tvm-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <...@apache.org>
Subject [GitHub] [incubator-tvm] comaniac commented on a change in pull request #5752: Make batch matrix multiplication on GPU tunable
Date Tue, 09 Jun 2020 20:25:13 GMT

comaniac commented on a change in pull request #5752:
URL: https://github.com/apache/incubator-tvm/pull/5752#discussion_r437696668



##########
File path: topi/python/topi/cuda/batch_matmul.py
##########
@@ -51,55 +62,73 @@ def _schedule(op):
             C = s.outputs[0].output(0)
 
         b, y, x = s[C].op.axis
-        y_bn = get_max_power2_factor(M, 64)
-        x_bn = get_max_power2_factor(N, 64)
-        by, y = s[C].split(y, y_bn)
-        bx, x = s[C].split(x, x_bn)
-        y_nthreads = min(y_bn, 8)
-        x_nthreads = min(x_bn, 8)
-        ty, yi = s[C].split(y, nparts=y_nthreads)
-        tx, xi = s[C].split(x, nparts=x_nthreads)
-        thread_x = te.thread_axis((0, x_nthreads), "threadIdx.x")
-        thread_y = te.thread_axis((0, y_nthreads), "threadIdx.y")
+
+        cfg.define_split("tile_y", y, num_outputs=3)
+        cfg.define_split("tile_x", x, num_outputs=3)
+        cfg.define_knob("auto_unroll_max_step", [8, 16, 32, 64])
+        target = tvm.target.Target.current()
+        if target.target_name in ['nvptx', 'rocm']:
+            # llvm-based backends cannot do non-explicit unrolling
+            cfg.define_knob("unroll_explicit", [1])
+        else:
+            cfg.define_knob("unroll_explicit", [0, 1])
+
+        if cfg.is_fallback:

Review comment:
       Sorry just found that the `unroll_explicit` is still missing in fallback. Also would
you mind moving L103 (define `tile_k`) up together with other tuning parameters?




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org



Mime
View raw message