_VERSION, 0, 0, ¶m_value_size);
> if(param_value_size == 0) {
> printf("Not OpenCL 2.0 device, ");
> + if(or_beignet){
> if(cl_check_beignet()) {
> printf("Beignet extension test!");
> return 1;
> @@ -953,6 +954,10 @@ int cl_check_ocl20(void)
> printf("Not beignet device , Skip!");
> return 0;
> }
> + }else{
> + printf("Skip!");
> + return 0;
> + }
> }
> char* device_version_str = (char* )malloc(param_value_size * sizeof(char) );
> OCL_CALL(clGetDeviceInfo, device, CL_DEVICE_OPENCL_C_VERSION, param_value_size, (void*)device_version_str, &ret_sz); @@ -961,6 +966,7 @@ int cl_check_ocl20(void)
> if(!strstr(device_version_str, "2.0")) {
> free(device_version_str);
> printf("Not OpenCL 2.0 device, ");
> + if(or_beignet){
> if(cl_check_beignet()) {
> printf("Beignet extension test!");
> return 1;
> @@ -968,6 +974,10 @@ int cl_check_ocl20(void)
> printf("Not beignet device , Skip!");
> return 0;
> }
> + }else{
> + printf("Skip!");
> + return 0;
> + }
> }
> free(device_version_str);
> return 1;
> --- a/docs/Beignet.mdwn
> +++ b/docs/Beignet.mdwn
> @@ -154,15 +154,9 @@ Supported Targets
>
> OpenCL 2.0
> ----------
> -From release v1.3.0, beignet support OpenCL 2.0. By default, OpenCL 2.0 support is disabled, you can enable it when cmake with option --DENABLE_OPENCL_20=1. Please remember that to enable OpenCL 2.0, there are some dependencies. First, OpenCL 2.0 only support the targets -from Skylake, include Skylake, Kabylake and Apollolake. Then, clang supports all OpenCL 2.0 feature from 3.9. So to enable OpenCL 2.0, -you must update LLVM/clang to 3.9 or later. And also requeires libdrm at least 2.4.66.
> -After enable OpenCL 2.0, beignet complies with OpenCL 2.0 spec, but some OpenCL 2.0 features are simulated by software, there is no performance -gain, such as pipe and device queues, especially device queues.
> -If you build beignet with OpenCL 2.0 enabled and your kernel don't use the OpenCL 2.0 features, please pass a build option -cl-std=CL1.2 for -performance, the OpenCL 2.0 uses more registers and has lots of int64 operations, which may hurt performance, and beignet will continue to improve -OpenCL 2.0 performance.
> +From release v1.3.0, beignet supports OpenCL 2.0 on Skylake and later hardware.
> +This requires LLVM/Clang 3.9 or later and libdrm 2.4.66 or later; if you do not have these or do not want 2.0, you can disable it by passing -DENABLE_OPENCL_20=0 to cmake.
> +As required by the OpenCL specification, kernels are compiled as OpenCL C 1.2 by default; to use 2.0 they must explicitly request it with the -cl-std=CL2.0 build option. As OpenCL 2.0 is likely to be slower than 1.2, we recommend that this is used only where needed. (This is because 2.0 uses more registers and has lots of int64 operations, and some of the 2.0 features (pipes and especially device queues) are implemented in software so do not provide any performance gain.) Beignet will continue to improve OpenCL 2.0 performance.
>
> Known Issues
> ------------
>
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/beignet
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 520 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/beignet/attachments/20170123/94bf6139/attachment.sig>
More information about the Beignet
mailing list