From c1c4bc5d209280e4ec9be5c0a26f7c94077a6b20 Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Mon, 3 Oct 2016 19:32:01 +0200
Subject: Re-organised GEMM direct kernel and added faster fall-back version
 for incomplete rectangles

---
 src/kernels/level3/xgemm_direct_part2.opencl | 432 ++++++++++++++-------------
 1 file changed, 219 insertions(+), 213 deletions(-)

(limited to 'src/kernels/level3/xgemm_direct_part2.opencl')

diff --git a/src/kernels/level3/xgemm_direct_part2.opencl b/src/kernels/level3/xgemm_direct_part2.opencl
index 0d066186..953a3b3c 100644
--- a/src/kernels/level3/xgemm_direct_part2.opencl
+++ b/src/kernels/level3/xgemm_direct_part2.opencl
@@ -7,7 +7,7 @@
 // Author(s):
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
-// This is part 2 of 2 of the GEMM kernel. See part 1 for more information.
+// This is part 2 of 3 of the GEMM kernel. See part 1 for more information.
 //
 // =================================================================================================
 
@@ -17,179 +17,222 @@ R"(
 
 // =================================================================================================
 
-// Merges the results in Cpm with the global array in Cgm. This also performs the multiplication
-// with the constants: Cgm = alpha*A*B + beta*Cgm = alpha*Cpm + beta*Cgm
-inline void StoreResultsDirect(__global real* cgm, real cpm[NWID][MWID],
-                               const int kSizeM, const int kSizeN,
-                               const real alpha, const real beta,
-                               const int c_ld, const int c_offset, const int c_transpose) {
+// Caches global off-chip memory into local (shared) memory on-chip. This function is specific for
+// caching the A input matrix.
+inline void GlobalToLocalDirectA(const __global realMD* restrict agm, __local real* alm,
+                                 const int a_ld, const int a_offset, const int kwg,
+                                 const int a_transpose, const int a_conjugate) {
+  #if MDIMCD == MDIMAD
+    const int la0 = get_local_id(0);
+    const int la1 = get_local_id(1);
+  #else
+    const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
+    const int la0 = tid % MDIMAD;
+    const int la1 = tid / MDIMAD;
+  #endif
   #pragma unroll
-  for (int ni=0; ni<NWID; ++ni) {
+  for (int mia=0; mia<MWAD/VWMD; ++mia) {
     #pragma unroll
-    for (int mi=0; mi<MWID; ++mi) {
-      int mg = mi + get_local_id(0)*MWID;
-      int ng = ni + get_local_id(1)*NWID;
-      int idm = mg + GetGroupID0() * WGD;
-      int idn = ng + GetGroupID1() * WGD;
-
-      // Determines the destination index
-      const int c_index = (c_transpose) ? idm*c_ld + idn : idn*c_ld + idm;
-
-      // The final multiplication with alpha (in case beta == 0)
-      real result;
-      if (IsZero(beta)) {
-        Multiply(result, alpha, cpm[ni][mi]);
-      }
-      // The final multiplication with alpha and the addition with beta*C
-      else {
-        AXPBY(result, alpha, cpm[ni][mi], beta, cgm[c_index + c_offset]);
+    for (int kia=0; kia<KWAD; ++kia) {
+
+      // Computes the indices for the global memory
+      int mg = mia + la0*(MWAD/VWMD);
+      int kg = kia + la1*KWAD;
+      int idm = (a_transpose) ? mg + kwg/VWMD : mg + GetGroupID0()*(WGD/VWMD);
+      int idk = (a_transpose) ? kg + GetGroupID0()*WGD : kg + kwg;
+
+      // Loads the data from global memory into the local memory
+      const realMD avec = agm[idk*(a_ld/VWMD) + idm + a_offset];
+      #if VWMD == 1
+         alm[kg*(WGD + PADA) + mg] = avec;
+      #elif VWMD == 2
+         alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.x;
+         alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.y;
+      #elif VWMD == 4
+         alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.x;
+         alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.y;
+         alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.z;
+         alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.w;
+      #elif VWMD == 8
+         alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.s0;
+         alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.s1;
+         alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.s2;
+         alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.s3;
+         alm[kg*(WGD + PADA) + mg*VWMD + 4] = avec.s4;
+         alm[kg*(WGD + PADA) + mg*VWMD + 5] = avec.s5;
+         alm[kg*(WGD + PADA) + mg*VWMD + 6] = avec.s6;
+         alm[kg*(WGD + PADA) + mg*VWMD + 7] = avec.s7;
+      #elif VWMD == 16
+         alm[kg*(WGD + PADA) + mg*VWMD + 0] = avec.s0;
+         alm[kg*(WGD + PADA) + mg*VWMD + 1] = avec.s1;
+         alm[kg*(WGD + PADA) + mg*VWMD + 2] = avec.s2;
+         alm[kg*(WGD + PADA) + mg*VWMD + 3] = avec.s3;
+         alm[kg*(WGD + PADA) + mg*VWMD + 4] = avec.s4;
+         alm[kg*(WGD + PADA) + mg*VWMD + 5] = avec.s5;
+         alm[kg*(WGD + PADA) + mg*VWMD + 6] = avec.s6;
+         alm[kg*(WGD + PADA) + mg*VWMD + 7] = avec.s7;
+         alm[kg*(WGD + PADA) + mg*VWMD + 8] = avec.s8;
+         alm[kg*(WGD + PADA) + mg*VWMD + 9] = avec.s9;
+         alm[kg*(WGD + PADA) + mg*VWMD + 10] = avec.sA;
+         alm[kg*(WGD + PADA) + mg*VWMD + 11] = avec.sB;
+         alm[kg*(WGD + PADA) + mg*VWMD + 12] = avec.sC;
+         alm[kg*(WGD + PADA) + mg*VWMD + 13] = avec.sD;
+         alm[kg*(WGD + PADA) + mg*VWMD + 14] = avec.sE;
+         alm[kg*(WGD + PADA) + mg*VWMD + 15] = avec.sF;
+      #endif
+      if (a_conjugate) {
+        for (int vm=0; vm<VWMD; ++vm) {
+          COMPLEX_CONJUGATE(alm[kg*(WGD + PADA) + mg*VWMD + vm]);
+        }
       }
-      cgm[c_index + c_offset] = result;
     }
   }
 }
 
-// =================================================================================================
-
-// Main body of the kernel. This is the direct version without restrictions.
-inline void XgemmDirect(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 __global realND* restrict bgm, const int b_offset, const int b_ld,
-                        __global real* cgm, const int c_offset, const int c_ld,
-                        __local real* alm, __local real* blm,
-                        const int a_transpose, const int b_transpose, const int c_transpose,
-                        const int a_conjugate, const int b_conjugate) {
-  const real alpha = GetRealArg(arg_alpha);
-  const real beta = GetRealArg(arg_beta);
-
-  // Extra pointers to scalar versions of global memory
-  const __global real* restrict agms = (const __global real* restrict) agm;
-  const __global real* restrict bgms = (const __global real* restrict) bgm;
-
-  // Allocates workitem-private memory (registers)
-  real apm[MWID];
-  real bpm[NWID];
-  real cpm[NWID][MWID];
-
-  // Initializes the accumulation registers
-  InitAccRegistersDirect(cpm);
-
-  // The faster version of GEMM is not allowed on the (incomplete) borders. Therefore, this section
-  // processes only the main parts: output blocks of WGD by WGD.
-  const int idm = get_local_id(0) * MWID + GetGroupID0() * WGD;
-  const int idn = get_local_id(1) * NWID + GetGroupID1() * WGD;
-  if ((idm < (kSizeM/WGD)*WGD) && (idn < (kSizeN/WGD)*WGD) &&
-      (a_ld % VWMD == 0) && (b_ld % VWND == 0)) {
-
-    // Loops over all complete workgroup tiles
-    int kwg = 0;
-    for (; kwg < (kSizeK/WGD) * WGD; kwg+=WGD) {
-
-      // Loads data: off-chip --> local (matrix A and B)
-      GlobalToLocalDirectA(agm, alm, a_ld, a_offset, kwg, a_transpose, a_conjugate);
-      GlobalToLocalDirectB(bgm, blm, b_ld, b_offset, kwg, b_transpose, b_conjugate);
-      barrier(CLK_LOCAL_MEM_FENCE);
-
-      // Loops over all workitem tiles, unrolled by a factor KWID
-      for (int pwi=0; pwi<WGD; pwi+=KWID) {
-        #pragma unroll
-        for (int pit=0; pit<KWID; ++pit) {
-          int kg = pwi + pit;
-
-          // Loads data: local --> private (matrix A)
-          LocalToPrivateDirectA(alm, apm, kg, a_transpose);
-
-          // Loads data: local --> private (matrix B)
-          LocalToPrivateDirectB(blm, bpm, kg, b_transpose);
-
-          // Performs the accumulation (Cpm += Apm * Bpm)
-          MultiplyAccumulateDirect(cpm, apm, bpm);
+// Same as above, but now for the B input matrix
+inline void GlobalToLocalDirectB(const __global realND* restrict bgm, __local real* blm,
+                                 const int b_ld, const int b_offset, const int kwg,
+                                 const int b_transpose, const int b_conjugate) {
+  #if MDIMCD == NDIMBD
+    const int lb0 = get_local_id(0);
+    const int lb1 = get_local_id(1);
+  #else
+    const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
+    const int lb0 = tid % NDIMBD;
+    const int lb1 = tid / NDIMBD;
+  #endif
+  #pragma unroll
+  for (int kib=0; kib<KWBD; ++kib) {
+    #pragma unroll
+    for (int nib=0; nib<NWBD/VWND; ++nib) {
+
+      // Computes the indices for the global memory
+      int ng = nib + lb0*(NWBD/VWND);
+      int kg = kib + lb1*KWBD;
+      int idn = (b_transpose) ? ng + kwg/VWND : ng + GetGroupID1()*(WGD/VWND);
+      int idk = (b_transpose) ? kg + GetGroupID1()*WGD : kg + kwg;
+
+      // Loads the data from global memory into the local memory
+      const realND bvec = bgm[idk*(b_ld/VWND) + idn + b_offset];
+      #if VWND == 1
+         blm[kg*(WGD + PADB) + ng] = bvec;
+      #elif VWND == 2
+         blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.x;
+         blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.y;
+      #elif VWND == 4
+         blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.x;
+         blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.y;
+         blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.z;
+         blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.w;
+      #elif VWND == 8
+         blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.s0;
+         blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.s1;
+         blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.s2;
+         blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.s3;
+         blm[kg*(WGD + PADB) + ng*VWND + 4] = bvec.s4;
+         blm[kg*(WGD + PADB) + ng*VWND + 5] = bvec.s5;
+         blm[kg*(WGD + PADB) + ng*VWND + 6] = bvec.s6;
+         blm[kg*(WGD + PADB) + ng*VWND + 7] = bvec.s7;
+      #elif VWND == 16
+         blm[kg*(WGD + PADB) + ng*VWND + 0] = bvec.s0;
+         blm[kg*(WGD + PADB) + ng*VWND + 1] = bvec.s1;
+         blm[kg*(WGD + PADB) + ng*VWND + 2] = bvec.s2;
+         blm[kg*(WGD + PADB) + ng*VWND + 3] = bvec.s3;
+         blm[kg*(WGD + PADB) + ng*VWND + 4] = bvec.s4;
+         blm[kg*(WGD + PADB) + ng*VWND + 5] = bvec.s5;
+         blm[kg*(WGD + PADB) + ng*VWND + 6] = bvec.s6;
+         blm[kg*(WGD + PADB) + ng*VWND + 7] = bvec.s7;
+         blm[kg*(WGD + PADB) + ng*VWND + 8] = bvec.s8;
+         blm[kg*(WGD + PADB) + ng*VWND + 9] = bvec.s9;
+         blm[kg*(WGD + PADB) + ng*VWND + 10] = bvec.sA;
+         blm[kg*(WGD + PADB) + ng*VWND + 11] = bvec.sB;
+         blm[kg*(WGD + PADB) + ng*VWND + 12] = bvec.sC;
+         blm[kg*(WGD + PADB) + ng*VWND + 13] = bvec.sD;
+         blm[kg*(WGD + PADB) + ng*VWND + 14] = bvec.sE;
+         blm[kg*(WGD + PADB) + ng*VWND + 15] = bvec.sF;
+      #endif
+      if (b_conjugate) {
+        for (int vn=0; vn<VWND; ++vn) {
+          COMPLEX_CONJUGATE(blm[kg*(WGD + PADB) + ng*VWND + vn]);
         }
       }
-      barrier(CLK_LOCAL_MEM_FENCE);
-    }
-
-    // Loop over the remaining part (incomplete tile in K-dimension)
-    for (; kwg < kSizeK; ++kwg) {
-      const int idk = kwg;
-
-      // Loads A into register memory
-      #pragma unroll
-      for (int mi=0; mi<MWID; ++mi) {
-        const int a_index = (a_transpose) ? (idm + mi)*a_ld + idk : idk*a_ld + (idm + mi);
-        apm[mi] = agms[a_index + a_offset];
-        if (a_conjugate) { COMPLEX_CONJUGATE(apm[mi]); }
-      }
-
-      // Loads B into register memory
-      #pragma unroll
-      for (int ni=0; ni<NWID; ++ni) {
-        const int b_index = (b_transpose) ? (idn + ni)*b_ld + idk : idk*b_ld + (idn + ni);
-        bpm[ni] = bgms[b_index + b_offset];
-        if (b_conjugate) { COMPLEX_CONJUGATE(bpm[ni]); }
-      }
-
-      // Performs the accumulation (Cpm += Apm * Bpm)
-      MultiplyAccumulateDirect(cpm, apm, bpm);
     }
-
-    // Stores a tile of results and performs the multiplication with alpha and beta
-    StoreResultsDirect(cgm, cpm, kSizeM, kSizeN, alpha, beta, c_ld, c_offset, c_transpose);
   }
+}
 
-  // Simple but slow version for the parts on the edge (incomplete tiles in M and N-dimensions)
-  else {
-
-    // Loop over the K-dimension
-    for (int idk = 0; idk < kSizeK; ++idk) {
-
-      // Loads A into register memory
-      #pragma unroll
-      for (int mi=0; mi<MWID; ++mi) {
-        if (idm + mi < kSizeM) {
-          const int a_index = (a_transpose) ? (idm + mi)*a_ld + idk : idk*a_ld + (idm + mi);
-          apm[mi] = agms[a_index + a_offset];
-          if (a_conjugate) { COMPLEX_CONJUGATE(apm[mi]); }
-        }
-        else {
-          SetToZero(apm[mi]);
-        }
+// Caches global off-chip memory into local (shared) memory on-chip. This function is specific for
+// caching the A input matrix. In contrast to the functions above, this function performs bounds
+// checks and doesn't use the vector data-types.
+inline void GlobalToLocalCheckedA(const __global real* restrict agms, __local real* alm,
+                                  const int a_ld, const int a_offset, const int kwg,
+                                  const int a_transpose, const int a_conjugate,
+                                  const int kSizeM, const int kSizeK) {
+  #if MDIMCD == MDIMAD
+    const int la0 = get_local_id(0);
+    const int la1 = get_local_id(1);
+  #else
+    const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
+    const int la0 = tid % MDIMAD;
+    const int la1 = tid / MDIMAD;
+  #endif
+  #pragma unroll
+  for (int mia=0; mia<MWAD; ++mia) {
+    #pragma unroll
+    for (int kia=0; kia<KWAD; ++kia) {
+
+      // Computes the indices for the global memory
+      int mg = mia + la0*MWAD;
+      int kg = kia + la1*KWAD;
+      int idm = (a_transpose) ? mg + kwg : mg + GetGroupID0()*WGD;
+      int idk = (a_transpose) ? kg + GetGroupID0()*WGD : kg + kwg;
+
+      // Loads the data from global memory into the local memory
+      int condition = (a_transpose) ? idm < kSizeK : idm < kSizeM;
+      if (condition) {
+        const real result = agms[idk*a_ld + idm + a_offset];
+        if (a_conjugate) { COMPLEX_CONJUGATE(result); }
+        alm[kg*(WGD + PADA) + mg] = result;
       }
-
-      // Loads B into register memory
-      #pragma unroll
-      for (int ni=0; ni<NWID; ++ni) {
-        if (idn + ni < kSizeN) {
-          const int b_index = (b_transpose) ? (idn + ni)*b_ld + idk : idk*b_ld + (idn + ni);
-          bpm[ni] = bgms[b_index + b_offset];
-          if (b_conjugate) { COMPLEX_CONJUGATE(bpm[ni]); }
-        }
-        else {
-          SetToZero(bpm[ni]);
-        }
+      else {
+        SetToZero(alm[kg*(WGD + PADA) + mg]);
       }
-
-      // Performs the accumulation (Cpm += Apm * Bpm)
-      MultiplyAccumulateDirect(cpm, apm, bpm);
     }
+  }
+}
 
-    // Stores the results
+// Same as above, but now for the B input matrix
+inline void GlobalToLocalCheckedB(const __global real* restrict bgms, __local real* blm,
+                                  const int b_ld, const int b_offset, const int kwg,
+                                  const int b_transpose, const int b_conjugate,
+                                  const int kSizeN, const int kSizeK) {
+  #if MDIMCD == NDIMBD
+    const int lb0 = get_local_id(0);
+    const int lb1 = get_local_id(1);
+  #else
+    const int tid = get_local_id(0) + MDIMCD*get_local_id(1);
+    const int lb0 = tid % NDIMBD;
+    const int lb1 = tid / NDIMBD;
+  #endif
+  #pragma unroll
+  for (int kib=0; kib<KWBD; ++kib) {
     #pragma unroll
-    for (int ni=0; ni<NWID; ++ni) {
-      #pragma unroll
-      for (int mi=0; mi<MWID; ++mi) {
-        if ((idm + mi) < kSizeM && (idn + ni) < kSizeN) {
-
-          // Determines the destination index
-          const int c_index = (c_transpose) ? (idm + mi)*c_ld + (idn + ni) : (idn + ni)*c_ld + (idm + mi);
-
-          // Computes and stores the result
-          real result;
-          AXPBY(result, alpha, cpm[ni][mi], beta, cgm[c_index + c_offset]);
-          cgm[c_index + c_offset] = result;
-        }
+    for (int nib=0; nib<NWBD; ++nib) {
+
+      // Computes the indices for the global memory
+      int ng = nib + lb0*NWBD;
+      int kg = kib + lb1*KWBD;
+      int idn = (b_transpose) ? ng + kwg : ng + GetGroupID1()*WGD;
+      int idk = (b_transpose) ? kg + GetGroupID1()*WGD : kg + kwg;
+
+      // Loads the data from global memory into the local memory
+      int condition = (b_transpose) ? idn < kSizeK : idn < kSizeN;
+      if (condition) {
+        const real result = bgms[idk*b_ld + idn + b_offset];
+        if (b_conjugate) { COMPLEX_CONJUGATE(result); }
+        blm[kg*(WGD + PADB) + ng] = result;
+      }
+      else {
+        SetToZero(blm[kg*(WGD + PADB) + ng]);
       }
     }
   }
@@ -197,64 +240,27 @@ inline void XgemmDirect(const int kSizeM, const int kSizeN, const int kSizeK,
 
 // =================================================================================================
 
-// Direct version of the GEMM kernel with [A, B] = [non-transposed, non-transposed]
-__attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
-__kernel void XgemmDirectNN(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 __global realND* restrict bgm, const int b_offset, const int b_ld,
-                            __global real* cgm, const int c_offset, const int c_ld,
-                            const int c_transpose, const int a_conjugate, const int b_conjugate) {
-  __local real alm[WGD * (WGD + PADA)];
-  __local real blm[WGD * (WGD + PADB)];
-  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
-              agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
-              alm, blm, 0, 0, c_transpose, a_conjugate, b_conjugate);
-}
-
-// Direct version of the GEMM kernel with [A, B] = [non-transposed, transposed]
-__attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
-__kernel void XgemmDirectNT(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 __global realND* restrict bgm, const int b_offset, const int b_ld,
-                            __global real* cgm, const int c_offset, const int c_ld,
-                            const int c_transpose, const int a_conjugate, const int b_conjugate) {
-  __local real alm[WGD * (WGD + PADA)];
-  __local real blm[WGD * (WGD + PADB)];
-  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
-              agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
-              alm, blm, 0, 1, c_transpose, a_conjugate, b_conjugate);
-}
-
-// Direct version of the GEMM kernel with [A, B] = [transposed, non-transposed]
-__attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
-__kernel void XgemmDirectTN(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 __global realND* restrict bgm, const int b_offset, const int b_ld,
-                            __global real* cgm, const int c_offset, const int c_ld,
-                            const int c_transpose, const int a_conjugate, const int b_conjugate) {
-  __local real alm[WGD * (WGD + PADA)];
-  __local real blm[WGD * (WGD + PADB)];
-  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
-              agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
-              alm, blm, 1, 0, c_transpose, a_conjugate, b_conjugate);
+// Caches on-chip local memory into per-thread private memory (registers). This function is specific
+// for caching the A input matrix.
+inline void LocalToPrivateDirectA(__local real* alm, real apm[MWID], const int kg,
+                                  const int a_transpose) {
+  #pragma unroll
+  for (int mi=0; mi<MWID; ++mi) {
+    const int mg = mi + get_local_id(0)*MWID;
+    const int index = (a_transpose) ? mg*(WGD + PADA) + kg : kg*(WGD + PADA) + mg;
+    apm[mi] = alm[index];
+  }
 }
 
-// Direct version of the GEMM kernel with [A, B] = [transposed, transposed]
-__attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
-__kernel void XgemmDirectTT(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 __global realND* restrict bgm, const int b_offset, const int b_ld,
-                            __global real* cgm, const int c_offset, const int c_ld,
-                            const int c_transpose, const int a_conjugate, const int b_conjugate) {
-  __local real alm[WGD * (WGD + PADA)];
-  __local real blm[WGD * (WGD + PADB)];
-  XgemmDirect(kSizeM, kSizeN, kSizeK, arg_alpha, arg_beta,
-              agm, a_offset, a_ld, bgm, b_offset, b_ld, cgm, c_offset, c_ld,
-              alm, blm, 1, 1, c_transpose, a_conjugate, b_conjugate);
+// Same as above, but now for the B input matrix
+inline void LocalToPrivateDirectB(__local real* blm, real bpm[NWID], const int kg,
+                                  const int b_transpose) {
+  #pragma unroll
+  for (int ni=0; ni<NWID; ++ni) {
+    const int ng = ni + get_local_id(1)*NWID;
+    const int index = (b_transpose) ? ng*(WGD + PADB) + kg : kg*(WGD + PADB) + ng;
+    bpm[ni] = blm[index];
+  }
 }
 
 // =================================================================================================
-- 
cgit v1.2.3