1
2 // =================================================================================================
3 // This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
4 // project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
5 // width of 100 characters per line.
6 //
7 // Author(s):
8 // Cedric Nugteren <www.cedricnugteren.nl>
9 //
10 // This file implements the Routine base class (see the header for information about the class).
11 //
12 // =================================================================================================
13
14 #include <string>
15 #include <vector>
16 #include <chrono>
17 #include <cstdlib>
18
19 #include "routine.hpp"
20
21 namespace clblast {
22 // =================================================================================================
23
24 // For each kernel this map contains a list of routines it is used in
25 const std::vector<std::string> Routine::routines_axpy = {"AXPY", "COPY", "SCAL", "SWAP"};
26 const std::vector<std::string> Routine::routines_dot = {"AMAX", "ASUM", "DOT", "DOTC", "DOTU", "MAX", "MIN", "NRM2", "SUM"};
27 const std::vector<std::string> Routine::routines_ger = {"GER", "GERC", "GERU", "HER", "HER2", "HPR", "HPR2", "SPR", "SPR2", "SYR", "SYR2"};
28 const std::vector<std::string> Routine::routines_gemv = {"GBMV", "GEMV", "HBMV", "HEMV", "HPMV", "SBMV", "SPMV", "SYMV", "TMBV", "TPMV", "TRMV", "TRSV"};
29 const std::vector<std::string> Routine::routines_gemm = {"GEMM", "HEMM", "SYMM", "TRMM"};
30 const std::vector<std::string> Routine::routines_gemm_syrk = {"GEMM", "HEMM", "HER2K", "HERK", "SYMM", "SYR2K", "SYRK", "TRMM", "TRSM"};
31 const std::vector<std::string> Routine::routines_trsm = {"TRSM"};
32 const std::unordered_map<std::string, const std::vector<std::string>> Routine::routines_by_kernel = {
33 {"Xaxpy", routines_axpy},
34 {"Xdot", routines_dot},
35 {"Xgemv", routines_gemv},
36 {"XgemvFast", routines_gemv},
37 {"XgemvFastRot", routines_gemv},
38 {"Xtrsv", routines_gemv},
39 {"Xger", routines_ger},
40 {"Copy", routines_gemm_syrk},
41 {"Pad", routines_gemm_syrk},
42 {"Transpose", routines_gemm_syrk},
43 {"Padtranspose", routines_gemm_syrk},
44 {"Xgemm", routines_gemm_syrk},
45 {"XgemmDirect", routines_gemm},
46 {"KernelSelection", routines_gemm},
47 {"Invert", routines_trsm},
48 };
49 // =================================================================================================
50
51 // The constructor does all heavy work, errors are returned as exceptions
Routine(Queue & queue,EventPointer event,const std::string & name,const std::vector<std::string> & kernel_names,const Precision precision,const std::vector<database::DatabaseEntry> & userDatabase,std::initializer_list<const char * > source)52 Routine::Routine(Queue &queue, EventPointer event, const std::string &name,
53 const std::vector<std::string> &kernel_names, const Precision precision,
54 const std::vector<database::DatabaseEntry> &userDatabase,
55 std::initializer_list<const char *> source):
56 precision_(precision),
57 routine_name_(name),
58 kernel_names_(kernel_names),
59 queue_(queue),
60 event_(event),
61 context_(queue_.GetContext()),
62 device_(queue_.GetDevice()),
63 platform_(device_.Platform()),
64 db_(kernel_names) {
65
66 InitDatabase(userDatabase);
67 InitProgram(source);
68 }
69
InitDatabase(const std::vector<database::DatabaseEntry> & userDatabase)70 void Routine::InitDatabase(const std::vector<database::DatabaseEntry> &userDatabase) {
71 for (const auto &kernel_name : kernel_names_) {
72
73 // Queries the cache to see whether or not the kernel parameter database is already there
74 bool has_db;
75 db_(kernel_name) = DatabaseCache::Instance().Get(DatabaseKeyRef{ platform_, device_(), precision_, kernel_name },
76 &has_db);
77 if (has_db) { continue; }
78
79 // Builds the parameter database for this device and routine set and stores it in the cache
80 db_(kernel_name) = Database(device_, kernel_name, precision_, userDatabase);
81 DatabaseCache::Instance().Store(DatabaseKey{ platform_, device_(), precision_, kernel_name },
82 Database{ db_(kernel_name) });
83 }
84 }
85
InitProgram(std::initializer_list<const char * > source)86 void Routine::InitProgram(std::initializer_list<const char *> source) {
87
88 // Queries the cache to see whether or not the program (context-specific) is already there
89 bool has_program;
90 program_ = ProgramCache::Instance().Get(ProgramKeyRef{ context_(), device_(), precision_, routine_name_ },
91 &has_program);
92 if (has_program) { return; }
93
94 // Sets the build options from an environmental variable (if set)
95 auto options = std::vector<std::string>();
96 const auto environment_variable = std::getenv("CLBLAST_BUILD_OPTIONS");
97 if (environment_variable != nullptr) {
98 options.push_back(std::string(environment_variable));
99 }
100
101 // Queries the cache to see whether or not the binary (device-specific) is already there. If it
102 // is, a program is created and stored in the cache
103 const auto device_name = GetDeviceName(device_);
104 bool has_binary;
105 auto binary = BinaryCache::Instance().Get(BinaryKeyRef{ precision_, routine_name_, device_name },
106 &has_binary);
107 if (has_binary) {
108 program_ = Program(device_, context_, binary);
109 program_.Build(device_, options);
110 ProgramCache::Instance().Store(ProgramKey{ context_(), device_(), precision_, routine_name_ },
111 Program{ program_ });
112 return;
113 }
114
115 // Otherwise, the kernel will be compiled and program will be built. Both the binary and the
116 // program will be added to the cache.
117
118 // Inspects whether or not cl_khr_fp64 is supported in case of double precision
119 if ((precision_ == Precision::kDouble && !PrecisionSupported<double>(device_)) ||
120 (precision_ == Precision::kComplexDouble && !PrecisionSupported<double2>(device_))) {
121 throw RuntimeErrorCode(StatusCode::kNoDoublePrecision);
122 }
123
124 // As above, but for cl_khr_fp16 (half precision)
125 if (precision_ == Precision::kHalf && !PrecisionSupported<half>(device_)) {
126 throw RuntimeErrorCode(StatusCode::kNoHalfPrecision);
127 }
128
129 // Collects the parameters for this device in the form of defines, and adds the precision
130 auto source_string = std::string{""};
131 for (const auto &kernel_name : kernel_names_) {
132 source_string += db_(kernel_name).GetDefines();
133 }
134 source_string += "#define PRECISION "+ToString(static_cast<int>(precision_))+"\n";
135
136 // Adds the name of the routine as a define
137 source_string += "#define ROUTINE_"+routine_name_+"\n";
138
139 // Not all OpenCL compilers support the 'inline' keyword. The keyword is only used for devices on
140 // which it is known to work with all OpenCL platforms.
141 if (device_.IsNVIDIA() || device_.IsARM()) {
142 source_string += "#define USE_INLINE_KEYWORD 1\n";
143 }
144
145 // For specific devices, use the non-IEE754 compliant OpenCL mad() instruction. This can improve
146 // performance, but might result in a reduced accuracy.
147 if (device_.IsAMD() && device_.IsGPU()) {
148 source_string += "#define USE_CL_MAD 1\n";
149 }
150
151 // For specific devices, use staggered/shuffled workgroup indices.
152 if (device_.IsAMD() && device_.IsGPU()) {
153 source_string += "#define USE_STAGGERED_INDICES 1\n";
154 }
155
156 // For specific devices add a global synchronisation barrier to the GEMM kernel to optimize
157 // performance through better cache behaviour
158 if (device_.IsARM() && device_.IsGPU()) {
159 source_string += "#define GLOBAL_MEM_FENCE 1\n";
160 }
161
162 // Loads the common header (typedefs and defines and such)
163 source_string +=
164 #include "kernels/common.opencl"
165 ;
166
167 // Adds routine-specific code to the constructed source string
168 for (const char *s: source) {
169 source_string += s;
170 }
171
172 // Prints details of the routine to compile in case of debugging in verbose mode
173 #ifdef VERBOSE
174 printf("[DEBUG] Compiling routine '%s-%s' for device '%s'\n",
175 routine_name_.c_str(), ToString(precision_).c_str(), device_name.c_str());
176 const auto start_time = std::chrono::steady_clock::now();
177 #endif
178
179 // Compiles the kernel
180 program_ = Program(context_, source_string);
181 try {
182 program_.Build(device_, options);
183 } catch (const CLError &e) {
184 if (e.status() == CL_BUILD_PROGRAM_FAILURE) {
185 fprintf(stdout, "OpenCL compiler error/warning: %s\n",
186 program_.GetBuildInfo(device_).c_str());
187 }
188 throw;
189 }
190
191 // Store the compiled binary and program in the cache
192 BinaryCache::Instance().Store(BinaryKey{ precision_, routine_name_, device_name },
193 program_.GetIR());
194
195 ProgramCache::Instance().Store(ProgramKey{ context_(), device_(), precision_, routine_name_ },
196 Program{ program_ });
197
198 // Prints the elapsed compilation time in case of debugging in verbose mode
199 #ifdef VERBOSE
200 const auto elapsed_time = std::chrono::steady_clock::now() - start_time;
201 const auto timing = std::chrono::duration<double,std::milli>(elapsed_time).count();
202 printf("[DEBUG] Completed compilation in %.2lf ms\n", timing);
203 #endif
204 }
205
206 // =================================================================================================
207 } // namespace clblast
208