You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/11/08 23:55:15 UTC
[GitHub] [incubator-tvm] comaniac commented on a change in pull request #6877: [AutoScheduler] Add winograd support in tuning networks
comaniac commented on a change in pull request #6877:
URL: https://github.com/apache/incubator-tvm/pull/6877#discussion_r519495596
##########
File path: src/auto_scheduler/search_policy/sketch_policy.cc
##########
@@ -116,8 +116,8 @@ SketchPolicy::SketchPolicy(SearchTask task, CostModel program_cost_model,
} else if (IsGPUTask(node->search_task)) {
// Sketch Generation Rules
node->sketch_rules.push_back(&rule_add_cache_read_stage);
- node->sketch_rules.push_back(&rule_always_inline);
node->sketch_rules.push_back(&rule_special_compute_location_gpu);
+ node->sketch_rules.push_back(&rule_always_inline);
Review comment:
Out of curiosity, why changed to this order? I saw that the condition of applying special compute location also checks should be always inlined, so then it should be fine to keep always inline to go first?
##########
File path: python/tvm/relay/op/strategy/cuda.py
##########
@@ -206,19 +198,39 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
name="conv2d_nhwc_winograd_direct.cuda",
plevel=5,
)
- if target.kind.name == "cuda":
- if nvcc.have_tensorcore(tvm.gpu(0).compute_version):
- if (
- (N % 16 == 0 and CI % 16 == 0 and CO % 16 == 0)
- or (N % 8 == 0 and CI % 16 == 0 and CO % 32 == 0)
- or (N % 32 == 0 and CI % 16 == 0 and CO % 8 == 0)
- ):
- strategy.add_implementation(
- wrap_compute_conv2d(topi.cuda.conv2d_nhwc_tensorcore),
- wrap_topi_schedule(topi.cuda.schedule_conv2d_nhwc_tensorcore),
- name="conv2d_nhwc_tensorcore.cuda",
- plevel=20,
- )
+ if (
+ target.kind.name == "cuda"
+ and nvcc.have_tensorcore(tvm.gpu(0).compute_version)
+ and (
+ (N % 16 == 0 and CI % 16 == 0 and CO % 16 == 0)
+ or (N % 8 == 0 and CI % 16 == 0 and CO % 32 == 0)
+ or (N % 32 == 0 and CI % 16 == 0 and CO % 8 == 0)
+ )
+ ):
+ strategy.add_implementation(
+ wrap_compute_conv2d(topi.cuda.conv2d_nhwc_tensorcore),
+ wrap_topi_schedule(topi.cuda.schedule_conv2d_nhwc_tensorcore),
+ name="conv2d_nhwc_tensorcore.cuda",
+ plevel=20,
+ )
+
+ # register auto-scheduler implementations
+ judge_winograd_auto_scheduler = (
Review comment:
Following the above style of judging Winograd, we can remove this variable.
##########
File path: python/tvm/relay/op/strategy/cuda.py
##########
@@ -206,19 +198,39 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
name="conv2d_nhwc_winograd_direct.cuda",
plevel=5,
)
- if target.kind.name == "cuda":
- if nvcc.have_tensorcore(tvm.gpu(0).compute_version):
- if (
- (N % 16 == 0 and CI % 16 == 0 and CO % 16 == 0)
- or (N % 8 == 0 and CI % 16 == 0 and CO % 32 == 0)
- or (N % 32 == 0 and CI % 16 == 0 and CO % 8 == 0)
- ):
- strategy.add_implementation(
- wrap_compute_conv2d(topi.cuda.conv2d_nhwc_tensorcore),
- wrap_topi_schedule(topi.cuda.schedule_conv2d_nhwc_tensorcore),
- name="conv2d_nhwc_tensorcore.cuda",
- plevel=20,
- )
+ if (
+ target.kind.name == "cuda"
+ and nvcc.have_tensorcore(tvm.gpu(0).compute_version)
+ and (
+ (N % 16 == 0 and CI % 16 == 0 and CO % 16 == 0)
+ or (N % 8 == 0 and CI % 16 == 0 and CO % 32 == 0)
+ or (N % 32 == 0 and CI % 16 == 0 and CO % 8 == 0)
+ )
+ ):
+ strategy.add_implementation(
+ wrap_compute_conv2d(topi.cuda.conv2d_nhwc_tensorcore),
+ wrap_topi_schedule(topi.cuda.schedule_conv2d_nhwc_tensorcore),
+ name="conv2d_nhwc_tensorcore.cuda",
+ plevel=20,
+ )
+
+ # register auto-scheduler implementations
+ judge_winograd_auto_scheduler = (
+ ("float" in data.dtype and "float" in kernel.dtype)
Review comment:
Will `dtype` be "float32"?
----------------------------------------------------------------
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