Implemented direct version of strided-batched GEMM kernel

1 files changed, 102 insertions, 20 deletions

diff --git a/src/kernels/level3/xgemm_direct_batched.opencl b/src/kernels/level3/xgemm_direct_batched.opencl
index d946a056..d15ed31e 100644
--- a/src/kernels/level3/xgemm_direct_batched.opencl
+++ b/src/kernels/level3/xgemm_direct_batched.opencl
@@ -17,15 +17,16 @@
 R"(
 
 // =================================================================================================
+#if defined(ROUTINE_GEMMBATCHED)
 
 // Direct version of the batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
 __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
-                                   const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
-                                   const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
-                                   const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
-                                   __global real* cgm, const __constant int* c_offsets, const int c_ld,
-                                   const int c_transpose, const int a_conjugate, const int b_conjugate) {
+                          const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
+                          const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
+                          const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
+                          __global real* cgm, const __constant int* c_offsets, const int c_ld,
+                          const int c_transpose, const int a_conjugate, const int b_conjugate) {
   const int batch = get_group_id(2);
   const real_arg arg_alpha = arg_alphas[batch];
   const real_arg arg_beta = arg_betas[batch];
@@ -42,11 +43,11 @@ void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the batched GEMM kernel with [A, B] = [non-transposed, transposed]
 __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
-                                   const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
-                                   const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
-                                   const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
-                                   __global real* cgm, const __constant int* c_offsets, const int c_ld,
-                                   const int c_transpose, const int a_conjugate, const int b_conjugate) {
+                          const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
+                          const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
+                          const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
+                          __global real* cgm, const __constant int* c_offsets, const int c_ld,
+                          const int c_transpose, const int a_conjugate, const int b_conjugate) {
   const int batch = get_group_id(2);
   const real_arg arg_alpha = arg_alphas[batch];
   const real_arg arg_beta = arg_betas[batch];
@@ -63,11 +64,11 @@ void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the batched GEMM kernel with [A, B] = [transposed, non-transposed]
 __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
-                                   const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
-                                   const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
-                                   const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
-                                   __global real* cgm, const __constant int* c_offsets, const int c_ld,
-                                   const int c_transpose, const int a_conjugate, const int b_conjugate) {
+                          const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
+                          const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
+                          const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
+                          __global real* cgm, const __constant int* c_offsets, const int c_ld,
+                          const int c_transpose, const int a_conjugate, const int b_conjugate) {
   const int batch = get_group_id(2);
   const real_arg arg_alpha = arg_alphas[batch];
   const real_arg arg_beta = arg_betas[batch];
@@ -84,11 +85,11 @@ void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the batched GEMM kernel with [A, B] = [transposed, transposed]
 __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
-                                   const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
-                                   const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
-                                   const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
-                                   __global real* cgm, const __constant int* c_offsets, const int c_ld,
-                                   const int c_transpose, const int a_conjugate, const int b_conjugate) {
+                          const __constant real_arg* arg_alphas, const __constant real_arg* arg_betas,
+                          const __global realMD* restrict agm, const __constant int* a_offsets, const int a_ld,
+                          const __global realND* restrict bgm, const __constant int* b_offsets, const int b_ld,
+                          __global real* cgm, const __constant int* c_offsets, const int c_ld,
+                          const int c_transpose, const int a_conjugate, const int b_conjugate) {
   const int batch = get_group_id(2);
   const real_arg arg_alpha = arg_alphas[batch];
   const real_arg arg_beta = arg_betas[batch];
@@ -102,6 +103,87 @@ void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
               alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate);
 }
 
+#endif
+// =================================================================================================
+#if defined(ROUTINE_GEMMSTRIDEDBATCHED)
+
+// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
+__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
+void XgemmDirectStridedBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
+                                 const real_arg arg_alpha, const real_arg arg_beta,
+                                 const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride,
+                                 const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
+                                 __global real* cgm, const int c_offset, const int c_ld, const int c_stride,
+                                 const int c_transpose, const int a_conjugate, const int b_conjugate) {
+  const int batch = get_group_id(2);
+  const int a_offset_batch = a_offset + a_stride * batch;
+  const int b_offset_batch = b_offset + b_stride * batch;
+  const int c_offset_batch = c_offset + c_stride * batch;
+  __local real alm[WGD * (WGD + PADA)];
+  __local real blm[WGD * (WGD + PADB)];
+  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
+              agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
+              alm, blm, 0, 0, c_transpose, a_conjugate, b_conjugate);
+}
+
+// Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, transposed]
+__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
+void XgemmDirectStridedBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
+                                 const real_arg arg_alpha, const real_arg arg_beta,
+                                 const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride,
+                                 const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
+                                 __global real* cgm, const int c_offset, const int c_ld, const int c_stride,
+                                 const int c_transpose, const int a_conjugate, const int b_conjugate) {
+  const int batch = get_group_id(2);
+  const int a_offset_batch = a_offset + a_stride * batch;
+  const int b_offset_batch = b_offset + b_stride * batch;
+  const int c_offset_batch = c_offset + c_stride * batch;
+  __local real alm[WGD * (WGD + PADA)];
+  __local real blm[WGD * (WGD + PADB)];
+  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
+              agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
+              alm, blm, 0, 1, c_transpose, a_conjugate, b_conjugate);
+}
+
+// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, non-transposed]
+__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
+void XgemmDirectStridedBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
+                                 const real_arg arg_alpha, const real_arg arg_beta,
+                                 const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride,
+                                 const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
+                                 __global real* cgm, const int c_offset, const int c_ld, const int c_stride,
+                                 const int c_transpose, const int a_conjugate, const int b_conjugate) {
+  const int batch = get_group_id(2);
+  const int a_offset_batch = a_offset + a_stride * batch;
+  const int b_offset_batch = b_offset + b_stride * batch;
+  const int c_offset_batch = c_offset + c_stride * batch;
+  __local real alm[WGD * (WGD + PADA)];
+  __local real blm[WGD * (WGD + PADB)];
+  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
+              agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
+              alm, blm, 1, 0, c_transpose, a_conjugate, b_conjugate);
+}
+
+// Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, transposed]
+__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
+void XgemmDirectStridedBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
+                                 const real_arg arg_alpha, const real_arg arg_beta,
+                                 const __global realMD* restrict agm, const int a_offset, const int a_ld, const int a_stride,
+                                 const __global realND* restrict bgm, const int b_offset, const int b_ld, const int b_stride,
+                                 __global real* cgm, const int c_offset, const int c_ld, const int c_stride,
+                                 const int c_transpose, const int a_conjugate, const int b_conjugate) {
+  const int batch = get_group_id(2);
+  const int a_offset_batch = a_offset + a_stride * batch;
+  const int b_offset_batch = b_offset + b_stride * batch;
+  const int c_offset_batch = c_offset + c_stride * batch;
+  __local real alm[WGD * (WGD + PADA)];
+  __local real blm[WGD * (WGD + PADB)];
+  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
+              agm, a_offset_batch, a_ld, bgm, b_offset_batch, b_ld, cgm, c_offset_batch, c_ld,
+              alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate);
+}
+
+#endif
 // =================================================================================================
 
 // End of the C++11 raw string literal