Implement flop support for int8 models (#2776)

Author: ajtulloch

python/tvm/autotvm/task/task.py

@@ -338,7 +338,7 @@ def _count_flop(exp):
                             expr.Max, expr.Min,
                             expr.EQ, expr.NE, expr.LT, expr.LE, expr.GT, expr.GE,
                             expr.And, expr.Or, expr.Not)):
-            base = 1 if "float" in exp.a.dtype else 0
+            base = 1
 
             if isinstance(exp, expr.Not):  # unary
                 return base + _count_flop(exp.a)
@@ -348,6 +348,10 @@ def _count_flop(exp):
             return _count_flop(exp.condition) + max(_count_flop(exp.true_value),
                                                     _count_flop(exp.false_value))
         if isinstance(exp, expr.Call):
+            if exp.call_type == expr.Call.Halide:
+                # Ignore flops from indexing expressions.
+                return 0
+
             return sum([_count_flop(x) for x in exp.args])
 
         raise FlopCalculationError("Found unsupported operator in the compute expr")

tests/python/unittest/test_autotvm_flop_calculator.py

@@ -5,11 +5,17 @@
 
 from tvm.autotvm.task.task import compute_flop
 
+def random_dtypes():
+    """Return pair of (input, accumulator) dtypes"""
+    candidates = [("float32", "float32"), ("float16", "float32"), ("int8", "int32")]
+    return candidates[np.random.choice(len(candidates))]
+
 def test_conv():
     for i in range(5):
         N, H, W, CO, CI, KH, KW = [np.random.randint(10, 32) for _ in range(7)]
-        D = tvm.placeholder((N, CI, H, W))
-        K = tvm.placeholder((CO, CI, KH, KW))
+        (input_dtype, acc_dtype) = random_dtypes()
+        D = tvm.placeholder((N, CI, H, W), dtype=input_dtype)
+        K = tvm.placeholder((CO, CI, KH, KW), dtype=input_dtype)
 
         KH = min(H, KH)
         KW = min(W, KW)
@@ -22,7 +28,8 @@ def test_conv():
         OW = (W - KW) + 1
 
         C = tvm.compute((N, CO, OH, OW), lambda n, co, h, w:
-        tvm.sum(D[n][ci][h][w] * K[co][ci][h][w], axis=[ci, kh, kw]))
+        tvm.sum(D[n][ci][h][w].astype(acc_dtype) * K[co][ci][h][w].astype(acc_dtype),
+                axis=[ci, kh, kw]))
 
         s = tvm.create_schedule([C.op])
 
@@ -31,15 +38,16 @@ def test_conv():
 def test_pack_gemm():
     for i in range(5):
         N, L, M = [np.random.randint(10, 128) * 4 for _ in range(3)]
-        A = tvm.placeholder((N, L))
-        B = tvm.placeholder((M, L))
+        (input_dtype, acc_dtype) = random_dtypes()
+        A = tvm.placeholder((N, L), dtype=input_dtype)
+        B = tvm.placeholder((M, L), dtype=input_dtype)
         k = tvm.reduce_axis((0, L))
 
         bn = 4
         A_pack = tvm.compute((N // bn, L, bn), lambda i, j, k: A[i * bn + k][j])
         B_pack = tvm.compute((M // bn, L, bn), lambda i, j, k: B[i * bn + k][j])
         C_pack = tvm.compute((N // bn, M // bn, bn, bn), lambda i, j, ii, jj:
-        tvm.sum(A_pack[i, k, ii] * B_pack[j, k, jj], axis=[k]))
+        tvm.sum(A_pack[i, k, ii].astype(acc_dtype) * B_pack[j, k, jj].astype(acc_dtype), axis=[k]))
         C = tvm.compute((N, M), lambda i, j: C_pack[i // bn][j // bn][i % bn][j % bn])
 
         s = tvm.create_schedule([C.op])
@@ -48,14 +56,61 @@ def test_pack_gemm():
 def test_outer_dot():
     for i in range(5):
         N, M = [np.random.randint(10, 128) * 4 for _ in range(2)]
-        A = tvm.placeholder((N,))
-        B = tvm.placeholder((M,))
+        (input_dtype, acc_dtype) = random_dtypes()
+        A = tvm.placeholder((N,), dtype=input_dtype)
+        B = tvm.placeholder((M,), dtype=input_dtype)
 
-        C = tvm.compute((N, M), lambda i, j: A[i] * B[j])
+        C = tvm.compute((N, M), lambda i, j: A[i].astype(acc_dtype) * B[j].astype(acc_dtype))
 
         s = tvm.create_schedule([C.op])
         assert compute_flop(s) == N * M
 
+def test_max_pool():
+    for i in range(5):
+        N, H, W, CO, CI, KH, KW = [np.random.randint(10, 32) for _ in range(7)]
+        (input_dtype, _) = random_dtypes()
+        D = tvm.placeholder((N, CI, H, W), dtype=input_dtype)
+
+        KH = min(H, KH)
+        KW = min(W, KW)
+
+        kh = tvm.reduce_axis((0, KH))
+        kw = tvm.reduce_axis((0, KW))
+
+        OH = (H - KH) + 1
+        OW = (W - KW) + 1
+
+        C = tvm.compute(
+            (N, CO, OH, OW),
+            lambda n, co, h, w: tvm.max(D[n][co][h + kh][w + kw], axis=[kh, kw]))
+
+        s = tvm.create_schedule([C.op])
+
+        assert compute_flop(s) == N * CO * OH * OW * KH * KW
+
+def test_average_pool():
+    for i in range(5):
+        N, H, W, CO, CI, KH, KW = [np.random.randint(10, 32) for _ in range(7)]
+        (input_dtype, acc_dtype) = random_dtypes()
+        D = tvm.placeholder((N, CI, H, W), dtype=input_dtype)
+
+        KH = min(H, KH)
+        KW = min(W, KW)
+
+        kh = tvm.reduce_axis((0, KH))
+        kw = tvm.reduce_axis((0, KW))
+
+        OH = (H - KH) + 1
+        OW = (W - KW) + 1
+
+        C = tvm.compute(
+            (N, CO, OH, OW),
+            lambda n, co, h, w: tvm.sum(D[n][co][h + kh][w + kw].astype(acc_dtype) / (KW * KH), axis=[kh, kw]))
+
+        s = tvm.create_schedule([C.op])
+
+        assert compute_flop(s) == 2 * N * CO * OH * OW * KH * KW
+
 def test_move():
     """No float number operation in simple move. So the estimator should raise an error """
     N = 1024