[TVM Discuss] [RFC] Introducing Hexagon backend
The special runtimes are mostly for offloading to extra devices from host. ARM and X86 are usually the hosts, so the runtime for them is fairly simple. It's still there, look at cpu_device_api.cc for example. --- [Visit Topic](https://discuss.tvm.ai/t/introducing-hexagon-backend/2421/18) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/1b8badc1432bf8cd41a1d0b9cc8085d18034dc6bfbf5eb7bd3d5247f2bfec4f0). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=P2IBsTi74G2tcY7pMUzQvQ2
Re: [dmlc/tvm] [RFC][EXPR] Formalize Integer Arithmetic Analysis (#2588)
https://github.com/dmlc/tvm/pull/3368 finishes the migration of the analysis stack. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/2588#issuecomment-507753372
Re: [dmlc/tvm] [RFC][EXPR] Formalize Integer Arithmetic Analysis (#2588)
Closed #2588. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/2588#event-2454985179
[dmlc/tvm] [RFC] Build and Evolve Internal Low-level IR (HalideIR) (#3474)
Dear Community: TVM makes use of HalideIR as our low-level IR structure, and it has served us well. One of the main gains we get is the integer simplification infrastructure provided by HalideIR. As the project moves forward, we start to introduce our own integer simplification infra to tackle new problems, and as of now, we start to use the new integer analysis infra in TVM. We also see increasing needs to evolve our own version of low-level IRs. Here is a brief summary of needs: - Introducing new IR nodes to enhance the analysis required by TVM (Reduce, ConstraintHints) - Different dialect and intrinsics support, mainly in the case of pragma, and region annotations, both of which will make the IR more coupled with the project. - Examples include dma_copy regions, thread bindings, and bound hints. - Unification of the object protocol with the rest of the project, in particular, runtime::Object, Node, and NDArray. - General refactoring concerns to reduce duplicated objects(Var, Range) - Long term: unification of IR infra, in particular between low-level IR and high-level IRs(relay). These technical needs will require us to evolve the IR more frequently, and it starts to make sense to build and evolve our own version of low-level IR. Action Items - [ ] Create include/tvm/node to introduce the node infra - As we start to unify node with objects, we might consider moving certain things to runtime. - [ ] Create include/tvm/ir folder to introduce IR structures, - The initial structure will directly reflect the structures in HalideIR and new additions like Reduce with clear acknowledgment to the original design (Halide). - Start to evolve new IRs, refactoring and remove duplicated objects and further changes as mentioned above. Please share your thoughts and suggestions about the proposal -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/3474
Re: [dmlc/tvm] [RFC][Codegen] Broadcast ops with symbolic shape (#3390)
close by https://github.com/dmlc/tvm/pull/3389 -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/3390#issuecomment-507793738
Re: [dmlc/tvm] [RFC][Codegen] Broadcast ops with symbolic shape (#3390)
Closed #3390. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/3390#event-2455283439
[dmlc/tvm] [RFC][ARITH] Improving Context-Dependent DivMod Simplifications (#3478)
Many parts of our code base relies on integer simplifications that involves
div/mod. These simplifications are in particular tricky due to the different
division mode, and the dependence on knowing the sign of the operand.
Here is a list of division variants:
## Trunc div/mod
IEEE standard integer division is truncated division, directly maps to the
hardware. This is the division used by C and TVM by default. The problem of
trunc div is that we usually relies on knowing the sign of both operands. For
example, we cannot simplify ```(x * 4 + 3)/4 => x ``` if x can be negative.
## Floor div/mod
Always round down. This is the division mode adopted by python. To simplify
floordiv, we usually need to know the sign of divisor, but not necessarily the
sign of dividend ```floordiv(x * 4 + 3,4) => x ```.
## Euclidean div/mod
Always ensures that the remainder is non-negative. This is the division mode
used by Halide by default. Note that when the divisor is positive, Euclidean
and floordiv are equivalent. But floordiv will produce a negative remainder
when the divisor is negative.
## Discussion
Different division mode has its own advantage and disadvantages. While Floor
and Euclidean mode enjoys better invariance in simplification. We will have
some trouble lowering them into the trunc div/mod, especially when we do not
know the sign. In general, it is good to know the sign when lowering. However,
this will leads to some slight differences, consider the following case:
```c++
for (x = 0; x < 100; x++) {
// trunc division
a [(x * 5 + 3)/ 4] = xyz.
}
```
```(x * 5 + 3)/4``` can be simplified to (x * 5 /4) , if we setup the
simplifier, with currect context(```x in [0, 100]```). However, there can be
cases where we forget to do so in our code, i that scenario the simplifcation
cannot be done. If instead we use
```c++
for (x = 0; x < 100; x++) {
// floor division
a [floordiv(x * 5 + 3, 4)] = xyz.
}
```
```(x * 5 + 3)/4``` can be simplified to floordiv(x * 5, 4) in the analysis.
However, to get the most efficient version of the code. We need to lower the
floor division to ```x * 5 / 4``` knowing that x is non-negative. The context
information is only needed during the floor division lowering.
In summary, using trunc div during analysis requires m
ore information each time we transform it, using floor/euclidean div makes
simplification easier, but still requires us to provide context information to
most efficiently lowers to trunc div/mod.
## Proposal to be Discussed: Adding Floor Div/Mod
It might be helpful to encourage the use of floor div/mod for integer analysis
while providing an effective way to lower to trunc div when necessary. To avoid
confusion between division mode, I propose we add an explicit floordiv/floormod
node and operator to the IR.
The reason to choose floordiv/mod over Euclidean is mainly due to floor is more
common(used by python) and easier to understand. The difference in behavior
happens when the divisor is negative, and in most cases, we will know the sign
of divisor.
The original truc div/mod simplification will be kept around for a while, but
we will encourage most index calculations to use floordiv/mod, and provide
rewriter to rewrite truc div/mod into floor version when the sign of the
dividend is positive
--
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/dmlc/tvm/issues/3478
Re: [dmlc/tvm] [RFC][ARITH] Improving Context-Dependent DivMod Simplifications (#3478)
cc @dmlc/tvm-team -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/3478#issuecomment-507892218
Re: [dmlc/tvm] [DEV] TVM v0.6 Roadmap (#2623)
# TVM Monthly - June 2019 https://discuss.tvm.ai/t/tvm-monthly-june-2019 -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/2623#issuecomment-507917006
[TVM Discuss] [Development] [solved][Relay] Broken case of type infer
solved in latest master --- [Visit Topic](https://discuss.tvm.ai/t/solved-relay-broken-case-of-type-infer/3169/2) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/a3f9cac923e2d47ecfc033511bb3a098a581149c5dcd5c783f341c7ad40ad00a). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=POy2p6tlX7yfz9M5vbwCdg2
Re: [dmlc/tvm] [RFC][ARITH] Introduce FloorDiv/Mod for Context-Independent Simplifications (#3478)
Can you clarify the example above? the simplifications seem invalid to me, how can you drop the addition by 3? -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/3478#issuecomment-507935931
Re: [dmlc/tvm] [RFC][ARITH] Introduce FloorDiv/Mod for Context-Independent Simplifications (#3478)
Sorry it should be ```(x * 4 + 3) / 4=>x``` -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/3478#issuecomment-507936537
