Added UVA write test

TESTS.md

@@ -17,6 +17,7 @@ Level-1
 * UVA: Tests DMA (Direct Memory Access) between each GPU and the host/other GPUs.
   * Special flags:
     * `--fullduplex`: Performs exchanges instead of uni-directional transfers
+    * `--write`: Performs DMA write instead of read
     * `--from`: Specify only one GPU to copy from (0 for host, 1...N for a specific GPU or -1 for all GPUs)
     * `--to`: Specify a single target GPU to copy to (1...N for a specific GPU or -1 for all GPUs)
 

run.sh

@@ -39,19 +39,25 @@ fi
 echo "L1 Tests"
 echo "--------"
 
-echo "1/5 Half-duplex (unidirectional) memory copy"
+echo "1/7 Half-duplex (unidirectional) memory copy"
 ./build/halfduplex > l1-halfduplex.log
 
-echo "2/5 Full-duplex (bidirectional) memory copy"
+echo "2/7 Full-duplex (bidirectional) memory copy"
 ./build/fullduplex > l1-fullduplex.log
 
-echo "3/5 Half-duplex DMA"
+echo "3/7 Half-duplex DMA Read"
 ./build/uva > l1-uvahalf.log
 
-echo "4/5 Full-duplex DMA"
+echo "4/7 Full-duplex DMA Read"
 ./build/uva --fullduplex > l1-uvafull.log
 
-echo "5/5 Scaling"
+echo "5/7 Half-duplex DMA Write"
+./build/uva --write > l1-uvawhalf.log
+
+echo "6/7 Full-duplex DMA Write"
+./build/uva --write --fullduplex > l1-uvawfull.log
+
+echo "7/7 Scaling"
 ./build/sgemm -n 4096 -k 4096 -m 4096 --repetitions=100 --regression=false --scaling=true > l1-scaling.log
 
 # Run L2 tests

src/L1/uva.cu

@@ -41,6 +41,7 @@ DEFINE_uint64(type_size, sizeof(float), "The size of the data chunk to "
 DEFINE_uint64(repetitions, 100, "Number of repetitions to average");
 DEFINE_uint64(block_size, 32, "Copy kernel block size");
 DEFINE_bool(fullduplex, false, "True for bi-directional copy");
+DEFINE_bool(write, false, "Perform DMA write instead of read");
 
 DEFINE_int32(from, -1, "Only copy from a single GPU index/host (Host is "
              "0, GPUs start from 1), or -1 for all");
@@ -144,7 +145,13 @@ void CopySegmentUVA(int a, int b)
     size_t sz = FLAGS_size / FLAGS_type_size, typesize = FLAGS_type_size;
     dim3 block_dim (FLAGS_block_size),
          grid_dim((sz + FLAGS_block_size - 1) / FLAGS_block_size);
-
+
+    // If using UVA to write, simply swap the buffers
+    if (FLAGS_write)
+    {
+        std::swap(deva_buff, devb_buff);
+        std::swap(deva_buff2, devb_buff2);
+    }
 
     // Copy or Exchange using UVA
     auto t1 = std::chrono::high_resolution_clock::now();
@@ -182,6 +189,15 @@ void CopySegmentUVA(int a, int b)
     double MBps = (FLAGS_size / 1024.0 / 1024.0) / (mstime / 1000.0);
 
     printf("%.2lf MB/s (%lf ms)\n", MBps, mstime);
+
+
+    // Swap buffers back, if necessary
+    if (FLAGS_write)
+    {
+        std::swap(deva_buff, devb_buff);
+        std::swap(deva_buff2, devb_buff2);
+    }
+
 
     // Free buffers
     if (a > 0)
@@ -283,17 +299,24 @@ int main(int argc, char **argv)
 
             if (FLAGS_fullduplex)
             {
-                if (i == 0)
-                    printf("Exchanging between host and GPU %d: ", j - 1);
-                else
-                    printf("Exchanging between GPU %d and GPU %d: ", i - 1, j - 1);
+                printf("Exchanging between GPU %d and GPU %d: ", i - 1, j - 1);
             }
             else
             {
-                if (i == 0)
-                    printf("Copying from host to GPU %d: ", j - 1);
+                if (!FLAGS_write)
+                {
+                    if (i == 0)
+                        printf("Copying from host to GPU %d: ", j - 1);
+                    else
+                        printf("Copying from GPU %d to GPU %d: ", i - 1, j - 1);
+                }
                 else
-                    printf("Copying from GPU %d to GPU %d: ", i - 1, j - 1);
+                {
+                    if (i == 0)
+                        printf("Copying from GPU %d to host: ", j - 1);
+                    else
+                        printf("Copying from GPU %d to GPU %d: ", j - 1, i - 1);
+                }
             }
 
             // Make sure that DMA access is possible