From: Haichen Shen <shenhaichen@gmail.com>
Date: Wed, 18 Dec 2019 21:17:18 +0000 (-0800)
Subject: [TOPI] Allow batch matmul to be fused into injective ops (#4537)
X-Git-Tag: upstream/0.7.0~1505
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=8ba9e8bc75f3c4bc869792a8e7873bb975199100;p=platform%2Fupstream%2Ftvm.git

[TOPI] Allow batch matmul to be fused into injective ops (#4537)
---

diff --git a/topi/python/topi/x86/batch_matmul.py b/topi/python/topi/x86/batch_matmul.py
index b505cbf..25b49d1 100644
--- a/topi/python/topi/x86/batch_matmul.py
+++ b/topi/python/topi/x86/batch_matmul.py
@@ -92,33 +92,40 @@ def schedule_batch_matmul(cfg, outs):
     def _callback(op):
         if "batch_matmul" in op.tag:
             C = op.output(0)
-            A, B = s[C].op.input_tensors
+            A, B = op.input_tensors
             _, M, K = get_const_tuple(A.shape)
             _, _, N = get_const_tuple(C.shape)
 
+            if op not in s.outputs:
+                s[C].compute_inline()
+                O = outs[0]
+            else:
+                O = C
+
+            CC = s.cache_write(C, "global")
+
             # create tuning space
             cfg.define_split("tile_y", M, num_outputs=2)
             cfg.define_split("tile_x", N, num_outputs=2)
             cfg.define_split("tile_k", K, num_outputs=2)
 
-            k, = s[C].op.reduce_axis
-
-            ko, ki = cfg["tile_k"].apply(s, C, k)
-            CC = s.rfactor(C, ki)
-
-            b, y, x = s[C].op.axis
-            yo, yi = cfg["tile_y"].apply(s, C, y)
-            xo, xi = cfg["tile_x"].apply(s, C, x)
-            s[C].reorder(b, yo, xo, yi, xi)
-            bxyo = s[C].fuse(b, yo, xo)
-            s[C].parallel(bxyo)
-            s[C].fuse(yi, xi)
-
-            s[CC].compute_at(s[C], bxyo)
-            _, _, y, x = s[CC].op.axis
-            s[CC].fuse(y, x)
-            s[CC].vectorize(s[CC].op.axis[0])
-            s[C].pragma(bxyo, 'auto_unroll_max_step', 16)
+            b, y, x = s[O].op.axis
+            yo, yi = cfg["tile_y"].apply(s, O, y)
+            xo, xi = cfg["tile_x"].apply(s, O, x)
+            s[O].reorder(b, yo, xo, yi, xi)
+            bxyo = s[O].fuse(b, yo, xo)
+            s[O].parallel(bxyo)
+
+            s[CC].compute_at(s[O], bxyo)
+            k, = s[CC].op.reduce_axis
+            ko, ki = cfg["tile_k"].apply(s, CC, k)
+
+            Crf = s.rfactor(CC, ki)
+            s[Crf].compute_at(s[CC], s[CC].op.axis[0])
+            _, _, y, x = s[Crf].op.axis
+            s[Crf].fuse(y, x)
+            s[Crf].vectorize(s[Crf].op.axis[0])
+            s[O].pragma(bxyo, 'auto_unroll_max_step', 16)
 
     traverse_inline(s, outs[0].op, _callback)
     return s