summaryrefslogtreecommitdiff
path: root/src/routines
diff options
context:
space:
mode:
Diffstat (limited to 'src/routines')
-rw-r--r--src/routines/common.cpp56
-rw-r--r--src/routines/common.hpp19
-rw-r--r--src/routines/level1/xaxpy.cpp8
-rw-r--r--src/routines/level2/xgemv.cpp14
-rw-r--r--src/routines/level2/xger.cpp6
-rw-r--r--src/routines/level2/xher.cpp6
-rw-r--r--src/routines/level2/xher2.cpp6
-rw-r--r--src/routines/level3/xgemm.cpp61
-rw-r--r--src/routines/level3/xher2k.cpp29
-rw-r--r--src/routines/level3/xherk.cpp19
-rw-r--r--src/routines/level3/xsyr2k.cpp22
-rw-r--r--src/routines/level3/xsyrk.cpp17
-rw-r--r--src/routines/levelx/xomatcopy.cpp2
13 files changed, 116 insertions, 149 deletions
diff --git a/src/routines/common.cpp b/src/routines/common.cpp
index c378df28..3969cf9f 100644
--- a/src/routines/common.cpp
+++ b/src/routines/common.cpp
@@ -12,6 +12,7 @@
// =================================================================================================
#include <vector>
+#include <chrono>
#include "routines/common.hpp"
@@ -21,45 +22,54 @@ 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
const auto local_mem_usage = kernel.LocalMemUsage(device);
if (!device.IsLocalMemoryValid(local_mem_usage)) { return StatusCode::kInvalidLocalMemUsage; }
+ // Prints the name of the kernel to launch in case of debugging in verbose mode
+ #ifdef VERBOSE
+ queue.Finish();
+ printf("[DEBUG] Running kernel '%s'\n", kernel.GetFunctionName().c_str());
+ const auto start_time = std::chrono::steady_clock::now();
+ #endif
+
// Launches the kernel (and checks for launch errors)
try {
kernel.Launch(queue, global, local, event, waitForEvents);
} catch (...) { return StatusCode::kKernelLaunchError; }
+ // Prints the elapsed execution time in case of debugging in verbose mode
+ #ifdef VERBOSE
+ queue.Finish();
+ const auto elapsed_time = std::chrono::steady_clock::now() - start_time;
+ const auto timing = std::chrono::duration<double,std::milli>(elapsed_time).count();
+ printf("[DEBUG] Completed kernel in %.2lf ms\n", timing);
+ #endif
+
// No errors, normal termination of this function
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 c99cd39d..9d8849c3 100644
--- a/src/routines/common.hpp
+++ b/src/routines/common.hpp
@@ -29,21 +29,16 @@ 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 = {});
// =================================================================================================
// Copies or transposes a matrix and optionally pads/unpads it with zeros. This method is also able
// to write to symmetric and triangular matrices through optional arguments.
template <typename T>
-StatusCode PadCopyTransposeMatrix(Queue &queue, const Device &device, const Context &context,
+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,
@@ -88,10 +83,6 @@ StatusCode PadCopyTransposeMatrix(Queue &queue, const Device &device, const Cont
}
}
- // Upload the scalar argument as a constant buffer to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context, 1);
- alpha_buffer.Write(queue, 1, &alpha);
-
// Retrieves the kernel from the compiled binary
try {
auto kernel = Kernel(program, kernel_name);
@@ -101,7 +92,7 @@ StatusCode PadCopyTransposeMatrix(Queue &queue, const Device &device, const Cont
kernel.SetArgument(0, static_cast<int>(src_ld));
kernel.SetArgument(1, src());
kernel.SetArgument(2, dest());
- kernel.SetArgument(3, alpha_buffer());
+ kernel.SetArgument(3, GetRealArg(alpha));
}
else {
kernel.SetArgument(0, static_cast<int>(src_one));
@@ -114,7 +105,7 @@ StatusCode PadCopyTransposeMatrix(Queue &queue, const Device &device, const Cont
kernel.SetArgument(7, static_cast<int>(dest_ld));
kernel.SetArgument(8, static_cast<int>(dest_offset));
kernel.SetArgument(9, dest());
- kernel.SetArgument(10, alpha_buffer());
+ kernel.SetArgument(10, GetRealArg(alpha));
if (do_pad) {
kernel.SetArgument(11, static_cast<int>(do_conjugate));
}
diff --git a/src/routines/level1/xaxpy.cpp b/src/routines/level1/xaxpy.cpp
index 5b6c9e77..3445e2b5 100644
--- a/src/routines/level1/xaxpy.cpp
+++ b/src/routines/level1/xaxpy.cpp
@@ -59,20 +59,16 @@ StatusCode Xaxpy<T>::DoAxpy(const size_t n, const T alpha,
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
auto kernel = Kernel(program, kernel_name);
- // Upload the scalar argument as a constant buffer to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
-
// Sets the kernel arguments
if (use_fast_kernel) {
kernel.SetArgument(0, static_cast<int>(n));
- kernel.SetArgument(1, alpha_buffer());
+ kernel.SetArgument(1, GetRealArg(alpha));
kernel.SetArgument(2, x_buffer());
kernel.SetArgument(3, y_buffer());
}
else {
kernel.SetArgument(0, static_cast<int>(n));
- kernel.SetArgument(1, alpha_buffer());
+ kernel.SetArgument(1, GetRealArg(alpha));
kernel.SetArgument(2, x_buffer());
kernel.SetArgument(3, static_cast<int>(x_offset));
kernel.SetArgument(4, static_cast<int>(x_inc));
diff --git a/src/routines/level2/xgemv.cpp b/src/routines/level2/xgemv.cpp
index 21fb397c..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"
@@ -122,16 +122,10 @@ 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"];
}
- // Upload the scalar arguments as constant buffers to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- auto beta_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
- beta_buffer.Write(queue_, 1, &beta);
-
// Retrieves the Xgemv kernel from the compiled binary
try {
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
@@ -140,8 +134,8 @@ StatusCode Xgemv<T>::MatVec(const Layout layout, const Transpose a_transpose,
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(m_real));
kernel.SetArgument(1, static_cast<int>(n_real));
- kernel.SetArgument(2, alpha_buffer());
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(2, GetRealArg(alpha));
+ kernel.SetArgument(3, GetRealArg(beta));
kernel.SetArgument(4, static_cast<int>(a_rotated));
kernel.SetArgument(5, a_buffer());
kernel.SetArgument(6, static_cast<int>(a_offset));
diff --git a/src/routines/level2/xger.cpp b/src/routines/level2/xger.cpp
index 353047d2..29cffe0c 100644
--- a/src/routines/level2/xger.cpp
+++ b/src/routines/level2/xger.cpp
@@ -56,10 +56,6 @@ StatusCode Xger<T>::DoGer(const Layout layout,
status = TestVectorY(n, y_buffer, y_offset, y_inc);
if (ErrorIn(status)) { return status; }
- // Upload the scalar argument as a constant buffer to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
-
// Retrieves the kernel from the compiled binary
try {
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
@@ -68,7 +64,7 @@ StatusCode Xger<T>::DoGer(const Layout layout,
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(a_one));
kernel.SetArgument(1, static_cast<int>(a_two));
- kernel.SetArgument(2, alpha_buffer());
+ kernel.SetArgument(2, GetRealArg(alpha));
kernel.SetArgument(3, x_buffer());
kernel.SetArgument(4, static_cast<int>(x_offset));
kernel.SetArgument(5, static_cast<int>(x_inc));
diff --git a/src/routines/level2/xher.cpp b/src/routines/level2/xher.cpp
index ed8ba9e9..6dd95938 100644
--- a/src/routines/level2/xher.cpp
+++ b/src/routines/level2/xher.cpp
@@ -70,10 +70,6 @@ StatusCode Xher<T,U>::DoHer(const Layout layout, const Triangle triangle,
// Creates a matching version of alpha
const auto matching_alpha = GetAlpha(alpha);
- // Upload the scalar argument as a constant buffer to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &matching_alpha);
-
// Retrieves the kernel from the compiled binary
try {
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
@@ -81,7 +77,7 @@ StatusCode Xher<T,U>::DoHer(const Layout layout, const Triangle triangle,
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n));
- kernel.SetArgument(1, alpha_buffer());
+ kernel.SetArgument(1, GetRealArg(matching_alpha));
kernel.SetArgument(2, x_buffer());
kernel.SetArgument(3, static_cast<int>(x_offset));
kernel.SetArgument(4, static_cast<int>(x_inc));
diff --git a/src/routines/level2/xher2.cpp b/src/routines/level2/xher2.cpp
index 50572cea..3d57a9b9 100644
--- a/src/routines/level2/xher2.cpp
+++ b/src/routines/level2/xher2.cpp
@@ -58,10 +58,6 @@ StatusCode Xher2<T>::DoHer2(const Layout layout, const Triangle triangle,
status = TestVectorY(n, y_buffer, y_offset, y_inc);
if (ErrorIn(status)) { return status; }
- // Upload the scalar argument as a constant buffer to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
-
// Retrieves the kernel from the compiled binary
try {
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
@@ -69,7 +65,7 @@ StatusCode Xher2<T>::DoHer2(const Layout layout, const Triangle triangle,
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n));
- kernel.SetArgument(1, alpha_buffer());
+ kernel.SetArgument(1, GetRealArg(alpha));
kernel.SetArgument(2, x_buffer());
kernel.SetArgument(3, static_cast<int>(x_offset));
kernel.SetArgument(4, static_cast<int>(x_inc));
diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp
index 9ea5559c..0b8e768f 100644
--- a/src/routines/level3/xgemm.cpp
+++ b/src/routines/level3/xgemm.cpp
@@ -34,6 +34,7 @@ Xgemm<T>::Xgemm(Queue &queue, EventPointer event, const std::string &name):
#include "../../kernels/level3/convert_hermitian.opencl"
#include "../../kernels/level3/xgemm_part1.opencl"
#include "../../kernels/level3/xgemm_part2.opencl"
+ #include "../../kernels/level3/xgemm_part3.opencl"
;
}
@@ -63,9 +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);
@@ -99,6 +103,15 @@ StatusCode Xgemm<T>::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 {
@@ -106,23 +119,17 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
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);
-
- // Upload the scalar arguments as constant buffers to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- auto beta_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
- beta_buffer.Write(queue_, 1, &beta);
+ 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>();
@@ -133,9 +140,9 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
// case nothing has to be done, these kernels can be skipped.
if (!a_no_temp) {
auto eventProcessA = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList,
+ 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; }
@@ -145,9 +152,9 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
// As above, but now for matrix B
if (!b_no_temp) {
auto eventProcessB = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList,
+ 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; }
@@ -157,9 +164,9 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
// As above, but now for matrix C. This is only necessary if C is used both as input and output.
if (!c_no_temp && beta != static_cast<T>(0)) {
auto eventProcessC = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList,
+ 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; }
@@ -174,16 +181,16 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
kernel.SetArgument(0, static_cast<int>(m_ceiled));
kernel.SetArgument(1, static_cast<int>(n_ceiled));
kernel.SetArgument(2, static_cast<int>(k_ceiled));
- kernel.SetArgument(3, alpha_buffer());
- kernel.SetArgument(4, beta_buffer());
+ kernel.SetArgument(3, GetRealArg(alpha));
+ kernel.SetArgument(4, GetRealArg(beta));
kernel.SetArgument(5, a_temp());
kernel.SetArgument(6, b_temp());
kernel.SetArgument(7, c_temp());
// 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"]};
@@ -196,8 +203,8 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
// Runs the post-processing kernel if needed
if (!c_no_temp) {
eventWaitList.push_back(eventKernel);
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList,
- m_ceiled, n_ceiled, m_ceiled, 0, c_temp,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
+ 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);
diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp
index bd7a053e..ba770065 100644
--- a/src/routines/level3/xher2k.cpp
+++ b/src/routines/level3/xher2k.cpp
@@ -31,6 +31,7 @@ Xher2k<T,U>::Xher2k(Queue &queue, EventPointer event, const std::string &name):
#include "../../kernels/level3/transpose_pad.opencl"
#include "../../kernels/level3/xgemm_part1.opencl"
#include "../../kernels/level3/xgemm_part2.opencl"
+ #include "../../kernels/level3/xgemm_part3.opencl"
;
}
@@ -107,12 +108,8 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
auto b2_temp = (b2_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
- // Upload the scalar arguments as constant buffers to the device (needed for half-precision)
+ // Convert the arguments to complex versions
auto complex_beta = T{beta, static_cast<U>(0.0)};
- auto alpha_buffer = Buffer<T>(context_, 1);
- auto beta_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
- beta_buffer.Write(queue_, 1, &complex_beta);
// Events of all kernels (including pre/post processing kernels)
auto eventWaitList = std::vector<Event>();
@@ -123,7 +120,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
// case nothing has to be done, these kernels can be skipped.
if (!a1_no_temp) {
auto eventProcessA1 = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA1.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA1.pointer(), emptyEventList,
ab_one, ab_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, a1_temp,
ConstantOne<T>(), program,
@@ -133,7 +130,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
}
if (!a2_no_temp) {
auto eventProcessA2 = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA2.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA2.pointer(), emptyEventList,
ab_one, ab_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, a2_temp,
ConstantOne<T>(), program,
@@ -143,7 +140,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
}
if (!b1_no_temp) {
auto eventProcessB1 = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB1.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB1.pointer(), emptyEventList,
ab_one, ab_two, b_ld, b_offset, b_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, b1_temp,
ConstantOne<T>(), program,
@@ -153,7 +150,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
}
if (!b2_no_temp) {
auto eventProcessB2 = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB2.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB2.pointer(), emptyEventList,
ab_one, ab_two, b_ld, b_offset, b_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, b2_temp,
ConstantOne<T>(), program,
@@ -165,7 +162,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
// Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to
// modify the other triangle.
auto eventProcessC = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
ConstantOne<T>(), program,
@@ -180,8 +177,8 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n_ceiled));
kernel.SetArgument(1, static_cast<int>(k_ceiled));
- kernel.SetArgument(2, alpha_buffer());
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(2, GetRealArg(alpha));
+ kernel.SetArgument(3, GetRealArg(complex_beta));
kernel.SetArgument(4, a1_temp());
kernel.SetArgument(5, b2_temp());
kernel.SetArgument(6, c_temp());
@@ -202,10 +199,8 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
// Swaps the arguments for matrices A and B, sets 'beta' to 1, and conjugate alpha
auto conjugate_alpha = T{alpha.real(), -alpha.imag()};
auto complex_one = T{static_cast<U>(1.0), static_cast<U>(0.0)};
- alpha_buffer.Write(queue_, 1, &conjugate_alpha);
- beta_buffer.Write(queue_, 1, &complex_one);
- kernel.SetArgument(2, alpha_buffer());
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(2, GetRealArg(conjugate_alpha));
+ kernel.SetArgument(3, GetRealArg(complex_one));
kernel.SetArgument(4, b1_temp());
kernel.SetArgument(5, a2_temp());
@@ -218,7 +213,7 @@ StatusCode Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, co
// Runs the post-processing kernel
auto upper = (triangle == Triangle::kUpper);
auto lower = (triangle == Triangle::kLower);
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
n, n, c_ld, c_offset, c_buffer,
ConstantOne<T>(), program,
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp
index 6ef7f21f..3063f3bc 100644
--- a/src/routines/level3/xherk.cpp
+++ b/src/routines/level3/xherk.cpp
@@ -31,6 +31,7 @@ Xherk<T,U>::Xherk(Queue &queue, EventPointer event, const std::string &name):
#include "../../kernels/level3/transpose_pad.opencl"
#include "../../kernels/level3/xgemm_part1.opencl"
#include "../../kernels/level3/xgemm_part2.opencl"
+ #include "../../kernels/level3/xgemm_part3.opencl"
;
}
@@ -98,13 +99,9 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
auto b_temp = (b_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
- // Upload the scalar arguments as constant buffers to the device (needed for half-precision)
+ // Convert the arguments to complex versions
auto complex_alpha = T{alpha, static_cast<U>(0.0)};
auto complex_beta = T{beta, static_cast<U>(0.0)};
- auto alpha_buffer = Buffer<T>(context_, 1);
- auto beta_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &complex_alpha);
- beta_buffer.Write(queue_, 1, &complex_beta);
// Events of all kernels (including pre/post processing kernels)
auto eventWaitList = std::vector<Event>();
@@ -115,7 +112,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
// case nothing has to be done, these kernels can be skipped. Two copies are created.
if (!a_no_temp) {
auto eventProcessA = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
ConstantOne<T>(), program,
@@ -125,7 +122,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
}
if (!b_no_temp) {
auto eventProcessB = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
ConstantOne<T>(), program,
@@ -137,7 +134,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
// Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to
// modify the other triangle.
auto eventProcessC = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
ConstantOne<T>(), program,
@@ -152,8 +149,8 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n_ceiled));
kernel.SetArgument(1, static_cast<int>(k_ceiled));
- kernel.SetArgument(2, alpha_buffer());
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(2, GetRealArg(complex_alpha));
+ kernel.SetArgument(3, GetRealArg(complex_beta));
kernel.SetArgument(4, a_temp());
kernel.SetArgument(5, b_temp());
kernel.SetArgument(6, c_temp());
@@ -174,7 +171,7 @@ StatusCode Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, cons
// Runs the post-processing kernel
auto upper = (triangle == Triangle::kUpper);
auto lower = (triangle == Triangle::kLower);
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
n, n, c_ld, c_offset, c_buffer,
ConstantOne<T>(), program,
diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp
index 424d4d2d..158cd9e5 100644
--- a/src/routines/level3/xsyr2k.cpp
+++ b/src/routines/level3/xsyr2k.cpp
@@ -31,6 +31,7 @@ Xsyr2k<T>::Xsyr2k(Queue &queue, EventPointer event, const std::string &name):
#include "../../kernels/level3/transpose_pad.opencl"
#include "../../kernels/level3/xgemm_part1.opencl"
#include "../../kernels/level3/xgemm_part2.opencl"
+ #include "../../kernels/level3/xgemm_part3.opencl"
;
}
@@ -97,12 +98,6 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
- // Upload the scalar arguments as constant buffers to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- auto beta_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
- beta_buffer.Write(queue_, 1, &beta);
-
// Events of all kernels (including pre/post processing kernels)
auto eventWaitList = std::vector<Event>();
auto emptyEventList = std::vector<Event>();
@@ -112,7 +107,7 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
// case nothing has to be done, these kernels can be skipped.
if (!a_no_temp) {
auto eventProcessA = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
ab_one, ab_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
ConstantOne<T>(), program,
@@ -122,7 +117,7 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
}
if (!b_no_temp) {
auto eventProcessB = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
ab_one, ab_two, b_ld, b_offset, b_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
ConstantOne<T>(), program,
@@ -134,7 +129,7 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
// Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to
// modify the other triangle.
auto eventProcessC = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
ConstantOne<T>(), program,
@@ -149,8 +144,8 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n_ceiled));
kernel.SetArgument(1, static_cast<int>(k_ceiled));
- kernel.SetArgument(2, alpha_buffer());
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(2, GetRealArg(alpha));
+ kernel.SetArgument(3, GetRealArg(beta));
kernel.SetArgument(4, a_temp());
kernel.SetArgument(5, b_temp());
kernel.SetArgument(6, c_temp());
@@ -170,8 +165,7 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
// Swaps the arguments for matrices A and B, and sets 'beta' to 1
auto one = static_cast<T>(1);
- beta_buffer.Write(queue_, 1, &one);
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(3, GetRealArg(one));
kernel.SetArgument(4, b_temp());
kernel.SetArgument(5, a_temp());
@@ -184,7 +178,7 @@ StatusCode Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, cons
// Runs the post-processing kernel
auto upper = (triangle == Triangle::kUpper);
auto lower = (triangle == Triangle::kLower);
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
n, n, c_ld, c_offset, c_buffer,
ConstantOne<T>(), program,
diff --git a/src/routines/level3/xsyrk.cpp b/src/routines/level3/xsyrk.cpp
index f56c232b..e1a72ef6 100644
--- a/src/routines/level3/xsyrk.cpp
+++ b/src/routines/level3/xsyrk.cpp
@@ -31,6 +31,7 @@ Xsyrk<T>::Xsyrk(Queue &queue, EventPointer event, const std::string &name):
#include "../../kernels/level3/transpose_pad.opencl"
#include "../../kernels/level3/xgemm_part1.opencl"
#include "../../kernels/level3/xgemm_part2.opencl"
+ #include "../../kernels/level3/xgemm_part3.opencl"
;
}
@@ -90,12 +91,6 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const
auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
- // Upload the scalar arguments as constant buffers to the device (needed for half-precision)
- auto alpha_buffer = Buffer<T>(context_, 1);
- auto beta_buffer = Buffer<T>(context_, 1);
- alpha_buffer.Write(queue_, 1, &alpha);
- beta_buffer.Write(queue_, 1, &beta);
-
// Events of all kernels (including pre/post processing kernels)
auto eventWaitList = std::vector<Event>();
auto emptyEventList = std::vector<Event>();
@@ -105,7 +100,7 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const
// case nothing has to be done, these kernels can be skipped.
if (!a_no_temp) {
auto eventProcessA = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
ConstantOne<T>(), program,
@@ -117,7 +112,7 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const
// Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to
// modify the other triangle.
auto eventProcessC = Event();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
ConstantOne<T>(), program,
@@ -132,8 +127,8 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n_ceiled));
kernel.SetArgument(1, static_cast<int>(k_ceiled));
- kernel.SetArgument(2, alpha_buffer());
- kernel.SetArgument(3, beta_buffer());
+ kernel.SetArgument(2, GetRealArg(alpha));
+ kernel.SetArgument(3, GetRealArg(beta));
kernel.SetArgument(4, a_temp());
kernel.SetArgument(5, a_temp());
kernel.SetArgument(6, c_temp());
@@ -154,7 +149,7 @@ StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const
// Runs the post-processing kernel
auto upper = (triangle == Triangle::kUpper);
auto lower = (triangle == Triangle::kLower);
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
n, n, c_ld, c_offset, c_buffer,
ConstantOne<T>(), program,
diff --git a/src/routines/levelx/xomatcopy.cpp b/src/routines/levelx/xomatcopy.cpp
index e8593301..af9080af 100644
--- a/src/routines/levelx/xomatcopy.cpp
+++ b/src/routines/levelx/xomatcopy.cpp
@@ -72,7 +72,7 @@ StatusCode Xomatcopy<T>::DoOmatcopy(const Layout layout, const Transpose a_trans
const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
auto emptyEventList = std::vector<Event>();
- status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, emptyEventList,
+ status = PadCopyTransposeMatrix(queue_, device_, db_, event_, emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
b_one, b_two, b_ld, b_offset, b_buffer,
alpha, program, false, transpose, conjugate);
generated by cgit v1.2.3 (git 2.39.1) at 2024-08-01 02:57:09 +0000