Added API and tests for new GemmStridedBatched routine

6 files changed, 635 insertions, 0 deletions

diff --git a/src/clblast.cpp b/src/clblast.cpp
index f5e2f1be..c4c51538 100644
--- a/src/clblast.cpp
+++ b/src/clblast.cpp
@@ -2336,6 +2336,77 @@ template StatusCode PUBLIC_API GemmBatched<half>(const Layout, const Transpose,
                                                  const size_t,
                                                  cl_command_queue*, cl_event*);
 
+// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED
+template <typename T>
+StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
+                              const size_t m, const size_t n, const size_t k,
+                              const T alpha,
+                              const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                              const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                              const T beta,
+                              cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                              const size_t batch_count,
+                              cl_command_queue* queue, cl_event* event) {
+  try {
+    auto queue_cpp = Queue(*queue);
+    auto routine = XgemmStridedBatched<T>(queue_cpp, event);
+    routine.DoGemmStridedBatched(layout, a_transpose, b_transpose,
+                                 m, n, k,
+                                 alpha,
+                                 Buffer<T>(a_buffer), a_offset, a_ld, a_stride,
+                                 Buffer<T>(b_buffer), b_offset, b_ld, b_stride,
+                                 beta,
+                                 Buffer<T>(c_buffer), c_offset, c_ld, c_stride,
+                                 batch_count);
+    return StatusCode::kSuccess;
+  } catch (...) { return DispatchException(); }
+}
+template StatusCode PUBLIC_API GemmStridedBatched<float>(const Layout, const Transpose, const Transpose,
+                                                         const size_t, const size_t, const size_t,
+                                                         const float,
+                                                         const cl_mem, const size_t, const size_t, const size_t,
+                                                         const cl_mem, const size_t, const size_t, const size_t,
+                                                         const float,
+                                                         cl_mem, const size_t, const size_t, const size_t,
+                                                         const size_t,
+                                                         cl_command_queue*, cl_event*);
+template StatusCode PUBLIC_API GemmStridedBatched<double>(const Layout, const Transpose, const Transpose,
+                                                          const size_t, const size_t, const size_t,
+                                                          const double,
+                                                          const cl_mem, const size_t, const size_t, const size_t,
+                                                          const cl_mem, const size_t, const size_t, const size_t,
+                                                          const double,
+                                                          cl_mem, const size_t, const size_t, const size_t,
+                                                          const size_t,
+                                                          cl_command_queue*, cl_event*);
+template StatusCode PUBLIC_API GemmStridedBatched<float2>(const Layout, const Transpose, const Transpose,
+                                                          const size_t, const size_t, const size_t,
+                                                          const float2,
+                                                          const cl_mem, const size_t, const size_t, const size_t,
+                                                          const cl_mem, const size_t, const size_t, const size_t,
+                                                          const float2,
+                                                          cl_mem, const size_t, const size_t, const size_t,
+                                                          const size_t,
+                                                          cl_command_queue*, cl_event*);
+template StatusCode PUBLIC_API GemmStridedBatched<double2>(const Layout, const Transpose, const Transpose,
+                                                           const size_t, const size_t, const size_t,
+                                                           const double2,
+                                                           const cl_mem, const size_t, const size_t, const size_t,
+                                                           const cl_mem, const size_t, const size_t, const size_t,
+                                                           const double2,
+                                                           cl_mem, const size_t, const size_t, const size_t,
+                                                           const size_t,
+                                                           cl_command_queue*, cl_event*);
+template StatusCode PUBLIC_API GemmStridedBatched<half>(const Layout, const Transpose, const Transpose,
+                                                        const size_t, const size_t, const size_t,
+                                                        const half,
+                                                        const cl_mem, const size_t, const size_t, const size_t,
+                                                        const cl_mem, const size_t, const size_t, const size_t,
+                                                        const half,
+                                                        cl_mem, const size_t, const size_t, const size_t,
+                                                        const size_t,
+                                                        cl_command_queue*, cl_event*);
+
 // =================================================================================================
 
 // Retrieves the required size of the temporary buffer for the GEMM kernel (optional)
diff --git a/src/clblast_c.cpp b/src/clblast_c.cpp
index 24697779..aa52cbca 100644
--- a/src/clblast_c.cpp
+++ b/src/clblast_c.cpp
@@ -3846,6 +3846,133 @@ CLBlastStatusCode CLBlastHgemmBatched(const CLBlastLayout layout, const CLBlastT
   } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); }
 }
 
+// GEMM
+CLBlastStatusCode CLBlastSgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose,
+                                             const size_t m, const size_t n, const size_t k,
+                                             const float alpha,
+                                             const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                             const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                                             const float beta,
+                                             cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                             const size_t batch_count,
+                                             cl_command_queue* queue, cl_event* event) {
+  try {
+    return static_cast<CLBlastStatusCode>(
+      clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout),
+                                  static_cast<clblast::Transpose>(a_transpose),
+                                  static_cast<clblast::Transpose>(b_transpose),
+                                  m, n, k,
+                                  alpha,
+                                  a_buffer, a_offset, a_ld, a_stride,
+                                  b_buffer, b_offset, b_ld, b_stride,
+                                  beta,
+                                  c_buffer, c_offset, c_ld, c_stride,
+                                  batch_count,
+                                  queue, event)
+    );
+  } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); }
+}
+CLBlastStatusCode CLBlastDgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose,
+                                             const size_t m, const size_t n, const size_t k,
+                                             const double alpha,
+                                             const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                             const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                                             const double beta,
+                                             cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                             const size_t batch_count,
+                                             cl_command_queue* queue, cl_event* event) {
+  try {
+    return static_cast<CLBlastStatusCode>(
+      clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout),
+                                  static_cast<clblast::Transpose>(a_transpose),
+                                  static_cast<clblast::Transpose>(b_transpose),
+                                  m, n, k,
+                                  alpha,
+                                  a_buffer, a_offset, a_ld, a_stride,
+                                  b_buffer, b_offset, b_ld, b_stride,
+                                  beta,
+                                  c_buffer, c_offset, c_ld, c_stride,
+                                  batch_count,
+                                  queue, event)
+    );
+  } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); }
+}
+CLBlastStatusCode CLBlastCgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose,
+                                             const size_t m, const size_t n, const size_t k,
+                                             const cl_float2 alpha,
+                                             const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                             const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                                             const cl_float2 beta,
+                                             cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                             const size_t batch_count,
+                                             cl_command_queue* queue, cl_event* event) {
+  try {
+    return static_cast<CLBlastStatusCode>(
+      clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout),
+                                  static_cast<clblast::Transpose>(a_transpose),
+                                  static_cast<clblast::Transpose>(b_transpose),
+                                  m, n, k,
+                                  float2{alpha.s[0], alpha.s[1]},
+                                  a_buffer, a_offset, a_ld, a_stride,
+                                  b_buffer, b_offset, b_ld, b_stride,
+                                  float2{beta.s[0], beta.s[1]},
+                                  c_buffer, c_offset, c_ld, c_stride,
+                                  batch_count,
+                                  queue, event)
+    );
+  } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); }
+}
+CLBlastStatusCode CLBlastZgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose,
+                                             const size_t m, const size_t n, const size_t k,
+                                             const cl_double2 alpha,
+                                             const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                             const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                                             const cl_double2 beta,
+                                             cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                             const size_t batch_count,
+                                             cl_command_queue* queue, cl_event* event) {
+  try {
+    return static_cast<CLBlastStatusCode>(
+      clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout),
+                                  static_cast<clblast::Transpose>(a_transpose),
+                                  static_cast<clblast::Transpose>(b_transpose),
+                                  m, n, k,
+                                  double2{alpha.s[0], alpha.s[1]},
+                                  a_buffer, a_offset, a_ld, a_stride,
+                                  b_buffer, b_offset, b_ld, b_stride,
+                                  double2{beta.s[0], beta.s[1]},
+                                  c_buffer, c_offset, c_ld, c_stride,
+                                  batch_count,
+                                  queue, event)
+    );
+  } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); }
+}
+CLBlastStatusCode CLBlastHgemmStridedBatched(const CLBlastLayout layout, const CLBlastTranspose a_transpose, const CLBlastTranspose b_transpose,
+                                             const size_t m, const size_t n, const size_t k,
+                                             const cl_half alpha,
+                                             const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                             const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                                             const cl_half beta,
+                                             cl_mem c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                             const size_t batch_count,
+                                             cl_command_queue* queue, cl_event* event) {
+  try {
+    return static_cast<CLBlastStatusCode>(
+      clblast::GemmStridedBatched(static_cast<clblast::Layout>(layout),
+                                  static_cast<clblast::Transpose>(a_transpose),
+                                  static_cast<clblast::Transpose>(b_transpose),
+                                  m, n, k,
+                                  alpha,
+                                  a_buffer, a_offset, a_ld, a_stride,
+                                  b_buffer, b_offset, b_ld, b_stride,
+                                  beta,
+                                  c_buffer, c_offset, c_ld, c_stride,
+                                  batch_count,
+                                  queue, event)
+    );
+  } catch (...) { return static_cast<CLBlastStatusCode>(clblast::DispatchExceptionForC()); }
+}
+
 // =================================================================================================
 
 // Clears the cache of stored binaries
diff --git a/src/clblast_cuda.cpp b/src/clblast_cuda.cpp
index 348ff3f5..0aa57087 100644
--- a/src/clblast_cuda.cpp
+++ b/src/clblast_cuda.cpp
@@ -2436,6 +2436,79 @@ template StatusCode PUBLIC_API GemmBatched<half>(const Layout, const Transpose,
                                                  const size_t,
                                                  const CUcontext, const CUdevice);
 
+// StridedBatched version of GEMM: SGEMMSTRIDEDBATCHED/DGEMMSTRIDEDBATCHED/CGEMMSTRIDEDBATCHED/ZGEMMSTRIDEDBATCHED/HGEMMSTRIDEDBATCHED
+template <typename T>
+StatusCode GemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
+                              const size_t m, const size_t n, const size_t k,
+                              const T alpha,
+                              const CUdeviceptr a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                              const CUdeviceptr b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride,
+                              const T beta,
+                              CUdeviceptr c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                              const size_t batch_count,
+                              const CUcontext context, const CUdevice device) {
+  try {
+    const auto context_cpp = Context(context);
+    const auto device_cpp = Device(device);
+    auto queue_cpp = Queue(context_cpp, device_cpp);
+    auto routine = XgemmStridedBatched<T>(queue_cpp, nullptr);
+    routine.DoGemmStridedBatched(layout, a_transpose, b_transpose,
+                                 m, n, k,
+                                 alpha,
+                                 Buffer<T>(a_buffer), a_offset, a_ld, a_stride,
+                                 Buffer<T>(b_buffer), b_offset, b_ld, b_stride,
+                                 beta,
+                                 Buffer<T>(c_buffer), c_offset, c_ld, c_stride,
+                                 batch_count);
+    return StatusCode::kSuccess;
+  } catch (...) { return DispatchException(); }
+}
+template StatusCode PUBLIC_API GemmStridedBatched<float>(const Layout, const Transpose, const Transpose,
+                                                         const size_t, const size_t, const size_t,
+                                                         const float,
+                                                         const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                         const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                         const float,
+                                                         CUdeviceptr, const size_t, const size_t, const size_t,
+                                                         const size_t,
+                                                         const CUcontext, const CUdevice);
+template StatusCode PUBLIC_API GemmStridedBatched<double>(const Layout, const Transpose, const Transpose,
+                                                          const size_t, const size_t, const size_t,
+                                                          const double,
+                                                          const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                          const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                          const double,
+                                                          CUdeviceptr, const size_t, const size_t, const size_t,
+                                                          const size_t,
+                                                          const CUcontext, const CUdevice);
+template StatusCode PUBLIC_API GemmStridedBatched<float2>(const Layout, const Transpose, const Transpose,
+                                                          const size_t, const size_t, const size_t,
+                                                          const float2,
+                                                          const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                          const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                          const float2,
+                                                          CUdeviceptr, const size_t, const size_t, const size_t,
+                                                          const size_t,
+                                                          const CUcontext, const CUdevice);
+template StatusCode PUBLIC_API GemmStridedBatched<double2>(const Layout, const Transpose, const Transpose,
+                                                           const size_t, const size_t, const size_t,
+                                                           const double2,
+                                                           const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                           const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                           const double2,
+                                                           CUdeviceptr, const size_t, const size_t, const size_t,
+                                                           const size_t,
+                                                           const CUcontext, const CUdevice);
+template StatusCode PUBLIC_API GemmStridedBatched<half>(const Layout, const Transpose, const Transpose,
+                                                        const size_t, const size_t, const size_t,
+                                                        const half,
+                                                        const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                        const CUdeviceptr, const size_t, const size_t, const size_t,
+                                                        const half,
+                                                        CUdeviceptr, const size_t, const size_t, const size_t,
+                                                        const size_t,
+                                                        const CUcontext, const CUdevice);
+
 // =================================================================================================
 
 // Retrieves the required size of the temporary buffer for the GEMM kernel (optional)
diff --git a/src/routines/levelx/xgemmstridedbatched.cpp b/src/routines/levelx/xgemmstridedbatched.cpp
new file mode 100644
index 00000000..3ea52980
--- /dev/null
+++ b/src/routines/levelx/xgemmstridedbatched.cpp
@@ -0,0 +1,297 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file implements the XgemmStridedBatched class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "routines/levelx/xgemmstridedbatched.hpp"
+#include "routines/level3/xgemm.hpp"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+XgemmStridedBatched<T>::XgemmStridedBatched(Queue &queue, EventPointer event, const std::string &name):
+    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm","XgemmDirect","GemmRoutine"},
+        PrecisionValue<T>(), {}, {
+            #include "../../kernels/level3/level3.opencl"
+            #include "../../kernels/level3/copy_fast.opencl"
+            #include "../../kernels/level3/copy_pad.opencl"
+            #include "../../kernels/level3/transpose_fast.opencl"
+            #include "../../kernels/level3/transpose_pad.opencl"
+            , // separated in multiple parts to prevent C1091 in MSVC 2013
+            #include "../../kernels/level3/xgemm_direct_part1.opencl"
+            #include "../../kernels/level3/xgemm_direct_part2.opencl"
+            #include "../../kernels/level3/xgemm_direct_part3.opencl"
+            , // separated in multiple parts to prevent C1091 in MSVC 2013
+            #include "../../kernels/level3/xgemm_part1.opencl"
+            #include "../../kernels/level3/xgemm_part2.opencl"
+            #include "../../kernels/level3/xgemm_part3.opencl"
+            #include "../../kernels/level3/xgemm_part4.opencl"
+            , // separated in multiple parts to prevent C1091 in MSVC 2013
+            #include "../../kernels/level3/xgemm_batched.opencl"
+            #include "../../kernels/level3/xgemm_direct_batched.opencl"
+        }) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+void XgemmStridedBatched<T>::DoGemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
+                                                  const size_t m, const size_t n, const size_t k, const T alpha,
+                                                  const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                                  const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta,
+                                                  const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                                  const size_t batch_count) {
+
+  // Tests for a valid batch count
+  if (batch_count < 1) {
+    throw BLASError(StatusCode::kInvalidBatchCount);
+  }
+
+  // Computes the transpose/conjugate options and sets the a/b/c sizes based on that
+  bool a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate;
+  size_t a_one, a_two, b_one, b_two, c_one, c_two;
+  Xgemm<T>::ProcessArguments(layout, a_transpose, b_transpose, m, n, k,
+                             a_one, a_two, b_one, b_two, c_one, c_two,
+                             a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate);
+
+  // Tests the matrices for validity
+  for (auto batch = size_t{0}; batch < batch_count; ++batch) {
+    TestMatrixA(a_one, a_two, a_buffer, a_offset + a_stride * batch, a_ld);
+    TestMatrixB(b_one, b_two, b_buffer, b_offset + b_stride * batch, b_ld);
+    TestMatrixC(c_one, c_two, c_buffer, c_offset + c_stride * batch, c_ld);
+  }
+
+  // Selects which version of the batched GEMM to run
+  const auto do_gemm_direct = true;
+  if (do_gemm_direct) { // single generic kernel
+    BatchedGemmDirect(m, n, k, alpha,
+                      a_buffer, a_offset, a_ld, a_stride,
+                      b_buffer, b_offset, b_ld, b_stride, beta,
+                      c_buffer, c_offset, c_ld, c_stride,
+                      a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate,
+                      batch_count);
+  }
+  else { // pre/post-processing plus a very fast kernel
+    BatchedGemmIndirect(m, n, k, alpha,
+                        a_buffer, a_offset, a_ld, a_stride,
+                        b_buffer, b_offset, b_ld, b_stride, beta,
+                        c_buffer, c_offset, c_ld, c_stride,
+                        a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate,
+                        a_one, a_two, b_one, b_two, c_one, c_two, batch_count);
+  }
+}
+
+
+// =================================================================================================
+
+// The indirect version of batched GEMM. This uses the faster but non-general kernel. It has specific
+// requirements, but several pre and post-processing kernels take care of those. However, the
+// overhead of these extra kernels might not be ideal for certain devices/arguments.
+template <typename T>
+void XgemmStridedBatched<T>::BatchedGemmIndirect(const size_t m, const size_t n, const size_t k, const T alpha,
+                                                 const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                                 const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta,
+                                                 const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                                 const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
+                                                 const bool a_conjugate, const bool b_conjugate,
+                                                 const size_t a_one, const size_t a_two,
+                                                 const size_t b_one, const size_t b_two,
+                                                 const size_t c_one, const size_t c_two,
+                                                 const size_t batch_count) {
+  // Calculates the ceiled versions of m, n, and k
+  const auto m_ceiled = Ceil(Ceil(m, db_["MWG"]), db_["VWM"]);
+  const auto n_ceiled = Ceil(Ceil(n, db_["NWG"]), db_["VWN"]);
+  const auto k_ceiled = Ceil(Ceil(k, db_["KWG"]), db_["VWM"]);
+
+  // 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.
+  size_t a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i;
+  Xgemm<T>::CalculateInternalDimensions(m, n, k, db_["MWG"], db_["NWG"], db_["KWG"],
+                                        a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i);
+
+  /* TODO
+  // Sets the "internal" offsets, i.e. the perfect offsets
+  auto a_offsets_i = 0;//std::vector<int>(batch_count);
+  auto b_offsets_i = 0;//std::vector<int>(batch_count);
+  auto c_offsets_i = 0;//std::vector<int>(batch_count);
+
+  // Determines whether or not temporary matrices are needed
+  auto a_no_temp = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && a_offsets == a_offsets_i &&
+                   !a_do_transpose && !a_conjugate;
+  auto b_no_temp = b_one == b_one_i && b_two == b_two_i && b_ld == b_one && b_offsets == b_offsets_i &&
+                   !b_do_transpose && !b_conjugate;
+  auto c_no_temp = c_one == c_one_i && c_two == c_two_i && c_ld == c_one && c_offsets == c_offsets_i &&
+                   !c_do_transpose;
+
+  // Creates the temporary matrices
+  const auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, batch_count * a_one_i * a_two_i);
+  const auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, batch_count * b_one_i * b_two_i);
+  const auto c_temp = (c_no_temp) ? c_buffer : Buffer<T>(context_, batch_count * c_one_i * c_two_i);
+
+  // Events of all kernels (including pre/post processing kernels)
+  auto eventWaitList = std::vector<Event>();
+  auto emptyEventList = std::vector<Event>();
+
+  // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros
+  // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In
+  // case nothing has to be done, these kernels can be skipped.
+  if (!a_no_temp) {
+    auto a_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
+    auto a_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
+    a_offsets_device.Write(queue_, batch_count, a_offsets);
+    a_offsets_i_device.Write(queue_, batch_count, a_offsets_i);
+    auto eventProcessA = Event();
+    PadCopyTransposeMatrixBatched(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
+                                  a_one, a_two, a_ld, a_offsets_device, a_buffer,
+                                  a_one_i, a_two_i, a_one_i, a_offsets_i_device, a_temp,
+                                  program_, true, a_do_transpose, a_conjugate, batch_count);
+    eventWaitList.push_back(eventProcessA);
+  }
+
+  // As above, but now for matrix B
+  if (!b_no_temp) {
+    auto b_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
+    auto b_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
+    b_offsets_device.Write(queue_, batch_count, b_offsets);
+    b_offsets_i_device.Write(queue_, batch_count, b_offsets_i);
+    auto eventProcessB = Event();
+    PadCopyTransposeMatrixBatched(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
+                                  b_one, b_two, b_ld, b_offsets_device, b_buffer,
+                                  b_one_i, b_two_i, b_one_i, b_offsets_i_device, b_temp,
+                                  program_, true, b_do_transpose, b_conjugate, batch_count);
+    eventWaitList.push_back(eventProcessB);
+  }
+
+  // As above, but now for matrix C
+  auto c_offsets_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
+  auto c_offsets_i_device = Buffer<int>(context_, BufferAccess::kReadWrite, batch_count);
+  if (!c_no_temp) {
+    c_offsets_device.Write(queue_, batch_count, c_offsets);
+    c_offsets_i_device.Write(queue_, batch_count, c_offsets_i);
+    auto eventProcessC = Event();
+    PadCopyTransposeMatrixBatched(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
+                                  c_one, c_two, c_ld, c_offsets_device, c_buffer,
+                                  c_one_i, c_two_i, c_one_i, c_offsets_i_device, c_temp,
+                                  program_, true, c_do_transpose, false, batch_count);
+    eventWaitList.push_back(eventProcessC);
+  }
+
+  // Retrieves the Xgemm kernel from the compiled binary
+  auto kernel = Kernel(program_, "XgemmStridedBatched");
+
+  // Sets the kernel arguments
+  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);
+  kernel.SetArgument(4, beta);
+  kernel.SetArgument(5, a_temp());
+  kernel.SetArgument(6, static_cast<int>(a_one_i));
+  kernel.SetArgument(7, static_cast<int>(a_two_i));
+  kernel.SetArgument(8, b_temp());
+  kernel.SetArgument(9, static_cast<int>(b_one_i));
+  kernel.SetArgument(10, static_cast<int>(b_two_i));
+  kernel.SetArgument(11, c_temp());
+  kernel.SetArgument(12, static_cast<int>(c_one_i));
+  kernel.SetArgument(13, static_cast<int>(c_two_i));
+
+  // Computes the global and local thread sizes
+  const auto global = std::vector<size_t>{
+      (c_one_i * db_["MDIMC"]) / db_["MWG"],
+      (c_two_i * db_["NDIMC"]) / db_["NWG"],
+      batch_count
+  };
+  const auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"], 1};
+
+  // Launches the kernel
+  auto eventKernel = Event();
+  auto eventPointer = eventKernel.pointer();
+  RunKernel(kernel, queue_, device_, global, local, eventPointer, eventWaitList);
+
+  // Runs the post-processing kernel if needed
+  if (!c_no_temp) {
+    eventWaitList.push_back(eventKernel);
+    PadCopyTransposeMatrixBatched(queue_, device_, db_, event_, eventWaitList,
+                                  c_one_i, c_two_i, c_one_i, c_offsets_i_device, c_temp,
+                                  c_one, c_two, c_ld, c_offsets_device, c_buffer,
+                                  program_, false, c_do_transpose, false, batch_count);
+  }
+  */
+}
+
+// =================================================================================================
+
+// The direct version of batched GEMM, requiring just one kernel, no pre or post-processing kernels.
+template <typename T>
+void XgemmStridedBatched<T>::BatchedGemmDirect(const size_t m, const size_t n, const size_t k, const T alpha,
+                                               const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                                               const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta,
+                                               const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                                               const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
+                                               const bool a_conjugate, const bool b_conjugate,
+                                               const size_t batch_count) {
+/* TODO
+  // Retrieves the proper XgemmDirect kernel from the compiled binary
+  const auto name = (a_do_transpose) ? (b_do_transpose ? "XgemmDirectBatchedTT" : "XgemmDirectBatchedTN") :
+                    (b_do_transpose ? "XgemmDirectBatchedNT" : "XgemmDirectBatchedNN");
+  auto kernel = Kernel(program_, name);
+
+  // Sets the kernel arguments
+  kernel.SetArgument(0, static_cast<int>(m));
+  kernel.SetArgument(1, static_cast<int>(n));
+  kernel.SetArgument(2, static_cast<int>(k));
+  kernel.SetArgument(3, alpha);
+  kernel.SetArgument(4, beta);
+  kernel.SetArgument(5, a_buffer());
+  kernel.SetArgument(6, a_offset);
+  kernel.SetArgument(7, static_cast<int>(a_ld));
+  kernel.SetArgument(8, b_buffer());
+  kernel.SetArgument(9, b_offset);
+  kernel.SetArgument(10, static_cast<int>(b_ld));
+  kernel.SetArgument(11, c_buffer());
+  kernel.SetArgument(12, c_offset);
+  kernel.SetArgument(13, static_cast<int>(c_ld));
+  kernel.SetArgument(14, static_cast<int>(c_do_transpose));
+  kernel.SetArgument(15, static_cast<int>(a_conjugate));
+  kernel.SetArgument(16, static_cast<int>(b_conjugate));
+
+  // Computes the global and local thread sizes
+  const auto m_ceiled = Ceil(m, db_["WGD"]);
+  const auto n_ceiled = Ceil(n, db_["WGD"]);
+  const auto global = std::vector<size_t>{
+      (m_ceiled * db_["MDIMCD"]) / db_["WGD"],
+      (n_ceiled * db_["NDIMCD"]) / db_["WGD"],
+      batch_count
+  };
+  const auto local = std::vector<size_t>{db_["MDIMCD"], db_["NDIMCD"], 1};
+
+  // Launches the kernel
+  RunKernel(kernel, queue_, device_, global, local, event_);
+  */
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class XgemmStridedBatched<half>;
+template class XgemmStridedBatched<float>;
+template class XgemmStridedBatched<double>;
+template class XgemmStridedBatched<float2>;
+template class XgemmStridedBatched<double2>;
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/routines/levelx/xgemmstridedbatched.hpp b/src/routines/levelx/xgemmstridedbatched.hpp
new file mode 100644
index 00000000..0dbbcb10
--- /dev/null
+++ b/src/routines/levelx/xgemmstridedbatched.hpp
@@ -0,0 +1,66 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file implements the XgemmStridedBatched routine. This is a non-blas batched version of GEMM.
+//
+// =================================================================================================
+
+#ifndef CLBLAST_ROUTINES_XGEMMSTRIDEDBATCHED_H_
+#define CLBLAST_ROUTINES_XGEMMSTRIDEDBATCHED_H_
+
+#include <vector>
+
+#include "routine.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// See comment at top of file for a description of the class
+template <typename T>
+class XgemmStridedBatched: public Routine {
+public:
+
+  // Constructor
+  XgemmStridedBatched(Queue &queue, EventPointer event, const std::string &name = "GEMMSTRIDEDBATCHED");
+
+  // Templated-precision implementation of the routine
+  void DoGemmStridedBatched(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
+                            const size_t m, const size_t n, const size_t k, const T alpha,
+                            const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                            const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta,
+                            const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                            const size_t batch_count);
+
+  // Indirect version of strided batched GEMM (with pre and post-processing kernels)
+  void BatchedGemmIndirect(const size_t m, const size_t n, const size_t k, const T alpha,
+                           const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                           const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta,
+                           const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                           const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
+                           const bool a_conjugate, const bool b_conjugate,
+                           const size_t a_one, const size_t a_two,
+                           const size_t b_one, const size_t b_two,
+                           const size_t c_one, const size_t c_two,
+                           const size_t batch_count);
+
+  // Direct version of strided batched GEMM (no pre and post-processing kernels)
+  void BatchedGemmDirect(const size_t m, const size_t n, const size_t k, const T alpha,
+                         const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, const size_t a_stride,
+                         const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld, const size_t b_stride, const T beta,
+                         const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld, const size_t c_stride,
+                         const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
+                         const bool a_conjugate, const bool b_conjugate,
+                         const size_t batch_count);
+};
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_ROUTINES_XGEMMSTRIDEDBATCHED_H_
+#endif
diff --git a/src/routines/routines.hpp b/src/routines/routines.hpp
index 9e7768b9..0aeff707 100644
--- a/src/routines/routines.hpp
+++ b/src/routines/routines.hpp
@@ -71,6 +71,7 @@
 #include "routines/levelx/xim2col.hpp"
 #include "routines/levelx/xaxpybatched.hpp"
 #include "routines/levelx/xgemmbatched.hpp"
+#include "routines/levelx/xgemmstridedbatched.hpp"
 
 // CLBLAST_ROUTINES_ROUTINES_H_
 #endif