diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2017-02-26 14:51:45 +0100 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2017-02-26 14:51:45 +0100 |
commit | ea6790665d228e9ff9ba39983a60cd91611ee1fe (patch) | |
tree | 043ca277a867507f97f804cc4057fe50e548b9b1 /src/database | |
parent | a145890aaac0087d36b414bd59c247ae4b70b3e5 (diff) | |
parent | 0643a29af51f9eb13e2b276d0a0e74590c699d3b (diff) |
Merge branch 'development' into triangular_solvers
Diffstat (limited to 'src/database')
-rw-r--r-- | src/database/database.cpp | 32 | ||||
-rw-r--r-- | src/database/database.hpp | 33 | ||||
-rw-r--r-- | src/database/kernels/copy.hpp | 32 | ||||
-rw-r--r-- | src/database/kernels/pad.hpp | 18 | ||||
-rw-r--r-- | src/database/kernels/padtranspose.hpp | 16 | ||||
-rw-r--r-- | src/database/kernels/transpose.hpp | 18 | ||||
-rw-r--r-- | src/database/kernels/xaxpy.hpp | 20 | ||||
-rw-r--r-- | src/database/kernels/xdot.hpp | 18 | ||||
-rw-r--r-- | src/database/kernels/xgemm.hpp | 26 | ||||
-rw-r--r-- | src/database/kernels/xgemm_direct.hpp | 27 | ||||
-rw-r--r-- | src/database/kernels/xgemv.hpp | 15 | ||||
-rw-r--r-- | src/database/kernels/xgemv_fast.hpp | 12 | ||||
-rw-r--r-- | src/database/kernels/xgemv_fast_rot.hpp | 12 | ||||
-rw-r--r-- | src/database/kernels/xger.hpp | 28 |
14 files changed, 252 insertions, 55 deletions
diff --git a/src/database/database.cpp b/src/database/database.cpp index aff6490d..f1d1dc66 100644 --- a/src/database/database.cpp +++ b/src/database/database.cpp @@ -67,7 +67,7 @@ const std::unordered_map<std::string, std::string> Database::kVendorNames{ // Constructor, computing device properties and populating the parameter-vector from the database. // This takes an optional overlay database in case of custom tuning or custom kernels. -Database::Database(const Device &device, const std::vector<std::string> &kernels, +Database::Database(const Device &device, const std::string &kernel_name, const Precision precision, const std::vector<const DatabaseEntry*> &overlay): parameters_(std::make_shared<Parameters>()) { @@ -83,20 +83,17 @@ Database::Database(const Device &device, const std::vector<std::string> &kernels } } - // Iterates over all kernels to include, and retrieves the parameters for each of them - for (auto &kernel: kernels) { - auto search_result = ParametersPtr{}; - - for (auto &db: { database, overlay}) { - search_result = Search(kernel, device_type, device_vendor, device_name, precision, db); - if (search_result) { - parameters_->insert(search_result->begin(), search_result->end()); - break; - } + // Searches potentially multiple databases + auto search_result = ParametersPtr{}; + for (auto &db: { overlay, database}) { + search_result = Search(kernel_name, device_type, device_vendor, device_name, precision, db); + if (search_result) { + parameters_->insert(search_result->begin(), search_result->end()); + break; } - - if (!search_result) { throw RuntimeErrorCode(StatusCode::kDatabaseError); } } + + if (!search_result) { throw RuntimeErrorCode(StatusCode::kDatabaseError); } } // ================================================================================================= @@ -110,6 +107,15 @@ std::string Database::GetDefines() const { return defines; } +// Retrieves the names of all the parameters +std::vector<std::string> Database::GetParameterNames() const { + auto parameter_names = std::vector<std::string>(); + for (auto ¶meter: *parameters_) { + parameter_names.push_back(parameter.first); + } + return parameter_names; +} + // ================================================================================================= // Searches a particular database for the right kernel and precision diff --git a/src/database/database.hpp b/src/database/database.hpp index 87c12293..b34e0d8a 100644 --- a/src/database/database.hpp +++ b/src/database/database.hpp @@ -75,15 +75,19 @@ class Database { Database() = default; // The constructor with a user-provided database overlay (potentially an empty vector) - explicit Database(const Device &device, const std::vector<std::string> &routines, + explicit Database(const Device &device, const std::string &kernel_name, const Precision precision, const std::vector<const DatabaseEntry*> &overlay); // Accessor of values by key - size_t operator[](const std::string key) const { return parameters_->find(key)->second; } + size_t operator[](const std::string &key) const { return parameters_->find(key)->second; } + bool exists(const std::string &key) const { return (parameters_->count(key) == 1); } // Obtain a list of OpenCL pre-processor defines based on the parameters std::string GetDefines() const; + // Retrieves the names of all the parameters + std::vector<std::string> GetParameterNames() const; + private: // Search method for a specified database, returning pointer (possibly a nullptr) ParametersPtr Search(const std::string &this_kernel, const std::string &this_type, @@ -96,6 +100,31 @@ class Database { }; // ================================================================================================= + +// Multiple databases together in a map +class Databases { + public: + + explicit Databases(const std::vector<std::string> &kernel_names): kernel_names_(kernel_names) { } + + // Database accessor + Database& operator()(const std::string &kernel_name) { return databases_[kernel_name]; } + + // Retrieves a parameter from the database + size_t operator[](const std::string &key) const { + for (const auto &kernel_name : kernel_names_) { + const auto &kernel_db = databases_.find(kernel_name)->second; + if (kernel_db.exists(key)) { return kernel_db[key]; } + } + throw RuntimeErrorCode(StatusCode::kDatabaseError); + } + + private: + const std::vector<std::string> kernel_names_; + std::unordered_map<std::string, Database> databases_; +}; + +// ================================================================================================= } // namespace clblast // CLBLAST_DATABASE_H_ diff --git a/src/database/kernels/copy.hpp b/src/database/kernels/copy.hpp index f0431933..b2aa736f 100644 --- a/src/database/kernels/copy.hpp +++ b/src/database/kernels/copy.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry CopyHalf = { "Copy", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",8}, {"COPY_WPT",4} } }, @@ -26,7 +32,7 @@ const Database::DatabaseEntry CopyHalf = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",4}, {"COPY_WPT",8} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, } }, } @@ -39,13 +45,14 @@ const Database::DatabaseEntry CopySingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, + { "Ellesmere", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",8} } }, { "Hawaii", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "Oland", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",2} } }, { "Pitcairn", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "Tahiti", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "Tonga", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, { "Turks", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",2} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, } }, { // ARM GPUs @@ -96,6 +103,7 @@ const Database::DatabaseEntry CopySingle = { { "GeForce GTX TITAN", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",4} } }, { "GeForce GTX TITAN Black", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",4}, {"COPY_WPT",8} } }, { "GeForce GTX TITAN X", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, + { "TITAN X (Pascal)", { {"COPY_DIMX",32}, {"COPY_DIMY",32}, {"COPY_VW",4}, {"COPY_WPT",1} } }, { "Tesla K20m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",2} } }, { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",4}, {"COPY_WPT",1} } }, @@ -103,7 +111,7 @@ const Database::DatabaseEntry CopySingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",4} } }, } }, } @@ -116,13 +124,14 @@ const Database::DatabaseEntry CopyComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "Ellesmere", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",4} } }, { "Hawaii", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Oland", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "Pitcairn", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Tahiti", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "Tonga", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Turks", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",4}, {"COPY_WPT",1} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",4} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, } }, { // Intel CPUs @@ -165,14 +174,15 @@ const Database::DatabaseEntry CopyComplexSingle = { { "GeForce GTX 980", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX TITAN Black", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX TITAN X", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "TITAN X (Pascal)", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "Tesla K20m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",4} } }, { "Tesla K40m", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, - { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, } }, } @@ -185,12 +195,13 @@ const Database::DatabaseEntry CopyDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "Ellesmere", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",4} } }, { "Hawaii", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Oland", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",8} } }, { "Pitcairn", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "Tahiti", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "Tonga", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",2}, {"COPY_WPT",4} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",4} } }, } }, { // ARM GPUs @@ -229,14 +240,15 @@ const Database::DatabaseEntry CopyDouble = { { "GeForce GTX TITAN", { {"COPY_DIMX",16}, {"COPY_DIMY",32}, {"COPY_VW",2}, {"COPY_WPT",2} } }, { "GeForce GTX TITAN Black", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",8} } }, { "GeForce GTX TITAN X", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "TITAN X (Pascal)", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "Tesla K20m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",1} } }, { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",1} } }, + { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",16}, {"COPY_VW",2}, {"COPY_WPT",1} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",2} } }, + { "default", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",1} } }, } }, } @@ -249,6 +261,7 @@ const Database::DatabaseEntry CopyComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "Ellesmere", { {"COPY_DIMX",8}, {"COPY_DIMY",32}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Hawaii", { {"COPY_DIMX",32}, {"COPY_DIMY",8}, {"COPY_VW",2}, {"COPY_WPT",8} } }, { "Oland", { {"COPY_DIMX",8}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "Pitcairn", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, @@ -293,6 +306,7 @@ const Database::DatabaseEntry CopyComplexDouble = { { "GeForce GTX TITAN", { {"COPY_DIMX",16}, {"COPY_DIMY",16}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "GeForce GTX TITAN Black", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "GeForce GTX TITAN X", { {"COPY_DIMX",16}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, + { "TITAN X (Pascal)", { {"COPY_DIMX",32}, {"COPY_DIMY",32}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "Tesla K20m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",2} } }, { "Tesla K40m", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, { "default", { {"COPY_DIMX",8}, {"COPY_DIMY",8}, {"COPY_VW",1}, {"COPY_WPT",1} } }, diff --git a/src/database/kernels/pad.hpp b/src/database/kernels/pad.hpp index 3378709c..e4005527 100644 --- a/src/database/kernels/pad.hpp +++ b/src/database/kernels/pad.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry PadHalf = { "Pad", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, + { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, @@ -39,6 +45,7 @@ const Database::DatabaseEntry PadSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "Ellesmere", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, { "Hawaii", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",4} } }, { "Oland", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Pitcairn", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, @@ -96,6 +103,7 @@ const Database::DatabaseEntry PadSingle = { { "GeForce GTX TITAN", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "GeForce GTX TITAN Black", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "GeForce GTX TITAN X", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "TITAN X (Pascal)", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tesla K20m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Tesla K40m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",2} } }, @@ -116,6 +124,7 @@ const Database::DatabaseEntry PadComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "Ellesmere", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",4} } }, { "Hawaii", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Oland", { {"PAD_DIMX",8}, {"PAD_DIMY",32}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Pitcairn", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, @@ -173,6 +182,7 @@ const Database::DatabaseEntry PadComplexSingle = { { "GeForce GTX TITAN", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "GeForce GTX TITAN Black", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "GeForce GTX TITAN X", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "TITAN X (Pascal)", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tesla K20m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Tesla K40m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, @@ -180,7 +190,7 @@ const Database::DatabaseEntry PadComplexSingle = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, } }, } @@ -193,12 +203,13 @@ const Database::DatabaseEntry PadDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "Ellesmere", { {"PAD_DIMX",8}, {"PAD_DIMY",32}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Hawaii", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Oland", { {"PAD_DIMX",8}, {"PAD_DIMY",32}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Pitcairn", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Tahiti", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tonga", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",4}, {"PAD_WPTY",1} } }, - { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, } }, { // ARM GPUs @@ -237,6 +248,7 @@ const Database::DatabaseEntry PadDouble = { { "GeForce GTX TITAN", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX TITAN Black", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "GeForce GTX TITAN X", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "TITAN X (Pascal)", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tesla K20m", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Tesla K40m", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "default", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, @@ -257,6 +269,7 @@ const Database::DatabaseEntry PadComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "Ellesmere", { {"PAD_DIMX",8}, {"PAD_DIMY",16}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Hawaii", { {"PAD_DIMX",32}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "Oland", { {"PAD_DIMX",8}, {"PAD_DIMY",16}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Pitcairn", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, @@ -301,6 +314,7 @@ const Database::DatabaseEntry PadComplexDouble = { { "GeForce GTX TITAN", { {"PAD_DIMX",8}, {"PAD_DIMY",32}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "GeForce GTX TITAN Black", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",4} } }, { "GeForce GTX TITAN X", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, + { "TITAN X (Pascal)", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",2}, {"PAD_WPTY",1} } }, { "Tesla K20m", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",2} } }, { "Tesla K40m", { {"PAD_DIMX",8}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, { "default", { {"PAD_DIMX",16}, {"PAD_DIMY",8}, {"PAD_WPTX",1}, {"PAD_WPTY",1} } }, diff --git a/src/database/kernels/padtranspose.hpp b/src/database/kernels/padtranspose.hpp index 212723c7..92aa4f7b 100644 --- a/src/database/kernels/padtranspose.hpp +++ b/src/database/kernels/padtranspose.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry PadtransposeHalf = { "Padtranspose", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, + { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",1} } }, @@ -39,6 +45,7 @@ const Database::DatabaseEntry PadtransposeSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, + { "Ellesmere", { {"PADTRA_PAD",1}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Hawaii", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "Oland", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "Pitcairn", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, @@ -96,6 +103,7 @@ const Database::DatabaseEntry PadtransposeSingle = { { "GeForce GTX TITAN", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "GeForce GTX TITAN Black", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } }, { "GeForce GTX TITAN X", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "Tesla K20m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } }, { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",2} } }, @@ -116,6 +124,7 @@ const Database::DatabaseEntry PadtransposeComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, + { "Ellesmere", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Hawaii", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "Oland", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Pitcairn", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, @@ -173,6 +182,7 @@ const Database::DatabaseEntry PadtransposeComplexSingle = { { "GeForce GTX TITAN", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX TITAN Black", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX TITAN X", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "Tesla K20m", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, @@ -193,6 +203,7 @@ const Database::DatabaseEntry PadtransposeDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, + { "Ellesmere", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "Hawaii", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, { "Oland", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",4} } }, { "Pitcairn", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, @@ -237,6 +248,7 @@ const Database::DatabaseEntry PadtransposeDouble = { { "GeForce GTX TITAN", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX TITAN Black", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX TITAN X", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",2} } }, { "Tesla K20m", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, @@ -244,7 +256,7 @@ const Database::DatabaseEntry PadtransposeDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",2} } }, + { "default", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, } }, } @@ -257,6 +269,7 @@ const Database::DatabaseEntry PadtransposeComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, + { "Ellesmere", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Hawaii", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Oland", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, { "Pitcairn", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",4} } }, @@ -301,6 +314,7 @@ const Database::DatabaseEntry PadtransposeComplexDouble = { { "GeForce GTX TITAN", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX TITAN Black", { {"PADTRA_PAD",0}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "GeForce GTX TITAN X", { {"PADTRA_PAD",1}, {"PADTRA_TILE",32}, {"PADTRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"PADTRA_PAD",0}, {"PADTRA_TILE",8}, {"PADTRA_WPT",1} } }, { "Tesla K20m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "Tesla K40m", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, { "default", { {"PADTRA_PAD",1}, {"PADTRA_TILE",16}, {"PADTRA_WPT",1} } }, diff --git a/src/database/kernels/transpose.hpp b/src/database/kernels/transpose.hpp index f33f2a04..0ee2d83b 100644 --- a/src/database/kernels/transpose.hpp +++ b/src/database/kernels/transpose.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry TransposeHalf = { "Transpose", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, + { "default", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, @@ -26,7 +32,7 @@ const Database::DatabaseEntry TransposeHalf = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",8} } }, + { "default", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, } }, } @@ -39,13 +45,14 @@ const Database::DatabaseEntry TransposeSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, + { "Ellesmere", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Hawaii", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",8} } }, { "Oland", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Pitcairn", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Tahiti", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Tonga", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "Turks", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, - { "default", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, + { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, } }, { // ARM GPUs @@ -96,6 +103,7 @@ const Database::DatabaseEntry TransposeSingle = { { "GeForce GTX TITAN", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "GeForce GTX TITAN Black", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "GeForce GTX TITAN X", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, + { "TITAN X (Pascal)", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Tesla K20m", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "Tesla K40m", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, { "default", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",4} } }, @@ -116,6 +124,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, + { "Ellesmere", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Hawaii", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Oland", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "Pitcairn", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, @@ -167,6 +176,7 @@ const Database::DatabaseEntry TransposeComplexSingle = { { "GeForce GTX TITAN", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX TITAN Black", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX TITAN X", { {"TRA_DIM",32}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, @@ -187,6 +197,7 @@ const Database::DatabaseEntry TransposeDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, + { "Ellesmere", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",4} } }, { "Hawaii", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Oland", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "Pitcairn", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, @@ -231,6 +242,7 @@ const Database::DatabaseEntry TransposeDouble = { { "GeForce GTX TITAN", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "GeForce GTX TITAN Black", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "GeForce GTX TITAN X", { {"TRA_DIM",32}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"TRA_DIM",8}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",2} } }, { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, @@ -251,6 +263,7 @@ const Database::DatabaseEntry TransposeComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"TRA_DIM",4}, {"TRA_PAD",1}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, + { "Ellesmere", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Hawaii", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",2} } }, { "Oland", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, { "Pitcairn", { {"TRA_DIM",4}, {"TRA_PAD",0}, {"TRA_SHUFFLE",1}, {"TRA_WPT",1} } }, @@ -289,6 +302,7 @@ const Database::DatabaseEntry TransposeComplexDouble = { { "GeForce GTX TITAN", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX TITAN Black", { {"TRA_DIM",16}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "GeForce GTX TITAN X", { {"TRA_DIM",32}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, + { "TITAN X (Pascal)", { {"TRA_DIM",8}, {"TRA_PAD",0}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "Tesla K20m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "Tesla K40m", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, { "default", { {"TRA_DIM",16}, {"TRA_PAD",1}, {"TRA_SHUFFLE",0}, {"TRA_WPT",1} } }, diff --git a/src/database/kernels/xaxpy.hpp b/src/database/kernels/xaxpy.hpp index e4e3c621..4b747f0a 100644 --- a/src/database/kernels/xaxpy.hpp +++ b/src/database/kernels/xaxpy.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XaxpyHalf = { "Xaxpy", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"VW",4}, {"WGS",128}, {"WPT",4} } }, + { "default", { {"VW",4}, {"WGS",128}, {"WPT",4} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, @@ -26,7 +32,7 @@ const Database::DatabaseEntry XaxpyHalf = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"VW",8}, {"WGS",64}, {"WPT",1} } }, + { "default", { {"VW",8}, {"WGS",256}, {"WPT",4} } }, } }, } @@ -39,13 +45,14 @@ const Database::DatabaseEntry XaxpySingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, + { "Ellesmere", { {"VW",1}, {"WGS",64}, {"WPT",4} } }, { "Hawaii", { {"VW",2}, {"WGS",64}, {"WPT",2} } }, { "Oland", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "Pitcairn", { {"VW",2}, {"WGS",128}, {"WPT",1} } }, { "Tahiti", { {"VW",2}, {"WGS",64}, {"WPT",1} } }, { "Tonga", { {"VW",1}, {"WGS",256}, {"WPT",8} } }, { "Turks", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, - { "default", { {"VW",2}, {"WGS",64}, {"WPT",2} } }, + { "default", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, } }, { // ARM GPUs @@ -96,6 +103,7 @@ const Database::DatabaseEntry XaxpySingle = { { "GeForce GTX TITAN", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX TITAN Black", { {"VW",4}, {"WGS",128}, {"WPT",4} } }, { "GeForce GTX TITAN X", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, + { "TITAN X (Pascal)", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "Tesla K20m", { {"VW",4}, {"WGS",128}, {"WPT",1} } }, { "Tesla K40m", { {"VW",4}, {"WGS",128}, {"WPT",1} } }, { "default", { {"VW",4}, {"WGS",256}, {"WPT",1} } }, @@ -116,6 +124,7 @@ const Database::DatabaseEntry XaxpyComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW",2}, {"WGS",64}, {"WPT",8} } }, + { "Ellesmere", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, { "Hawaii", { {"VW",1}, {"WGS",128}, {"WPT",2} } }, { "Oland", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "Pitcairn", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, @@ -173,6 +182,7 @@ const Database::DatabaseEntry XaxpyComplexSingle = { { "GeForce GTX TITAN", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "GeForce GTX TITAN Black", { {"VW",1}, {"WGS",128}, {"WPT",2} } }, { "GeForce GTX TITAN X", { {"VW",1}, {"WGS",512}, {"WPT",1} } }, + { "TITAN X (Pascal)", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "Tesla K20m", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "Tesla K40m", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "default", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, @@ -193,6 +203,7 @@ const Database::DatabaseEntry XaxpyDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, + { "Ellesmere", { {"VW",2}, {"WGS",64}, {"WPT",4} } }, { "Hawaii", { {"VW",1}, {"WGS",64}, {"WPT",2} } }, { "Oland", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "Pitcairn", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, @@ -237,6 +248,7 @@ const Database::DatabaseEntry XaxpyDouble = { { "GeForce GTX TITAN", { {"VW",2}, {"WGS",1024}, {"WPT",1} } }, { "GeForce GTX TITAN Black", { {"VW",2}, {"WGS",128}, {"WPT",1} } }, { "GeForce GTX TITAN X", { {"VW",1}, {"WGS",512}, {"WPT",1} } }, + { "TITAN X (Pascal)", { {"VW",2}, {"WGS",256}, {"WPT",4} } }, { "Tesla K20m", { {"VW",2}, {"WGS",128}, {"WPT",1} } }, { "Tesla K40m", { {"VW",2}, {"WGS",128}, {"WPT",1} } }, { "default", { {"VW",2}, {"WGS",1024}, {"WPT",1} } }, @@ -244,7 +256,7 @@ const Database::DatabaseEntry XaxpyDouble = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"VW",1}, {"WGS",128}, {"WPT",2} } }, + { "default", { {"VW",2}, {"WGS",256}, {"WPT",1} } }, } }, } @@ -257,6 +269,7 @@ const Database::DatabaseEntry XaxpyComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, + { "Ellesmere", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "Hawaii", { {"VW",2}, {"WGS",64}, {"WPT",1} } }, { "Oland", { {"VW",1}, {"WGS",256}, {"WPT",1} } }, { "Pitcairn", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, @@ -301,6 +314,7 @@ const Database::DatabaseEntry XaxpyComplexDouble = { { "GeForce GTX TITAN", { {"VW",1}, {"WGS",64}, {"WPT",4} } }, { "GeForce GTX TITAN Black", { {"VW",1}, {"WGS",128}, {"WPT",4} } }, { "GeForce GTX TITAN X", { {"VW",1}, {"WGS",1024}, {"WPT",1} } }, + { "TITAN X (Pascal)", { {"VW",1}, {"WGS",128}, {"WPT",1} } }, { "Tesla K20m", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "Tesla K40m", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, { "default", { {"VW",1}, {"WGS",64}, {"WPT",1} } }, diff --git a/src/database/kernels/xdot.hpp b/src/database/kernels/xdot.hpp index 30d98e5d..ea154b6e 100644 --- a/src/database/kernels/xdot.hpp +++ b/src/database/kernels/xdot.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XdotHalf = { "Xdot", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"WGS1",256}, {"WGS2",64} } }, + { "default", { {"WGS1",256}, {"WGS2",64} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"WGS1",256}, {"WGS2",32} } }, @@ -39,12 +45,13 @@ const Database::DatabaseEntry XdotSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",128}, {"WGS2",32} } }, + { "Ellesmere", { {"WGS1",128}, {"WGS2",32} } }, { "Oland", { {"WGS1",256}, {"WGS2",32} } }, { "Pitcairn", { {"WGS1",128}, {"WGS2",32} } }, { "Tahiti", { {"WGS1",128}, {"WGS2",32} } }, { "Tonga", { {"WGS1",64}, {"WGS2",32} } }, { "Turks", { {"WGS1",128}, {"WGS2",64} } }, - { "default", { {"WGS1",128}, {"WGS2",64} } }, + { "default", { {"WGS1",128}, {"WGS2",32} } }, } }, { // Intel CPUs @@ -79,8 +86,9 @@ const Database::DatabaseEntry XdotSingle = { { "GeForce GTX 980", { {"WGS1",256}, {"WGS2",32} } }, { "GeForce GTX TITAN Black", { {"WGS1",512}, {"WGS2",64} } }, { "GeForce GTX TITAN X", { {"WGS1",256}, {"WGS2",32} } }, + { "TITAN X (Pascal)", { {"WGS1",256}, {"WGS2",512} } }, { "Tesla K20m", { {"WGS1",1024}, {"WGS2",32} } }, - { "default", { {"WGS1",256}, {"WGS2",64} } }, + { "default", { {"WGS1",256}, {"WGS2",512} } }, } }, { // Default @@ -98,6 +106,7 @@ const Database::DatabaseEntry XdotComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WGS2",32} } }, + { "Ellesmere", { {"WGS1",256}, {"WGS2",32} } }, { "Oland", { {"WGS1",128}, {"WGS2",32} } }, { "Pitcairn", { {"WGS1",256}, {"WGS2",32} } }, { "Tahiti", { {"WGS1",64}, {"WGS2",32} } }, @@ -138,6 +147,7 @@ const Database::DatabaseEntry XdotComplexSingle = { { "GeForce GTX 980", { {"WGS1",256}, {"WGS2",64} } }, { "GeForce GTX TITAN Black", { {"WGS1",128}, {"WGS2",64} } }, { "GeForce GTX TITAN X", { {"WGS1",256}, {"WGS2",32} } }, + { "TITAN X (Pascal)", { {"WGS1",256}, {"WGS2",32} } }, { "Tesla K20m", { {"WGS1",512}, {"WGS2",32} } }, { "default", { {"WGS1",512}, {"WGS2",64} } }, } @@ -157,6 +167,7 @@ const Database::DatabaseEntry XdotDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WGS2",128} } }, + { "Ellesmere", { {"WGS1",128}, {"WGS2",64} } }, { "Oland", { {"WGS1",256}, {"WGS2",32} } }, { "Pitcairn", { {"WGS1",128}, {"WGS2",32} } }, { "Tahiti", { {"WGS1",256}, {"WGS2",32} } }, @@ -185,6 +196,7 @@ const Database::DatabaseEntry XdotDouble = { { "GeForce GTX 980", { {"WGS1",128}, {"WGS2",32} } }, { "GeForce GTX TITAN Black", { {"WGS1",128}, {"WGS2",64} } }, { "GeForce GTX TITAN X", { {"WGS1",256}, {"WGS2",32} } }, + { "TITAN X (Pascal)", { {"WGS1",256}, {"WGS2",32} } }, { "Tesla K20m", { {"WGS1",512}, {"WGS2",32} } }, { "default", { {"WGS1",128}, {"WGS2",128} } }, } @@ -204,6 +216,7 @@ const Database::DatabaseEntry XdotComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WGS2",32} } }, + { "Ellesmere", { {"WGS1",256}, {"WGS2",32} } }, { "Oland", { {"WGS1",256}, {"WGS2",32} } }, { "Pitcairn", { {"WGS1",256}, {"WGS2",32} } }, { "Tahiti", { {"WGS1",256}, {"WGS2",32} } }, @@ -232,6 +245,7 @@ const Database::DatabaseEntry XdotComplexDouble = { { "GeForce GTX 980", { {"WGS1",64}, {"WGS2",32} } }, { "GeForce GTX TITAN Black", { {"WGS1",128}, {"WGS2",32} } }, { "GeForce GTX TITAN X", { {"WGS1",128}, {"WGS2",32} } }, + { "TITAN X (Pascal)", { {"WGS1",128}, {"WGS2",512} } }, { "Tesla K20m", { {"WGS1",128}, {"WGS2",32} } }, { "default", { {"WGS1",128}, {"WGS2",64} } }, } diff --git a/src/database/kernels/xgemm.hpp b/src/database/kernels/xgemm.hpp index d9414f8b..751f403b 100644 --- a/src/database/kernels/xgemm.hpp +++ b/src/database/kernels/xgemm.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XgemmHalf = { "Xgemm", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics Skylake ULT GT2", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, @@ -38,6 +44,7 @@ const Database::DatabaseEntry XgemmSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",8} } }, + { "Ellesmere", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "Hawaii", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",32}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",2} } }, { "Oland", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "Pitcairn", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, @@ -95,9 +102,10 @@ const Database::DatabaseEntry XgemmSingle = { { "GeForce GTX TITAN", { {"KWG",16}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, { "GeForce GTX TITAN Black", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",4}, {"VWN",1} } }, { "GeForce GTX TITAN X", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",8} } }, + { "TITAN X (Pascal)", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",4}, {"VWN",2} } }, { "Tesla K20m", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, { "Tesla K40m", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, } }, { // Default @@ -115,6 +123,7 @@ const Database::DatabaseEntry XgemmComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",8} } }, + { "Ellesmere", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Hawaii", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",32}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Oland", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",1}, {"VWM",2}, {"VWN",4} } }, { "Pitcairn", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",4}, {"VWN",2} } }, @@ -172,14 +181,15 @@ const Database::DatabaseEntry XgemmComplexSingle = { { "GeForce GTX TITAN", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX TITAN Black", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "GeForce GTX TITAN X", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",4} } }, + { "TITAN X (Pascal)", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",1} } }, { "Tesla K20m", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Tesla K40m", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",32}, {"NDIMC",32}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, } }, } @@ -192,6 +202,7 @@ const Database::DatabaseEntry XgemmDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",4}, {"VWN",4} } }, + { "Ellesmere", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, { "Hawaii", { {"KWG",16}, {"KWI",8}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",128}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Oland", { {"KWG",16}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",1} } }, { "Pitcairn", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, @@ -236,6 +247,7 @@ const Database::DatabaseEntry XgemmDouble = { { "GeForce GTX TITAN", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",2} } }, { "GeForce GTX TITAN Black", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",16}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX TITAN X", { {"KWG",16}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, + { "TITAN X (Pascal)", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",8}, {"NDIMC",32}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",4}, {"VWN",1} } }, { "Tesla K20m", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tesla K40m", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",64}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",1}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",4} } }, { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",4} } }, @@ -256,12 +268,13 @@ const Database::DatabaseEntry XgemmComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWG",32}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",2} } }, + { "Ellesmere", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",16}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Hawaii", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",32}, {"SA",1}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, { "Oland", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",32}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",4} } }, { "Pitcairn", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",32}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tahiti", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tonga", { {"KWG",16}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",16}, {"SA",1}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",1} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",16}, {"SA",1}, {"SB",1}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, } }, { // ARM GPUs @@ -299,14 +312,15 @@ const Database::DatabaseEntry XgemmComplexDouble = { { "GeForce GTX 980", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",16}, {"NWG",128}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",1}, {"VWM",2}, {"VWN",2} } }, { "GeForce GTX TITAN Black", { {"KWG",16}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",1}, {"STRM",1}, {"STRN",1}, {"VWM",1}, {"VWN",1} } }, { "GeForce GTX TITAN X", { {"KWG",32}, {"KWI",8}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",128}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, + { "TITAN X (Pascal)", { {"KWG",32}, {"KWI",2}, {"MDIMA",16}, {"MDIMC",16}, {"MWG",16}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",16}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tesla K20m", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",16}, {"NDIMC",16}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, { "Tesla K40m", { {"KWG",16}, {"KWI",8}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",32}, {"NDIMC",16}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",1}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",32}, {"MDIMC",32}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",1} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",32}, {"NDIMC",32}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",1} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",32}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",64}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",2}, {"VWN",2} } }, + { "default", { {"KWG",32}, {"KWI",2}, {"MDIMA",8}, {"MDIMC",8}, {"MWG",16}, {"NDIMB",8}, {"NDIMC",8}, {"NWG",32}, {"SA",0}, {"SB",0}, {"STRM",0}, {"STRN",0}, {"VWM",1}, {"VWN",2} } }, } }, } diff --git a/src/database/kernels/xgemm_direct.hpp b/src/database/kernels/xgemm_direct.hpp index c0cd2c04..4984475f 100644 --- a/src/database/kernels/xgemm_direct.hpp +++ b/src/database/kernels/xgemm_direct.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XgemmDirectHalf = { "XgemmDirect", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"KWID",8}, {"MDIMAD",32}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + { "default", { {"KWID",8}, {"MDIMAD",32}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics Skylake ULT GT2", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, @@ -25,7 +31,7 @@ const Database::DatabaseEntry XgemmDirectHalf = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, + { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, } @@ -38,9 +44,10 @@ const Database::DatabaseEntry XgemmDirectSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, + { "Ellesmere", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",32}, {"NDIMCD",32}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",32} } }, { "Tonga", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",32}, {"NDIMCD",8}, {"PADA",0}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "Turks", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, } }, { // Intel CPUs @@ -62,12 +69,13 @@ const Database::DatabaseEntry XgemmDirectSingle = { { "GeForce GTX 1080", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, { "GeForce GTX TITAN Black", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, + { "TITAN X (Pascal)", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",2}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",4}, {"VWND",4}, {"WGD",32} } }, } }, } @@ -104,12 +112,13 @@ const Database::DatabaseEntry XgemmDirectComplexSingle = { { "GeForce GTX 1080", { {"KWID",8}, {"MDIMAD",8}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, { "GeForce GTX 750 Ti", { {"KWID",16}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",16} } }, { "GeForce GTX TITAN Black", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "TITAN X (Pascal)", { {"KWID",16}, {"MDIMAD",16}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, } }, } @@ -122,6 +131,7 @@ const Database::DatabaseEntry XgemmDirectDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "Ellesmere", { {"KWID",8}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",32} } }, { "Tonga", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } @@ -138,12 +148,13 @@ const Database::DatabaseEntry XgemmDirectDouble = { { "GeForce GTX 1080", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",4}, {"WGD",32} } }, { "GeForce GTX TITAN Black", { {"KWID",8}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "TITAN X (Pascal)", { {"KWID",8}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",1}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",32} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",2}, {"VWND",2}, {"WGD",16} } }, } }, } @@ -156,6 +167,7 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "Ellesmere", { {"KWID",16}, {"MDIMAD",32}, {"MDIMCD",32}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",0}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "Tonga", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",16}, {"NDIMCD",16}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } @@ -172,12 +184,13 @@ const Database::DatabaseEntry XgemmDirectComplexDouble = { { "GeForce GTX 1080", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "GeForce GTX 750 Ti", { {"KWID",2}, {"MDIMAD",32}, {"MDIMCD",32}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",32} } }, { "GeForce GTX TITAN Black", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",8} } }, + { "TITAN X (Pascal)", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",16}, {"NDIMCD",8}, {"PADA",1}, {"PADB",0}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, { "default", { {"KWID",2}, {"MDIMAD",16}, {"MDIMCD",16}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",1}, {"WGD",16} } }, + { "default", { {"KWID",2}, {"MDIMAD",8}, {"MDIMCD",8}, {"NDIMBD",8}, {"NDIMCD",8}, {"PADA",1}, {"PADB",1}, {"VWMD",1}, {"VWND",2}, {"WGD",16} } }, } }, } diff --git a/src/database/kernels/xgemv.hpp b/src/database/kernels/xgemv.hpp index 52b17d94..05f47554 100644 --- a/src/database/kernels/xgemv.hpp +++ b/src/database/kernels/xgemv.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XgemvHalf = { "Xgemv", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"WGS1",256}, {"WPT1",1} } }, + { "default", { {"WGS1",256}, {"WPT1",1} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"WGS1",64}, {"WPT1",1} } }, @@ -39,6 +45,7 @@ const Database::DatabaseEntry XgemvSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",128}, {"WPT1",1} } }, + { "Ellesmere", { {"WGS1",256}, {"WPT1",1} } }, { "Hawaii", { {"WGS1",128}, {"WPT1",1} } }, { "Oland", { {"WGS1",128}, {"WPT1",1} } }, { "Pitcairn", { {"WGS1",256}, {"WPT1",1} } }, @@ -89,9 +96,10 @@ const Database::DatabaseEntry XgemvSingle = { { "GeForce GTX TITAN", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX TITAN Black", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX TITAN X", { {"WGS1",256}, {"WPT1",1} } }, + { "TITAN X (Pascal)", { {"WGS1",32}, {"WPT1",1} } }, { "Tesla K20m", { {"WGS1",128}, {"WPT1",1} } }, { "Tesla K40m", { {"WGS1",256}, {"WPT1",1} } }, - { "default", { {"WGS1",256}, {"WPT1",1} } }, + { "default", { {"WGS1",128}, {"WPT1",1} } }, } }, { // Default @@ -109,6 +117,7 @@ const Database::DatabaseEntry XgemvComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WPT1",1} } }, + { "Ellesmere", { {"WGS1",32}, {"WPT1",1} } }, { "Hawaii", { {"WGS1",64}, {"WPT1",1} } }, { "Oland", { {"WGS1",64}, {"WPT1",1} } }, { "Pitcairn", { {"WGS1",64}, {"WPT1",1} } }, @@ -157,6 +166,7 @@ const Database::DatabaseEntry XgemvComplexSingle = { { "GeForce GTX 750 Ti", { {"WGS1",32}, {"WPT1",1} } }, { "GeForce GTX TITAN", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX TITAN Black", { {"WGS1",32}, {"WPT1",1} } }, + { "TITAN X (Pascal)", { {"WGS1",32}, {"WPT1",1} } }, { "default", { {"WGS1",64}, {"WPT1",1} } }, } }, @@ -175,6 +185,7 @@ const Database::DatabaseEntry XgemvDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WPT1",1} } }, + { "Ellesmere", { {"WGS1",32}, {"WPT1",1} } }, { "Hawaii", { {"WGS1",128}, {"WPT1",1} } }, { "Oland", { {"WGS1",256}, {"WPT1",1} } }, { "Pitcairn", { {"WGS1",256}, {"WPT1",1} } }, @@ -212,6 +223,7 @@ const Database::DatabaseEntry XgemvDouble = { { "GeForce GTX TITAN", { {"WGS1",256}, {"WPT1",1} } }, { "GeForce GTX TITAN Black", { {"WGS1",32}, {"WPT1",1} } }, { "GeForce GTX TITAN X", { {"WGS1",64}, {"WPT1",1} } }, + { "TITAN X (Pascal)", { {"WGS1",32}, {"WPT1",1} } }, { "Tesla K20m", { {"WGS1",256}, {"WPT1",1} } }, { "Tesla K40m", { {"WGS1",256}, {"WPT1",1} } }, { "default", { {"WGS1",128}, {"WPT1",1} } }, @@ -232,6 +244,7 @@ const Database::DatabaseEntry XgemvComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WPT1",1} } }, + { "Ellesmere", { {"WGS1",32}, {"WPT1",1} } }, { "Hawaii", { {"WGS1",64}, {"WPT1",1} } }, { "Oland", { {"WGS1",256}, {"WPT1",1} } }, { "Pitcairn", { {"WGS1",256}, {"WPT1",1} } }, diff --git a/src/database/kernels/xgemv_fast.hpp b/src/database/kernels/xgemv_fast.hpp index 2dd400bc..efc64d0c 100644 --- a/src/database/kernels/xgemv_fast.hpp +++ b/src/database/kernels/xgemv_fast.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XgemvFastHalf = { "XgemvFast", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"VW2",1}, {"WGS2",32}, {"WPT2",1} } }, + { "default", { {"VW2",1}, {"WGS2",32}, {"WPT2",1} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"VW2",1}, {"WGS2",16}, {"WPT2",1} } }, @@ -39,6 +45,7 @@ const Database::DatabaseEntry XgemvFastSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, + { "Ellesmere", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Hawaii", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Oland", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Pitcairn", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, @@ -89,6 +96,7 @@ const Database::DatabaseEntry XgemvFastSingle = { { "GeForce GTX TITAN", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "GeForce GTX TITAN Black", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "GeForce GTX TITAN X", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, + { "TITAN X (Pascal)", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "Tesla K20m", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "Tesla K40m", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "default", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, @@ -109,6 +117,7 @@ const Database::DatabaseEntry XgemvFastComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW2",2}, {"WGS2",256}, {"WPT2",2} } }, + { "Ellesmere", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Hawaii", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Oland", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Pitcairn", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, @@ -170,6 +179,7 @@ const Database::DatabaseEntry XgemvFastDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, + { "Ellesmere", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "Hawaii", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Oland", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Pitcairn", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, @@ -207,6 +217,7 @@ const Database::DatabaseEntry XgemvFastDouble = { { "GeForce GTX TITAN", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "GeForce GTX TITAN Black", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "GeForce GTX TITAN X", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, + { "TITAN X (Pascal)", { {"VW2",1}, {"WGS2",32}, {"WPT2",1} } }, { "Tesla K20m", { {"VW2",1}, {"WGS2",128}, {"WPT2",1} } }, { "Tesla K40m", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "default", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, @@ -227,6 +238,7 @@ const Database::DatabaseEntry XgemvFastComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, + { "Ellesmere", { {"VW2",1}, {"WGS2",16}, {"WPT2",1} } }, { "Hawaii", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, { "Oland", { {"VW2",1}, {"WGS2",256}, {"WPT2",1} } }, { "Pitcairn", { {"VW2",1}, {"WGS2",64}, {"WPT2",1} } }, diff --git a/src/database/kernels/xgemv_fast_rot.hpp b/src/database/kernels/xgemv_fast_rot.hpp index 36a435b5..aff83ff2 100644 --- a/src/database/kernels/xgemv_fast_rot.hpp +++ b/src/database/kernels/xgemv_fast_rot.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XgemvFastRotHalf = { "XgemvFastRot", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, + { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics Skylake ULT GT2", { {"VW3",8}, {"WGS3",128}, {"WPT3",32} } }, @@ -38,6 +44,7 @@ const Database::DatabaseEntry XgemvFastRotSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW3",8}, {"WGS3",64}, {"WPT3",32} } }, + { "Ellesmere", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, { "Tonga", { {"VW3",8}, {"WGS3",128}, {"WPT3",32} } }, { "Turks", { {"VW3",8}, {"WGS3",128}, {"WPT3",16} } }, { "default", { {"VW3",8}, {"WGS3",128}, {"WPT3",32} } }, @@ -67,6 +74,7 @@ const Database::DatabaseEntry XgemvFastRotSingle = { { "GeForce GTX 750 Ti", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, { "GeForce GTX TITAN", { {"VW3",1}, {"WGS3",16}, {"WPT3",16} } }, { "GeForce GTX TITAN Black", { {"VW3",4}, {"WGS3",128}, {"WPT3",16} } }, + { "TITAN X (Pascal)", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, } }, @@ -85,6 +93,7 @@ const Database::DatabaseEntry XgemvFastRotComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW3",8}, {"WGS3",16}, {"WPT3",16} } }, + { "Ellesmere", { {"VW3",2}, {"WGS3",32}, {"WPT3",16} } }, { "Tonga", { {"VW3",4}, {"WGS3",32}, {"WPT3",32} } }, { "Turks", { {"VW3",4}, {"WGS3",32}, {"WPT3",8} } }, { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",32} } }, @@ -123,6 +132,7 @@ const Database::DatabaseEntry XgemvFastRotDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, + { "Ellesmere", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, { "Tonga", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, { "default", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, } @@ -141,6 +151,7 @@ const Database::DatabaseEntry XgemvFastRotDouble = { { "GeForce GTX 750 Ti", { {"VW3",4}, {"WGS3",32}, {"WPT3",16} } }, { "GeForce GTX TITAN", { {"VW3",1}, {"WGS3",16}, {"WPT3",16} } }, { "GeForce GTX TITAN Black", { {"VW3",1}, {"WGS3",16}, {"WPT3",16} } }, + { "TITAN X (Pascal)", { {"VW3",8}, {"WGS3",32}, {"WPT3",16} } }, { "default", { {"VW3",4}, {"WGS3",32}, {"WPT3",16} } }, } }, @@ -159,6 +170,7 @@ const Database::DatabaseEntry XgemvFastRotComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"VW3",4}, {"WGS3",32}, {"WPT3",16} } }, + { "Ellesmere", { {"VW3",4}, {"WGS3",16}, {"WPT3",16} } }, { "Tonga", { {"VW3",4}, {"WGS3",16}, {"WPT3",8} } }, { "default", { {"VW3",8}, {"WGS3",32}, {"WPT3",16} } }, } diff --git a/src/database/kernels/xger.hpp b/src/database/kernels/xger.hpp index f99b7632..ab528800 100644 --- a/src/database/kernels/xger.hpp +++ b/src/database/kernels/xger.hpp @@ -17,6 +17,12 @@ namespace database { const Database::DatabaseEntry XgerHalf = { "Xger", Precision::kHalf, { + { // AMD GPUs + kDeviceTypeGPU, "AMD", { + { "Ellesmere", { {"WGS1",64}, {"WGS2",1}, {"WPT",2} } }, + { "default", { {"WGS1",64}, {"WGS2",1}, {"WPT",2} } }, + } + }, { // Intel GPUs kDeviceTypeGPU, "Intel", { { "Intel(R) HD Graphics 5500 BroadWell U-Processor GT2", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, @@ -26,7 +32,7 @@ const Database::DatabaseEntry XgerHalf = { }, { // Default kDeviceTypeAll, "default", { - { "default", { {"WGS1",4}, {"WGS2",8}, {"WPT",2} } }, + { "default", { {"WGS1",64}, {"WGS2",1}, {"WPT",2} } }, } }, } @@ -39,13 +45,14 @@ const Database::DatabaseEntry XgerSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",256}, {"WGS2",1}, {"WPT",1} } }, + { "Ellesmere", { {"WGS1",64}, {"WGS2",4}, {"WPT",2} } }, { "Hawaii", { {"WGS1",64}, {"WGS2",2}, {"WPT",1} } }, { "Oland", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, { "Pitcairn", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, { "Tahiti", { {"WGS1",256}, {"WGS2",1}, {"WPT",1} } }, { "Tonga", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, { "Turks", { {"WGS1",64}, {"WGS2",4}, {"WPT",2} } }, - { "default", { {"WGS1",32}, {"WGS2",8}, {"WPT",1} } }, + { "default", { {"WGS1",32}, {"WGS2",8}, {"WPT",2} } }, } }, { // ARM GPUs @@ -86,12 +93,13 @@ const Database::DatabaseEntry XgerSingle = { { "GeForce GTX 750 Ti", { {"WGS1",64}, {"WGS2",1}, {"WPT",2} } }, { "GeForce GTX TITAN", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, { "GeForce GTX TITAN Black", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, - { "default", { {"WGS1",128}, {"WGS2",1}, {"WPT",2} } }, + { "TITAN X (Pascal)", { {"WGS1",128}, {"WGS2",4}, {"WPT",1} } }, + { "default", { {"WGS1",256}, {"WGS2",1}, {"WPT",2} } }, } }, { // Default kDeviceTypeAll, "default", { - { "default", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, + { "default", { {"WGS1",32}, {"WGS2",8}, {"WPT",2} } }, } }, } @@ -104,13 +112,14 @@ const Database::DatabaseEntry XgerComplexSingle = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WGS2",4}, {"WPT",1} } }, + { "Ellesmere", { {"WGS1",16}, {"WGS2",8}, {"WPT",2} } }, { "Hawaii", { {"WGS1",64}, {"WGS2",1}, {"WPT",2} } }, { "Oland", { {"WGS1",4}, {"WGS2",8}, {"WPT",1} } }, { "Pitcairn", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, { "Tahiti", { {"WGS1",64}, {"WGS2",2}, {"WPT",1} } }, { "Tonga", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, { "Turks", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, - { "default", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, + { "default", { {"WGS1",64}, {"WGS2",4}, {"WPT",1} } }, } }, { // ARM GPUs @@ -151,7 +160,8 @@ const Database::DatabaseEntry XgerComplexSingle = { { "GeForce GTX 750 Ti", { {"WGS1",32}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX TITAN", { {"WGS1",16}, {"WGS2",16}, {"WPT",2} } }, { "GeForce GTX TITAN Black", { {"WGS1",16}, {"WGS2",16}, {"WPT",2} } }, - { "default", { {"WGS1",128}, {"WGS2",1}, {"WPT",2} } }, + { "TITAN X (Pascal)", { {"WGS1",32}, {"WGS2",2}, {"WPT",1} } }, + { "default", { {"WGS1",128}, {"WGS2",2}, {"WPT",2} } }, } }, { // Default @@ -169,12 +179,13 @@ const Database::DatabaseEntry XgerDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",32}, {"WGS2",4}, {"WPT",1} } }, + { "Ellesmere", { {"WGS1",64}, {"WGS2",1}, {"WPT",4} } }, { "Hawaii", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, { "Oland", { {"WGS1",128}, {"WGS2",1}, {"WPT",2} } }, { "Pitcairn", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, { "Tahiti", { {"WGS1",64}, {"WGS2",2}, {"WPT",1} } }, { "Tonga", { {"WGS1",8}, {"WGS2",16}, {"WPT",2} } }, - { "default", { {"WGS1",32}, {"WGS2",8}, {"WPT",1} } }, + { "default", { {"WGS1",128}, {"WGS2",2}, {"WPT",1} } }, } }, { // ARM GPUs @@ -204,6 +215,7 @@ const Database::DatabaseEntry XgerDouble = { { "GeForce GTX 750 Ti", { {"WGS1",32}, {"WGS2",16}, {"WPT",1} } }, { "GeForce GTX TITAN", { {"WGS1",16}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX TITAN Black", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, + { "TITAN X (Pascal)", { {"WGS1",32}, {"WGS2",16}, {"WPT",1} } }, { "default", { {"WGS1",64}, {"WGS2",2}, {"WPT",2} } }, } }, @@ -222,6 +234,7 @@ const Database::DatabaseEntry XgerComplexDouble = { { // AMD GPUs kDeviceTypeGPU, "AMD", { { "AMD Radeon R9 M370X Compute Engine", { {"WGS1",64}, {"WGS2",1}, {"WPT",1} } }, + { "Ellesmere", { {"WGS1",8}, {"WGS2",16}, {"WPT",1} } }, { "Hawaii", { {"WGS1",128}, {"WGS2",1}, {"WPT",1} } }, { "Oland", { {"WGS1",16}, {"WGS2",16}, {"WPT",2} } }, { "Pitcairn", { {"WGS1",64}, {"WGS2",4}, {"WPT",1} } }, @@ -257,6 +270,7 @@ const Database::DatabaseEntry XgerComplexDouble = { { "GeForce GTX 750 Ti", { {"WGS1",32}, {"WGS2",8}, {"WPT",2} } }, { "GeForce GTX TITAN", { {"WGS1",32}, {"WGS2",4}, {"WPT",2} } }, { "GeForce GTX TITAN Black", { {"WGS1",16}, {"WGS2",16}, {"WPT",2} } }, + { "TITAN X (Pascal)", { {"WGS1",64}, {"WGS2",4}, {"WPT",1} } }, { "default", { {"WGS1",16}, {"WGS2",8}, {"WPT",2} } }, } }, |