cgo ae update

Author: Jian Weng

apps/cpu/kernel/conv3d.py

@@ -11,17 +11,17 @@
 from tvm.relay import op
 
 workloads = [
-#(256, 16, 16, 256, 3, 1),
-#(512, 9, 9, 512, 3, 1),
+(256, 16, 16, 256, 3, 1),
+(512, 9, 9, 512, 3, 1),
 (128, 30, 30, 128, 3, 1),
-#(64, 56, 56, 128, 1, 2),
-#(64, 58, 58, 64, 3, 1),
-#(128, 28, 28, 256, 1, 2),
-#(256, 16, 16, 512, 3, 2),
-#(64, 58, 58, 128, 3, 2),
-#(4, 230, 230, 64, 7, 2),
-#(128, 30, 30, 256, 3, 2),
-#(256, 14, 14, 512, 1, 2)
+(64, 56, 56, 128, 1, 2),
+(64, 58, 58, 64, 3, 1),
+(128, 28, 28, 256, 1, 2),
+(256, 16, 16, 512, 3, 2),
+(64, 58, 58, 128, 3, 2),
+(4, 230, 230, 64, 7, 2),
+(128, 30, 30, 256, 3, 2),
+(256, 14, 14, 512, 1, 2)
 ]
 
 output = []
@@ -96,8 +96,8 @@
 
         relay.backend.compile_engine.get().clear()
         tune.cpu_idx += 1
-        #if tune.cpu_idx - target > 8:
-        #    break
+        if tune.cpu_idx - target > 8:
+            break
         if tune.cpu_idx >= tune.total_idx - 1:
             break
     #print(results)

apps/cpu/kernel/res

@@ -1 +1,16 @@
 (303464448, 861.4658333333336, 352.2652161674045)
+(303464448, 1169.7422000000001, 259.42848603735075)
+(303464448, 1008.5250666666667, 300.89926173376887)
+(303464448, 864.0738, 351.2020014957056)
+(303464448, 987.8491000000001, 307.1971700940963)
+(303464448, 1091.261066666667, 278.0860210902175)
+(303464448, 1014.0442, 299.2615588156808)
+(303464448, 1068.1393, 284.1056854663058)
+(3211264, 15.877699999999999, 202.24994804033332)
+(303464448, 991.7283666666665, 305.9955308326867)
+(3211264, 16.94766666666667, 189.48118718407642)
+(75866112, 289.8498333333333, 261.74281740142453)
+(75866112, 277.08956666666666, 273.7963500851169)
+(1376829440, 955.9776666666668, 1440.231804578419)
+(75866112, 280.5848333333334, 270.38564807197343)
+(3211264, 16.792066666666667, 191.23697301503486)

poc/vnni/input

@@ -1,8 +1,8 @@
-608
-14
-14
-192
-1
-1
+128
+28
+28
+256
 1
 1
+2
+2

poc/vnni/run-3d.py

@@ -0,0 +1,60 @@
+import subprocess
+
+#workloads = [(1024,16,16,2048,1024,1,1,1,1),
+#(960,8,8,160,960,1,1,1,1),
+#(256,56,56,512,256,1,1,1,1),
+#(64,56,56,128,64,1,1,1,1),
+#(512,10,10,512,512,3,3,1,1),
+#(192,16,16,160,192,3,3,1,1),
+#(64,65,65,128,64,3,3,1,1),
+#(320,65,65,384,320,3,3,1,1),
+#(384,8,8,384,384,1,3,1,1),
+#(192,23,23,192,192,7,1,1,1),
+#(3,224,224,64,3,7,7,2,2),
+#(1024,15,15,2048,1024,1,1,2,2),
+#(128,65,65,256,128,3,3,2,2)]
+#
+#workloads = [(1, 288, 35, 35, 384, 288, 3, 3, 2, 2), (1, 160, 9, 9, 224, 160, 3, 3, 1, 1), (1, 1056, 7, 7, 192, 1056, 1, 1, 1, 1), (1, 80, 73, 73, 192, 80, 3, 3, 1, 1), (1, 128, 16, 16, 128, 128, 3, 3, 1, 1), (1, 192, 16, 16, 192, 192, 3, 3, 1, 1), (1, 256, 16, 16, 256, 256, 3, 3, 1, 1), (1, 1024, 14, 14, 512, 1024, 1, 1, 1, 1), (1, 128, 16, 16, 160, 128, 3, 3, 1, 1), (1, 576, 14, 14, 192, 576, 1, 1, 1, 1), (1, 96, 16, 16, 128, 96, 3, 3, 1, 1), (1, 1024, 14, 14, 256, 1024, 1, 1, 1, 1), (1, 576, 14, 14, 128, 576, 1, 1, 1, 1), (1, 64, 29, 29, 96, 64, 3, 3, 1, 1), (1, 64, 56, 56, 128, 64, 1, 1, 2, 2), (1, 608, 14, 14, 192, 608, 1, 1, 1, 1)]
+#
+workloads = [
+(256, 16, 16, 256, 3, 1),
+(512, 9, 9, 512, 3, 1),
+(128, 30, 30, 128, 3, 1),
+(64, 56, 56, 128, 1, 2),
+(64, 58, 58, 64, 3, 1),
+(128, 28, 28, 256, 1, 2),
+(256, 16, 16, 512, 3, 2),
+(64, 58, 58, 128, 3, 2),
+(4, 230, 230, 64, 7, 2),
+(128, 30, 30, 256, 3, 2),
+(256, 14, 14, 512, 1, 2)
+]
+
+for i in workloads:
+    exec_time = []
+    with open('input', 'w') as f:
+        lst = list(i)
+        #for j in (i[1:5] + i[6:]):
+        #    f.write(str(j) + '\n')
+        f.write(str(i[0]) + '\n')
+        f.write(str(i[1]) + '\n')
+        f.write(str(i[2]) + '\n')
+        f.write(str(i[3]) + '\n')
+        f.write(str(i[4]) + '\n')
+        f.write(str(i[4]) + '\n')
+        f.write(str(i[5]) + '\n')
+        f.write(str(i[5]) + '\n')
+    try:
+        avg = []
+        for j in range(10):
+            res = subprocess.check_output('./mkldnn_conv3d.out < input', shell=True).decode('utf-8')
+            res = res.split('\n')[1]
+            res = res.lstrip('Exec time: ').rstrip('us')
+            res = float(res) * 1e6
+            res = int(res)
+            avg.append(res)
+        exec_time.append(sum(avg) / len(avg))
+    except:
+        print(i, 'fails')
+    print(i, min(exec_time))
+