diff options
Diffstat (limited to 'test/routines/level2')
-rw-r--r-- | test/routines/level2/xgbmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xgemv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xger.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xgerc.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xgeru.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xhbmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xhemv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xher.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xher2.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xhpmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xhpr.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xhpr2.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xsbmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xspmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xspr.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xspr2.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xsymv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xsyr.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xsyr2.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xtbmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xtpmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xtrmv.hpp | 5 | ||||
-rw-r--r-- | test/routines/level2/xtrsv.hpp | 53 |
23 files changed, 130 insertions, 33 deletions
diff --git a/test/routines/level2/xgbmv.hpp b/test/routines/level2/xgbmv.hpp index 5ed92aae..f371b9a7 100644 --- a/test/routines/level2/xgbmv.hpp +++ b/test/routines/level2/xgbmv.hpp @@ -80,6 +80,11 @@ class TestXgbmv { static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xgemv.hpp b/test/routines/level2/xgemv.hpp index 9ee6d535..2442be4c 100644 --- a/test/routines/level2/xgemv.hpp +++ b/test/routines/level2/xgemv.hpp @@ -80,6 +80,11 @@ class TestXgemv { static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xger.hpp b/test/routines/level2/xger.hpp index 42283107..3e7ccbc3 100644 --- a/test/routines/level2/xger.hpp +++ b/test/routines/level2/xger.hpp @@ -76,6 +76,11 @@ class TestXger { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xgerc.hpp b/test/routines/level2/xgerc.hpp index ef69c197..d880ae1f 100644 --- a/test/routines/level2/xgerc.hpp +++ b/test/routines/level2/xgerc.hpp @@ -76,6 +76,11 @@ class TestXgerc { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xgeru.hpp b/test/routines/level2/xgeru.hpp index b2afc6d8..1735e42a 100644 --- a/test/routines/level2/xgeru.hpp +++ b/test/routines/level2/xgeru.hpp @@ -76,6 +76,11 @@ class TestXgeru { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xhbmv.hpp b/test/routines/level2/xhbmv.hpp index 8bda4d0c..99538bf1 100644 --- a/test/routines/level2/xhbmv.hpp +++ b/test/routines/level2/xhbmv.hpp @@ -74,6 +74,11 @@ class TestXhbmv { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xhemv.hpp b/test/routines/level2/xhemv.hpp index 80565d04..3792cb66 100644 --- a/test/routines/level2/xhemv.hpp +++ b/test/routines/level2/xhemv.hpp @@ -74,6 +74,11 @@ class TestXhemv { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xher.hpp b/test/routines/level2/xher.hpp index d71c8009..c58eb189 100644 --- a/test/routines/level2/xher.hpp +++ b/test/routines/level2/xher.hpp @@ -70,6 +70,11 @@ class TestXher { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<U>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<U> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xher2.hpp b/test/routines/level2/xher2.hpp index 083dfa2f..8a7eb0b6 100644 --- a/test/routines/level2/xher2.hpp +++ b/test/routines/level2/xher2.hpp @@ -74,6 +74,11 @@ class TestXher2 { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xhpmv.hpp b/test/routines/level2/xhpmv.hpp index 1dd63562..0862b619 100644 --- a/test/routines/level2/xhpmv.hpp +++ b/test/routines/level2/xhpmv.hpp @@ -74,6 +74,11 @@ class TestXhpmv { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xhpr.hpp b/test/routines/level2/xhpr.hpp index a5c77811..5b454174 100644 --- a/test/routines/level2/xhpr.hpp +++ b/test/routines/level2/xhpr.hpp @@ -70,6 +70,11 @@ class TestXhpr { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<U>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<U> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xhpr2.hpp b/test/routines/level2/xhpr2.hpp index d09178f0..b770da2e 100644 --- a/test/routines/level2/xhpr2.hpp +++ b/test/routines/level2/xhpr2.hpp @@ -74,6 +74,11 @@ class TestXhpr2 { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xsbmv.hpp b/test/routines/level2/xsbmv.hpp index 8e0f8321..7a836170 100644 --- a/test/routines/level2/xsbmv.hpp +++ b/test/routines/level2/xsbmv.hpp @@ -74,6 +74,11 @@ class TestXsbmv { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xspmv.hpp b/test/routines/level2/xspmv.hpp index 977f733a..352c8cfd 100644 --- a/test/routines/level2/xspmv.hpp +++ b/test/routines/level2/xspmv.hpp @@ -74,6 +74,11 @@ class TestXspmv { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xspr.hpp b/test/routines/level2/xspr.hpp index 93da4b73..988bcdc2 100644 --- a/test/routines/level2/xspr.hpp +++ b/test/routines/level2/xspr.hpp @@ -70,6 +70,11 @@ class TestXspr { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xspr2.hpp b/test/routines/level2/xspr2.hpp index b835f2b0..ee517bc1 100644 --- a/test/routines/level2/xspr2.hpp +++ b/test/routines/level2/xspr2.hpp @@ -74,6 +74,11 @@ class TestXspr2 { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xsymv.hpp b/test/routines/level2/xsymv.hpp index 0ec96f1d..5eecfb74 100644 --- a/test/routines/level2/xsymv.hpp +++ b/test/routines/level2/xsymv.hpp @@ -74,6 +74,11 @@ class TestXsymv { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xsyr.hpp b/test/routines/level2/xsyr.hpp index b49132e3..ac4ee1ff 100644 --- a/test/routines/level2/xsyr.hpp +++ b/test/routines/level2/xsyr.hpp @@ -70,6 +70,11 @@ class TestXsyr { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xsyr2.hpp b/test/routines/level2/xsyr2.hpp index 7c65daa2..43644883 100644 --- a/test/routines/level2/xsyr2.hpp +++ b/test/routines/level2/xsyr2.hpp @@ -74,6 +74,11 @@ class TestXsyr2 { static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xtbmv.hpp b/test/routines/level2/xtbmv.hpp index cf30c2f7..ab9244af 100644 --- a/test/routines/level2/xtbmv.hpp +++ b/test/routines/level2/xtbmv.hpp @@ -69,6 +69,11 @@ class TestXtbmv { static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xtpmv.hpp b/test/routines/level2/xtpmv.hpp index d08e132f..3821e1a4 100644 --- a/test/routines/level2/xtpmv.hpp +++ b/test/routines/level2/xtpmv.hpp @@ -69,6 +69,11 @@ class TestXtpmv { static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xtrmv.hpp b/test/routines/level2/xtrmv.hpp index cf9a0063..7211c757 100644 --- a/test/routines/level2/xtrmv.hpp +++ b/test/routines/level2/xtrmv.hpp @@ -69,6 +69,11 @@ class TestXtrmv { static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T>&, Queue&, const int, std::vector<T>&, + std::vector<T>&, std::vector<T>&, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) {} // N/A for this routine + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { auto queue_plain = queue(); diff --git a/test/routines/level2/xtrsv.hpp b/test/routines/level2/xtrsv.hpp index fed4378a..78b9672f 100644 --- a/test/routines/level2/xtrsv.hpp +++ b/test/routines/level2/xtrsv.hpp @@ -29,36 +29,6 @@ namespace clblast { // ================================================================================================= -// Prepares the data -template <typename T> -void PrepareData(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { - if (args.a_ld < args.n) { return; } - if (args.a_size <= 0 || args.x_size <= 0) { return; } - - // Copies input buffers to the host - std::vector<T> a_mat_cpu(args.a_size, static_cast<T>(0)); - std::vector<T> x_vec_cpu(args.x_size, static_cast<T>(0)); - buffers.a_mat.Read(queue, args.a_size, a_mat_cpu); - buffers.x_vec.Read(queue, args.x_size, x_vec_cpu); - - // Generates 'proper' input for the TRSV routine - // TODO: Improve this, currently loosely based on clBLAS's implementation - for (auto i = size_t{0}; i < args.n; ++i) { - auto diagonal = a_mat_cpu[i*args.a_ld + i + args.a_offset]; - diagonal = AbsoluteValue(diagonal) + static_cast<T>(args.n / size_t{4}); - for (auto j = size_t{0}; j < args.n; ++j) { - a_mat_cpu[j*args.a_ld + i + args.a_offset] /= ConstantTwo<T>(); - } - a_mat_cpu[i*args.a_ld + i + args.a_offset] = diagonal; - x_vec_cpu[i * args.x_inc + args.x_offset] /= ConstantTwo<T>(); - } - - // Copies input buffers back to the OpenCL device - buffers.a_mat.Write(queue, args.a_size, a_mat_cpu); - buffers.x_vec.Write(queue, args.x_size, x_vec_cpu); - return; -} - // See comment at top of file for a description of the class template <typename T> class TestXtrsv { @@ -99,9 +69,28 @@ class TestXtrsv { static Transposes GetATransposes(const Transposes &all) { return all; } static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + // Describes how to prepare the input data + static void PrepareData(const Arguments<T> &args, Queue&, const int, std::vector<T> &x_source, + std::vector<T>&, std::vector<T> &a_source, std::vector<T>&, std::vector<T>&, + std::vector<T>&, std::vector<T>&) { + if (args.a_ld < args.n) { return; } + if (args.a_size <= 0 || args.x_size <= 0) { return; } + + // Generates 'proper' input for the TRSV routine + // TODO: Improve this, currently loosely based on clBLAS's implementation + for (auto i = size_t{0}; i < args.n; ++i) { + auto diagonal = a_source[i*args.a_ld + i + args.a_offset]; + diagonal = static_cast<T>(AbsoluteValue(diagonal)) + static_cast<T>(args.n / size_t{4}); + for (auto j = size_t{0}; j < args.n; ++j) { + a_source[j*args.a_ld + i + args.a_offset] /= Constant<T>(2.0); + } + a_source[i*args.a_ld + i + args.a_offset] = diagonal; + x_source[i * args.x_inc + args.x_offset] /= Constant<T>(2.0); + } + } + // Describes how to run the CLBlast routine static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { - PrepareData(args, buffers, queue); auto queue_plain = queue(); auto event = cl_event{}; auto status = Trsv<T>(args.layout, args.triangle, args.a_transpose, args.diagonal, @@ -116,7 +105,6 @@ class TestXtrsv { // Describes how to run the clBLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CLBLAS static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { - PrepareData(args, buffers, queue); auto queue_plain = queue(); auto event = cl_event{}; auto status = clblasXtrsv<T>(convertToCLBLAS(args.layout), @@ -135,7 +123,6 @@ class TestXtrsv { // Describes how to run the CPU BLAS routine (for correctness/performance comparison) #ifdef CLBLAST_REF_CBLAS static StatusCode RunReference2(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { - PrepareData(args, buffers, queue); std::vector<T> a_mat_cpu(args.a_size, static_cast<T>(0)); std::vector<T> x_vec_cpu(args.x_size, static_cast<T>(0)); buffers.a_mat.Read(queue, args.a_size, a_mat_cpu); |