commit working (but slower) double buffering version of K10

src/kernels/10_kernel_warptiling.cuh

@@ -112,8 +112,8 @@ __global__ void __launch_bounds__(NUM_THREADS)
   const uint threadRowInWarp = threadIdxInWarp / (WSUBN / TN); // i/4
 
   // allocate space for the current blocktile in SMEM
-  __shared__ float As[BM * BK];
-  __shared__ float Bs[BK * BN];
+  __shared__ float As[2 * BM * BK];
+  __shared__ float Bs[2 * BK * BN];
 
   // Move blocktile to beginning of A's row and B's column
   A += cRow * BM * K;
@@ -136,19 +136,44 @@ __global__ void __launch_bounds__(NUM_THREADS)
   float regM[WMITER * TM] = {0.0};
   float regN[WNITER * TN] = {0.0};
 
+  int As_offset = 0;
+  int Bs_offset = 0;
+
+  // double-buffering: load first blocktile into SMEM
+  wt::loadFromGmem<BM, BN, BK, rowStrideA, rowStrideB>(
+      N, K, A, B, As + As_offset * BM * BK, Bs + Bs_offset * BK * BN, innerRowA,
+      innerColA, innerRowB, innerColB);
+
+  __syncthreads();
+
   // outer-most loop over block tiles
-  for (uint bkIdx = 0; bkIdx < K; bkIdx += BK) {
+  for (uint bkIdx = 0; bkIdx < K - BK; bkIdx += BK) {
+    // double-buffering: load next blocktile into SMEM
     wt::loadFromGmem<BM, BN, BK, rowStrideA, rowStrideB>(
-        N, K, A, B, As, Bs, innerRowA, innerColA, innerRowB, innerColB);
-    __syncthreads();
+        N, K, A + BK, B + BK * N, As + (1 - As_offset) * BM * BK,
+        Bs + (1 - Bs_offset) * BK * BN, innerRowA, innerColA, innerRowB,
+        innerColB);
+
+    // compute the current blocktile
     wt::processFromSmem<BM, BN, BK, WM, WN, WMITER, WNITER, WSUBM, WSUBN, TM,
-                        TN>(regM, regN, threadResults, As, Bs, warpRow, warpCol,
+                        TN>(regM, regN, threadResults, As + As_offset * BM * BK,
+                            Bs + Bs_offset * BK * BN, warpRow, warpCol,
                             threadRowInWarp, threadColInWarp);
     A += BK;     // move BK columns to right
     B += BK * N; // move BK rows down
+
     __syncthreads();
+
+    As_offset = 1 - As_offset;
+    Bs_offset = 1 - Bs_offset;
   }
 
+  // compute the last blocktile
+  wt::processFromSmem<BM, BN, BK, WM, WN, WMITER, WNITER, WSUBM, WSUBN, TM, TN>(
+      regM, regN, threadResults, As + As_offset * BM * BK,
+      Bs + Bs_offset * BK * BN, warpRow, warpCol, threadRowInWarp,
+      threadColInWarp);
+
   // write out the results
   for (uint wSubRowIdx = 0; wSubRowIdx < WMITER; ++wSubRowIdx) {
     for (uint wSubColIdx = 0; wSubColIdx < WNITER; ++wSubColIdx) {