diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2016-07-16 10:56:37 +0200 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2016-07-16 10:56:37 +0200 |
commit | 066af4069ba5c92decc7652e5c6d36c27849ccab (patch) | |
tree | 6e86ca648d9b59acf920255b66ebe57106935d9a /src/routines/level3 | |
parent | c87e877bf23d2fe38a7da2898e1734a3cdeaf48c (diff) |
Removed an unused variable from the copy-transpose-pad function
Diffstat (limited to 'src/routines/level3')
-rw-r--r-- | src/routines/level3/xgemm.cpp | 8 | ||||
-rw-r--r-- | src/routines/level3/xher2k.cpp | 12 | ||||
-rw-r--r-- | src/routines/level3/xherk.cpp | 8 | ||||
-rw-r--r-- | src/routines/level3/xsyr2k.cpp | 8 | ||||
-rw-r--r-- | src/routines/level3/xsyrk.cpp | 6 |
5 files changed, 21 insertions, 21 deletions
diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp index 97e8db7e..0db28537 100644 --- a/src/routines/level3/xgemm.cpp +++ b/src/routines/level3/xgemm.cpp @@ -127,7 +127,7 @@ 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, ConstantOne<T>(), program, @@ -139,7 +139,7 @@ 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, ConstantOne<T>(), program, @@ -151,7 +151,7 @@ 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, ConstantOne<T>(), program, @@ -190,7 +190,7 @@ 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, + status = PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, m_ceiled, n_ceiled, m_ceiled, 0, c_temp, c_one, c_two, c_ld, c_offset, c_buffer, ConstantOne<T>(), program, diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp index 65e2be55..1ba6080f 100644 --- a/src/routines/level3/xher2k.cpp +++ b/src/routines/level3/xher2k.cpp @@ -119,7 +119,7 @@ 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, @@ -129,7 +129,7 @@ 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, @@ -139,7 +139,7 @@ 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, @@ -149,7 +149,7 @@ 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, @@ -161,7 +161,7 @@ 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, @@ -212,7 +212,7 @@ 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 cc87e3e9..0fa1b7b1 100644 --- a/src/routines/level3/xherk.cpp +++ b/src/routines/level3/xherk.cpp @@ -111,7 +111,7 @@ 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, @@ -121,7 +121,7 @@ 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, @@ -133,7 +133,7 @@ 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, @@ -170,7 +170,7 @@ 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 18a1eac7..5a90a5a2 100644 --- a/src/routines/level3/xsyr2k.cpp +++ b/src/routines/level3/xsyr2k.cpp @@ -106,7 +106,7 @@ 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, @@ -116,7 +116,7 @@ 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, @@ -128,7 +128,7 @@ 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, @@ -177,7 +177,7 @@ 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 1992cec1..46b96b76 100644 --- a/src/routines/level3/xsyrk.cpp +++ b/src/routines/level3/xsyrk.cpp @@ -99,7 +99,7 @@ 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, @@ -111,7 +111,7 @@ 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, @@ -148,7 +148,7 @@ 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, |