From 1ae71614ac227e9b80678505fcb5ecc18a3d3383 Mon Sep 17 00:00:00 2001 From: Ivan Shapovalov Date: Sun, 10 Jul 2016 23:11:27 +0300 Subject: xgemm: do not hardcode kernel requirements for internal matrix layout Do not hardcode the knowledge about "A and C col-major, B row-major". This allows for easier reuse of the DoGemm() routine with different kernels. --- src/routines/level3/xgemm.cpp | 42 +++++++++++++++++++++++++++--------------- 1 file changed, 27 insertions(+), 15 deletions(-) (limited to 'src/routines') diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp index 0db28537..fce59622 100644 --- a/src/routines/level3/xgemm.cpp +++ b/src/routines/level3/xgemm.cpp @@ -63,9 +63,12 @@ StatusCode Xgemm::DoGemm(const Layout layout, const auto b_rotated = (layout == Layout::kColMajor && b_transpose != Transpose::kNo) || (layout == Layout::kRowMajor && b_transpose == Transpose::kNo); const auto c_rotated = (layout == Layout::kRowMajor); - const auto a_do_transpose = a_rotated; - const auto b_do_transpose = !b_rotated; - const auto c_do_transpose = c_rotated; + static const auto a_want_rotated = false; + static const auto b_want_rotated = true; + static const auto c_want_rotated = false; + const auto a_do_transpose = a_rotated != a_want_rotated; + const auto b_do_transpose = b_rotated != b_want_rotated; + const auto c_do_transpose = c_rotated != c_want_rotated; // In case of complex data-types, the transpose can also become a conjugate transpose const auto a_conjugate = (a_transpose == Transpose::kConjugate); @@ -99,6 +102,15 @@ StatusCode Xgemm::DoGemm(const Layout layout, const auto n_ceiled = Ceil(n, db_["NWG"]); const auto k_ceiled = Ceil(k, db_["KWG"]); + // Computes the first and second "internal" (ceiled) dimensions of the 3 matrices taking into account + // whether the matrices need to be rotated or not for the kernel. + const auto a_one_i = (a_want_rotated) ? k_ceiled : m_ceiled; + const auto a_two_i = (a_want_rotated) ? m_ceiled : k_ceiled; + const auto b_one_i = (b_want_rotated) ? n_ceiled : k_ceiled; + const auto b_two_i = (b_want_rotated) ? k_ceiled : n_ceiled; + const auto c_one_i = (c_want_rotated) ? n_ceiled : m_ceiled; + const auto c_two_i = (c_want_rotated) ? m_ceiled : n_ceiled; + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception try { @@ -106,17 +118,17 @@ StatusCode Xgemm::DoGemm(const Layout layout, const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); // Determines whether or not temporary matrices are needed - auto a_no_temp = a_one == m_ceiled && a_two == k_ceiled && a_ld == m_ceiled && a_offset == 0 && + auto a_no_temp = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && a_offset == 0 && a_do_transpose == false && a_conjugate == false; - auto b_no_temp = b_one == n_ceiled && b_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + auto b_no_temp = b_one == b_one_i && b_two == b_two_i && b_ld == b_one && b_offset == 0 && b_do_transpose == false && b_conjugate == false; - auto c_no_temp = c_one == m_ceiled && c_two == n_ceiled && c_ld == m_ceiled && c_offset == 0 && + auto c_no_temp = c_one == c_one_i && c_two == c_two_i && c_ld == c_one && c_offset == 0 && c_do_transpose == false; // Creates the temporary matrices - const auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*m_ceiled); - const auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - const auto c_temp = (c_no_temp) ? c_buffer : Buffer(context_, m_ceiled*n_ceiled); + const auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, a_one_i*a_two_i); + const auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, b_one_i*b_two_i); + const auto c_temp = (c_no_temp) ? c_buffer : Buffer(context_, c_one_i*c_two_i); // Events of all kernels (including pre/post processing kernels) auto eventWaitList = std::vector(); @@ -129,7 +141,7 @@ StatusCode Xgemm::DoGemm(const Layout layout, auto eventProcessA = Event(); status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, a_one, a_two, a_ld, a_offset, a_buffer, - m_ceiled, k_ceiled, m_ceiled, 0, a_temp, + a_one_i, a_two_i, a_one_i, 0, a_temp, ConstantOne(), program, true, a_do_transpose, a_conjugate); if (ErrorIn(status)) { return status; } @@ -141,7 +153,7 @@ StatusCode Xgemm::DoGemm(const Layout layout, auto eventProcessB = Event(); status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, b_one, b_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + b_one_i, b_two_i, b_one_i, 0, b_temp, ConstantOne(), program, true, b_do_transpose, b_conjugate); if (ErrorIn(status)) { return status; } @@ -153,7 +165,7 @@ StatusCode Xgemm::DoGemm(const Layout layout, auto eventProcessC = Event(); status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, c_one, c_two, c_ld, c_offset, c_buffer, - m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + c_one_i, c_two_i, c_one_i, 0, c_temp, ConstantOne(), program, true, c_do_transpose, false); if (ErrorIn(status)) { return status; } @@ -176,8 +188,8 @@ StatusCode Xgemm::DoGemm(const Layout layout, // Computes the global and local thread sizes const auto global = std::vector{ - (m_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] + (c_one_i * db_["MDIMC"]) / db_["MWG"], + (c_two_i * db_["NDIMC"]) / db_["NWG"] }; const auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; @@ -191,7 +203,7 @@ StatusCode Xgemm::DoGemm(const Layout layout, if (!c_no_temp) { eventWaitList.push_back(eventKernel); status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, - m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + c_one_i, c_two_i, c_one_i, 0, c_temp, c_one, c_two, c_ld, c_offset, c_buffer, ConstantOne(), program, false, c_do_transpose, false); -- cgit v1.2.3 From 2dd5ee3f758f817238cf74c068c5cad6eb3d46dd Mon Sep 17 00:00:00 2001 From: Ivan Shapovalov Date: Mon, 11 Jul 2016 22:36:39 +0300 Subject: clblast::RunKernel, cl::Kernel: take const vector as waitForEvents --- src/clpp11.hpp | 4 +++- src/routines/common.cpp | 2 +- src/routines/common.hpp | 4 ++-- 3 files changed, 6 insertions(+), 4 deletions(-) (limited to 'src/routines') diff --git a/src/clpp11.hpp b/src/clpp11.hpp index af9d2ea4..9fa683ac 100644 --- a/src/clpp11.hpp +++ b/src/clpp11.hpp @@ -109,7 +109,9 @@ class Event { // Accessor to the private data-member cl_event& operator()() { return *event_; } + const cl_event& operator()() const { return *event_; } cl_event* pointer() { return &(*event_); } + const cl_event* pointer() const { return &(*event_); } private: std::shared_ptr event_; }; @@ -686,7 +688,7 @@ class Kernel { // As above, but with an event waiting list void Launch(const Queue &queue, const std::vector &global, const std::vector &local, EventPointer event, - std::vector& waitForEvents) { + const std::vector &waitForEvents) { if (waitForEvents.size() == 0) { return Launch(queue, global, local, event); } // Builds a plain version of the events waiting list diff --git a/src/routines/common.cpp b/src/routines/common.cpp index 2e82e04d..21e16954 100644 --- a/src/routines/common.cpp +++ b/src/routines/common.cpp @@ -22,7 +22,7 @@ namespace clblast { // Enqueues a kernel, waits for completion, and checks for errors StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, std::vector global, const std::vector &local, - EventPointer event, std::vector& waitForEvents) { + EventPointer event, const std::vector &waitForEvents) { // Tests for validity of the local thread sizes if (local.size() > device.MaxWorkItemDimensions()) { diff --git a/src/routines/common.hpp b/src/routines/common.hpp index d53bdc25..d0bbc707 100644 --- a/src/routines/common.hpp +++ b/src/routines/common.hpp @@ -29,7 +29,7 @@ namespace clblast { // Enqueues a kernel, waits for completion, and checks for errors StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, std::vector global, const std::vector &local, - EventPointer event, std::vector& waitForEvents); + EventPointer event, const std::vector &waitForEvents); // As above, but without an event waiting list StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, @@ -43,7 +43,7 @@ StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, template StatusCode PadCopyTransposeMatrix(Queue &queue, const Device &device, const Database &db, - EventPointer event, std::vector& waitForEvents, + EventPointer event, const std::vector &waitForEvents, const size_t src_one, const size_t src_two, const size_t src_ld, const size_t src_offset, const Buffer &src, -- cgit v1.2.3 From ae3299da302ba2c26f1e3490a1a7ee389d91feee Mon Sep 17 00:00:00 2001 From: Ivan Shapovalov Date: Mon, 11 Jul 2016 23:14:43 +0300 Subject: clblast::RunKernel, cl::Kernel: unify variants with/without waitForEvents, support empty LWS --- src/clpp11.hpp | 13 ++----------- src/routines/common.cpp | 38 ++++++++++++++++---------------------- src/routines/common.hpp | 7 +------ 3 files changed, 19 insertions(+), 39 deletions(-) (limited to 'src/routines') diff --git a/src/clpp11.hpp b/src/clpp11.hpp index debfff09..d57223dd 100644 --- a/src/clpp11.hpp +++ b/src/clpp11.hpp @@ -695,23 +695,14 @@ class Kernel { if (waitEvent()) { waitForEventsPlain.push_back(waitEvent()); } } - if (waitForEvents.size() == 0) { return Launch(queue, global, local, event); } - // Launches the kernel while waiting for other events CheckError(clEnqueueNDRangeKernel(queue(), *kernel_, static_cast(global.size()), - nullptr, global.data(), local.data(), + nullptr, global.data(), !local.empty() ? local.data() : nullptr, static_cast(waitForEventsPlain.size()), - waitForEventsPlain.data(), + !waitForEventsPlain.empty() ? waitForEventsPlain.data() : nullptr, event)); } - // As above, but with the default local workgroup size - void Launch(const Queue &queue, const std::vector &global, EventPointer event) { - CheckError(clEnqueueNDRangeKernel(queue(), *kernel_, static_cast(global.size()), - nullptr, global.data(), nullptr, - 0, nullptr, event)); - } - // Accessor to the private data-member const cl_kernel& operator()() const { return *kernel_; } private: diff --git a/src/routines/common.cpp b/src/routines/common.cpp index 21e16954..3969cf9f 100644 --- a/src/routines/common.cpp +++ b/src/routines/common.cpp @@ -24,21 +24,23 @@ StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, std::vector global, const std::vector &local, EventPointer event, const std::vector &waitForEvents) { - // Tests for validity of the local thread sizes - if (local.size() > device.MaxWorkItemDimensions()) { - return StatusCode::kInvalidLocalNumDimensions; - } - const auto max_work_item_sizes = device.MaxWorkItemSizes(); - for (auto i=size_t{0}; i max_work_item_sizes[i]) { return StatusCode::kInvalidLocalThreadsDim; } - } - auto local_size = size_t{1}; - for (auto &item: local) { local_size *= item; } - if (local_size > device.MaxWorkGroupSize()) { return StatusCode::kInvalidLocalThreadsTotal; } + if (!local.empty()) { + // Tests for validity of the local thread sizes + if (local.size() > device.MaxWorkItemDimensions()) { + return StatusCode::kInvalidLocalNumDimensions; + } + const auto max_work_item_sizes = device.MaxWorkItemSizes(); + for (auto i=size_t{0}; i max_work_item_sizes[i]) { return StatusCode::kInvalidLocalThreadsDim; } + } + auto local_size = size_t{1}; + for (auto &item: local) { local_size *= item; } + if (local_size > device.MaxWorkGroupSize()) { return StatusCode::kInvalidLocalThreadsTotal; } - // Make sure the global thread sizes are at least equal to the local sizes - for (auto i=size_t{0}; i global, const std::vector &local, - EventPointer event) { - auto emptyWaitingList = std::vector(); - return RunKernel(kernel, queue, device, global, local, event, emptyWaitingList); -} - // ================================================================================================= } // namespace clblast diff --git a/src/routines/common.hpp b/src/routines/common.hpp index d0bbc707..9d8849c3 100644 --- a/src/routines/common.hpp +++ b/src/routines/common.hpp @@ -29,12 +29,7 @@ namespace clblast { // Enqueues a kernel, waits for completion, and checks for errors StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, std::vector global, const std::vector &local, - EventPointer event, const std::vector &waitForEvents); - -// As above, but without an event waiting list -StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, - std::vector global, const std::vector &local, - EventPointer event); + EventPointer event, const std::vector &waitForEvents = {}); // ================================================================================================= -- cgit v1.2.3