1 /*
2  * Copyright (C) 2018-2021 Intel Corporation
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  */
7 
8 #include "shared/source/compiler_interface/oclc_extensions.h"
9 
10 #include "shared/source/debug_settings/debug_settings_manager.h"
11 #include "shared/source/helpers/hw_info.h"
12 #include "shared/source/helpers/string.h"
13 
14 #include <string>
15 
16 namespace NEO {
17 const char *deviceExtensionsList = "cl_khr_byte_addressable_store "
18                                    "cl_khr_fp16 "
19                                    "cl_khr_global_int32_base_atomics "
20                                    "cl_khr_global_int32_extended_atomics "
21                                    "cl_khr_icd "
22                                    "cl_khr_local_int32_base_atomics "
23                                    "cl_khr_local_int32_extended_atomics "
24                                    "cl_intel_command_queue_families "
25                                    "cl_intel_subgroups "
26                                    "cl_intel_required_subgroup_size "
27                                    "cl_intel_subgroups_short "
28                                    "cl_khr_spir "
29                                    "cl_intel_accelerator "
30                                    "cl_intel_driver_diagnostics "
31                                    "cl_khr_priority_hints "
32                                    "cl_khr_throttle_hints "
33                                    "cl_khr_create_command_queue "
34                                    "cl_intel_subgroups_char "
35                                    "cl_intel_subgroups_long "
36                                    "cl_khr_il_program "
37                                    "cl_intel_mem_force_host_memory "
38                                    "cl_khr_subgroup_extended_types "
39                                    "cl_khr_subgroup_non_uniform_vote "
40                                    "cl_khr_subgroup_ballot "
41                                    "cl_khr_subgroup_non_uniform_arithmetic "
42                                    "cl_khr_subgroup_shuffle "
43                                    "cl_khr_subgroup_shuffle_relative "
44                                    "cl_khr_subgroup_clustered_reduce "
45                                    "cl_intel_device_attribute_query "
46                                    "cl_khr_suggested_local_work_size ";
47 
getExtensionsList(const HardwareInfo & hwInfo)48 std::string getExtensionsList(const HardwareInfo &hwInfo) {
49     std::string allExtensionsList;
50     allExtensionsList.reserve(1000);
51 
52     allExtensionsList.append(deviceExtensionsList);
53 
54     if (hwInfo.capabilityTable.supportsOcl21Features) {
55         allExtensionsList += "cl_khr_subgroups ";
56         if (hwInfo.capabilityTable.supportsVme) {
57             allExtensionsList += "cl_intel_spirv_device_side_avc_motion_estimation ";
58         }
59         if (hwInfo.capabilityTable.supportsMediaBlock) {
60             allExtensionsList += "cl_intel_spirv_media_block_io ";
61         }
62         allExtensionsList += "cl_intel_spirv_subgroups ";
63         allExtensionsList += "cl_khr_spirv_no_integer_wrap_decoration ";
64     }
65 
66     if (hwInfo.capabilityTable.ftrSupportsFP64) {
67         allExtensionsList += "cl_khr_fp64 ";
68     }
69 
70     if (hwInfo.capabilityTable.ftrSupportsInteger64BitAtomics) {
71         allExtensionsList += "cl_khr_int64_base_atomics ";
72         allExtensionsList += "cl_khr_int64_extended_atomics ";
73     }
74 
75     if (hwInfo.capabilityTable.supportsImages) {
76         allExtensionsList += "cl_khr_3d_image_writes ";
77     }
78 
79     if (hwInfo.capabilityTable.supportsVme) {
80         allExtensionsList += "cl_intel_motion_estimation cl_intel_device_side_avc_motion_estimation ";
81     }
82 
83     return allExtensionsList;
84 }
85 
getOpenclCFeaturesList(const HardwareInfo & hwInfo,OpenClCFeaturesContainer & openclCFeatures)86 void getOpenclCFeaturesList(const HardwareInfo &hwInfo, OpenClCFeaturesContainer &openclCFeatures) {
87     cl_name_version openClCFeature;
88     openClCFeature.version = CL_MAKE_VERSION(3, 0, 0);
89 
90     strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_int64");
91     openclCFeatures.push_back(openClCFeature);
92 
93     if (hwInfo.capabilityTable.supportsImages) {
94         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_3d_image_writes");
95         openclCFeatures.push_back(openClCFeature);
96 
97         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_images");
98         openclCFeatures.push_back(openClCFeature);
99 
100         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_read_write_images");
101         openclCFeatures.push_back(openClCFeature);
102     }
103 
104     if (hwInfo.capabilityTable.supportsOcl21Features) {
105         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_atomic_order_acq_rel");
106         openclCFeatures.push_back(openClCFeature);
107 
108         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_atomic_order_seq_cst");
109         openclCFeatures.push_back(openClCFeature);
110 
111         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_atomic_scope_all_devices");
112         openclCFeatures.push_back(openClCFeature);
113 
114         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_atomic_scope_device");
115         openclCFeatures.push_back(openClCFeature);
116 
117         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_generic_address_space");
118         openclCFeatures.push_back(openClCFeature);
119 
120         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_program_scope_global_variables");
121         openclCFeatures.push_back(openClCFeature);
122 
123         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_work_group_collective_functions");
124         openclCFeatures.push_back(openClCFeature);
125 
126         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_subgroups");
127         openclCFeatures.push_back(openClCFeature);
128     }
129 
130     auto forceDeviceEnqueueSupport = DebugManager.flags.ForceDeviceEnqueueSupport.get();
131     if ((hwInfo.capabilityTable.supportsDeviceEnqueue && (forceDeviceEnqueueSupport == -1)) ||
132         (forceDeviceEnqueueSupport == 1)) {
133         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_device_enqueue");
134         openclCFeatures.push_back(openClCFeature);
135     }
136 
137     auto forcePipeSupport = DebugManager.flags.ForcePipeSupport.get();
138     if ((hwInfo.capabilityTable.supportsPipes && (forcePipeSupport == -1)) ||
139         (forcePipeSupport == 1)) {
140         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_pipes");
141         openclCFeatures.push_back(openClCFeature);
142     }
143 
144     auto forceFp64Support = DebugManager.flags.OverrideDefaultFP64Settings.get();
145     if ((hwInfo.capabilityTable.ftrSupportsFP64 && (forceFp64Support == -1)) ||
146         (forceFp64Support == 1)) {
147         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_fp64");
148         openclCFeatures.push_back(openClCFeature);
149     }
150 }
151 
convertEnabledExtensionsToCompilerInternalOptions(const char * enabledExtensions,OpenClCFeaturesContainer & openclCFeatures)152 std::string convertEnabledExtensionsToCompilerInternalOptions(const char *enabledExtensions,
153                                                               OpenClCFeaturesContainer &openclCFeatures) {
154 
155     std::string extensionsList = enabledExtensions;
156     extensionsList.reserve(1500);
157     extensionsList = " -cl-ext=-all,";
158     std::istringstream extensionsStringStream(enabledExtensions);
159     std::string extension;
160     while (extensionsStringStream >> extension) {
161         extensionsList.append("+");
162         extensionsList.append(extension);
163         extensionsList.append(",");
164     }
165     for (auto &feature : openclCFeatures) {
166         extensionsList.append("+");
167         extensionsList.append(feature.name);
168         extensionsList.append(",");
169     }
170     extensionsList[extensionsList.size() - 1] = ' ';
171 
172     return extensionsList;
173 }
174 
getOclVersionCompilerInternalOption(unsigned int oclVersion)175 std::string getOclVersionCompilerInternalOption(unsigned int oclVersion) {
176     switch (oclVersion) {
177     case 30:
178         return "-ocl-version=300 ";
179     case 21:
180         return "-ocl-version=210 ";
181     default:
182         return "-ocl-version=120 ";
183     }
184 }
185 
186 } // namespace NEO
187