Improved the benchmark scripts; added gemmstridedbatched benchmark

3 files changed, 63 insertions, 16 deletions

diff --git a/scripts/benchmark/benchmark.py b/scripts/benchmark/benchmark.py
index d0a9d80f..d84d5d98 100644
--- a/scripts/benchmark/benchmark.py
+++ b/scripts/benchmark/benchmark.py
@@ -22,6 +22,7 @@ EXPERIMENTS = {
     "gemm": settings.GEMM,
     "gemm_small": settings.GEMM_SMALL,
     "gemmbatched": settings.GEMMBATCHED,
+    "gemmstridedbatched": settings.GEMMSTRIDEDBATCHED,
     "symm": settings.SYMM,
     "syrk": settings.SYRK,
     "summary": settings.SUMMARY,
@@ -162,6 +163,11 @@ def benchmark_single(benchmark, comparisons, platform, device, num_runs, precisi
         label_names.append("CLBlast FP32")
         y_keys = [y_key + [y_key[0] + "_FP32"] for y_key in y_keys]
 
+    # For batched routines: comparison is non-batched
+    if benchmark in ["axpybatched", "gemmbatched", "gemmstridedbatched"]:
+        for index in range(1, len(label_names)):
+            label_names[index] += " (non-batched)"
+
     # Plots the graphs
     plot.plot_graphs(results["benchmarks"], pdf_file_name, results["num_rows"], results["num_cols"],
                      x_keys, y_keys, titles, x_labels, y_labels,
diff --git a/scripts/benchmark/plot.py b/scripts/benchmark/plot.py
index 06fb278f..66d71595 100644
--- a/scripts/benchmark/plot.py
+++ b/scripts/benchmark/plot.py
@@ -109,9 +109,20 @@ def plot_graphs(results, file_name, num_rows, num_cols,
             assert len(label_names) == len(y_keys[index])
             for i in range(len(y_keys[index])):
                 color = COLORS[i]
-                if label_names[i] == "cuBLAS":
+                marker = MARKERS[i]
+                if label_names[i] in ["CLBlast", "CLBlast FP32"]:
+                    color = BLUEISH
+                    marker = "o-"
+                elif label_names[i] in ["CLBlast FP16"]:
+                    color = PURPLISH
+                    marker = ".-"
+                elif label_names[i] in ["clBLAS", "clBLAS FP32"]:
+                    color = REDISH
+                    marker = "x-"
+                elif label_names[i] == "cuBLAS":
                     color = GREEN
-                ax.plot(x_location, y_list[i], MARKERS[i], label=label_names[i], color=color)
+                    marker = ".-"
+                ax.plot(x_location, y_list[i], marker, label=label_names[i], color=color)
 
             # Sets the legend
             leg = ax.legend(loc=(0.02, 1.0 - legend_from_top - legend_from_top_per_item * len(y_keys[index])),
diff --git a/scripts/benchmark/settings.py b/scripts/benchmark/settings.py
index d0d17178..98b97530 100644
--- a/scripts/benchmark/settings.py
+++ b/scripts/benchmark/settings.py
@@ -63,28 +63,28 @@ AXPYBATCHED = {
     "num_rows": 1, "num_cols": 3,
     "benchmarks": [
         {
-            "name": "axpybatched", "num_runs": 30,
-            "title": "8 AXPYs",
+            "name": "axpybatched", "num_runs": 10,
+            "title": "b=8",
             "x_label": "sizes (n)", "x_keys": ["n"],
             "y_label": "GB/s (higher is better)", "y_key": "GBs",
             "arguments": [{"batch_num": 8, "n": n, "incx": 1, "incy": 1, "step": 0, "num_steps": 1}
                           for n in utils.powers_of_2(utils.k(8), utils.m(4))],
         },
         {
-            "name": "axpybatched", "num_runs": 20,
-            "title": "64 AXPYs",
+            "name": "axpybatched", "num_runs": 5,
+            "title": "b=64",
             "x_label": "sizes (n)", "x_keys": ["n"],
             "y_label": "GB/s (higher is better)", "y_key": "GBs",
             "arguments": [{"batch_num": 64, "n": n, "incx": 1, "incy": 1, "step": 0, "num_steps": 1}
                           for n in utils.powers_of_2(utils.k(8), utils.m(4))],
         },
         {
-            "name": "axpybatched", "num_runs": 40,
+            "name": "axpybatched", "num_runs": 10,
             "title": "n=512K",
-            "x_label": "batch size", "x_keys": ["batch_num"],
+            "x_label": "num batches (b)", "x_keys": ["batch_num"],
             "y_label": "GB/s (higher is better)", "y_key": "GBs",
             "arguments": [{"batch_num": b, "n": utils.k(512), "incx": 1, "incy": 1, "step": 1, "num_steps": 1}
-                          for b in utils.powers_of_2(1, 512)],
+                          for b in utils.powers_of_2(1, 256)],
         }
     ]
 }
@@ -221,28 +221,58 @@ GEMMBATCHED = {
     "num_rows": 1, "num_cols": 3,
     "benchmarks": [
         {
-            "name": "gemmbatched", "num_runs": 40,
-            "title": "8 GEMMs",
+            "name": "gemmbatched", "num_runs": 20,
+            "title": "b=8",
             "x_label": "sizes (m=n=k)", "x_keys": ["m"],
             "y_label": "GFLOPS (higher is better)", "y_key": "GFLOPS",
             "arguments": [{"batch_num": 8, "m": 32, "n": 32, "k": 32, "layout": 102,
                            "transA": 111, "transB": 111, "step": 32, "num_steps": 20}],
         },
         {
-            "name": "gemmbatched", "num_runs": 20,
-            "title": "64 GEMMs",
+            "name": "gemmbatched", "num_runs": 10,
+            "title": "b=64",
+            "x_label": "sizes (m=n=k)", "x_keys": ["m"],
+            "y_label": "GFLOPS (higher is better)", "y_key": "GFLOPS",
+            "arguments": [{"batch_num": 64, "m": 32, "n": 32, "k": 32, "layout": 102,
+                           "transA": 111, "transB": 111, "step": 32, "num_steps": 20}],
+        },
+        {
+            "name": "gemmbatched", "num_runs": 10,
+            "title": "m=n=k=128",
+            "x_label": "num batches (b)", "x_keys": ["batch_num"],
+            "y_label": "GFLOPS (higher is better)", "y_key": "GFLOPS",
+            "arguments": [{"batch_num": b, "m": 128, "n": 128, "k": 128, "layout": 102,
+                           "transA": 111, "transB": 111} for b in utils.powers_of_2(1, utils.k(4))],
+        }
+    ]
+}
+
+GEMMSTRIDEDBATCHED = {
+    "num_rows": 1, "num_cols": 3,
+    "benchmarks": [
+        {
+            "name": "gemmstridedbatched", "num_runs": 20,
+            "title": "b=8",
+            "x_label": "sizes (m=n=k)", "x_keys": ["m"],
+            "y_label": "GFLOPS (higher is better)", "y_key": "GFLOPS",
+            "arguments": [{"batch_num": 8, "m": 32, "n": 32, "k": 32, "layout": 102,
+                           "transA": 111, "transB": 111, "step": 32, "num_steps": 20}],
+        },
+        {
+            "name": "gemmstridedbatched", "num_runs": 10,
+            "title": "b=64",
             "x_label": "sizes (m=n=k)", "x_keys": ["m"],
             "y_label": "GFLOPS (higher is better)", "y_key": "GFLOPS",
             "arguments": [{"batch_num": 64, "m": 32, "n": 32, "k": 32, "layout": 102,
                            "transA": 111, "transB": 111, "step": 32, "num_steps": 20}],
         },
         {
-            "name": "gemmbatched", "num_runs": 30,
+            "name": "gemmstridedbatched", "num_runs": 10,
             "title": "m=n=k=128",
-            "x_label": "batch size", "x_keys": ["batch_num"],
+            "x_label": "num batches (b)", "x_keys": ["batch_num"],
             "y_label": "GFLOPS (higher is better)", "y_key": "GFLOPS",
             "arguments": [{"batch_num": b, "m": 128, "n": 128, "k": 128, "layout": 102,
-                           "transA": 111, "transB": 111} for b in utils.powers_of_2(1, utils.k(16))],
+                           "transA": 111, "transB": 111} for b in utils.powers_of_2(1, utils.k(4))],
         }
     ]
 }