summaryrefslogtreecommitdiff
path: root/src/routines
diff options
context:
space:
mode:
Diffstat (limited to 'src/routines')
-rw-r--r--src/routines/common.cpp40
-rw-r--r--src/routines/common.hpp9
-rw-r--r--src/routines/level2/xgemv.cpp4
-rw-r--r--src/routines/level3/xgemm.cpp42
4 files changed, 48 insertions, 47 deletions
diff --git a/src/routines/common.cpp b/src/routines/common.cpp
index 2e82e04d..3969cf9f 100644
--- a/src/routines/common.cpp
+++ b/src/routines/common.cpp
@@ -22,23 +22,25 @@ namespace clblast {
// Enqueues a kernel, waits for completion, and checks for errors
StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device,
std::vector<size_t> global, const std::vector<size_t> &local,
- EventPointer event, std::vector<Event>& waitForEvents) {
+ EventPointer event, const std::vector<Event> &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<local.size(); ++i) {
- if (local[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<local.size(); ++i) {
+ if (local[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.size(); ++i) {
- if (global[i] < local[i]) { global[i] = local[i]; }
+ // Make sure the global thread sizes are at least equal to the local sizes
+ for (auto i=size_t{0}; i<global.size(); ++i) {
+ if (global[i] < local[i]) { global[i] = local[i]; }
+ }
}
// Tests for local memory usage
@@ -69,13 +71,5 @@ StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device,
return StatusCode::kSuccess;
}
-// As above, but without an event waiting list
-StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device,
- std::vector<size_t> global, const std::vector<size_t> &local,
- EventPointer event) {
- auto emptyWaitingList = std::vector<Event>();
- return RunKernel(kernel, queue, device, global, local, event, emptyWaitingList);
-}
-
// =================================================================================================
} // namespace clblast
diff --git a/src/routines/common.hpp b/src/routines/common.hpp
index d53bdc25..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<size_t> global, const std::vector<size_t> &local,
- EventPointer event, std::vector<Event>& waitForEvents);
-
-// As above, but without an event waiting list
-StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device,
- std::vector<size_t> global, const std::vector<size_t> &local,
- EventPointer event);
+ EventPointer event, const std::vector<Event> &waitForEvents = {});
// =================================================================================================
@@ -43,7 +38,7 @@ StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device,
template <typename T>
StatusCode PadCopyTransposeMatrix(Queue &queue, const Device &device,
const Database &db,
- EventPointer event, std::vector<Event>& waitForEvents,
+ EventPointer event, const std::vector<Event> &waitForEvents,
const size_t src_one, const size_t src_two,
const size_t src_ld, const size_t src_offset,
const Buffer<T> &src,
diff --git a/src/routines/level2/xgemv.cpp b/src/routines/level2/xgemv.cpp
index 2842ef07..4e32ba41 100644
--- a/src/routines/level2/xgemv.cpp
+++ b/src/routines/level2/xgemv.cpp
@@ -22,7 +22,7 @@ namespace clblast {
// Constructor: forwards to base class constructor
template <typename T>
Xgemv<T>::Xgemv(Queue &queue, EventPointer event, const std::string &name):
- Routine(queue, event, name, {"Pad", "Xgemv"}, PrecisionValue<T>()) {
+ Routine(queue, event, name, {"Pad", "Xgemv", "XgemvFast", "XgemvFastRot"}, PrecisionValue<T>()) {
source_string_ =
#include "../../kernels/level2/xgemv.opencl"
#include "../../kernels/level2/xgemv_fast.opencl"
@@ -122,7 +122,7 @@ StatusCode Xgemv<T>::MatVec(const Layout layout, const Transpose a_transpose,
}
if (fast_kernel_rot) {
kernel_name = "XgemvFastRot";
- global_size = m_real / db_["WPT3"];
+ global_size = m_real;
local_size = db_["WGS3"];
}
diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp
index d6ba2c32..e2e8647e 100644
--- a/src/routines/level3/xgemm.cpp
+++ b/src/routines/level3/xgemm.cpp
@@ -64,9 +64,12 @@ StatusCode Xgemm<T>::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);
@@ -136,6 +139,15 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
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 {
@@ -143,17 +155,17 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), 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<T>(context_, k_ceiled*m_ceiled);
- const auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
- const auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, m_ceiled*n_ceiled);
+ const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, a_one_i*a_two_i);
+ const auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, b_one_i*b_two_i);
+ const auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, c_one_i*c_two_i);
// Events of all kernels (including pre/post processing kernels)
auto eventWaitList = std::vector<Event>();
@@ -166,7 +178,7 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
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<T>(), program,
true, a_do_transpose, a_conjugate);
if (ErrorIn(status)) { return status; }
@@ -178,7 +190,7 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
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<T>(), program,
true, b_do_transpose, b_conjugate);
if (ErrorIn(status)) { return status; }
@@ -190,7 +202,7 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
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<T>(), program,
true, c_do_transpose, false);
if (ErrorIn(status)) { return status; }
@@ -213,8 +225,8 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
// Computes the global and local thread sizes
const auto global = std::vector<size_t>{
- (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<size_t>{db_["MDIMC"], db_["NDIMC"]};
@@ -228,7 +240,7 @@ StatusCode Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k
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<T>(), program,
false, c_do_transpose, false);
generated by cgit v1.2.3 (git 2.39.1) at 2024-08-01 07:03:22 +0000