[WebGPU] Support dot4I8Packed(int8x4, int8x4) as a pure extern method
#16976
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This patch adds the support of `__dp4a(int8x4, int8x4)` as a pure
extern method of WebGPU target. In the generated WGSL shader,
`int8x4` will be translated into `u32`, and `__dp4a(int8x4, int8x4)`
will be translated into the WGSL built-in function
`dot4I8Packed(u32, u32)`.
Here is an example to use `__dp4a` in WebGPU target:
```
n = te.var("n")
A = te.placeholder((n,), "int8x4", name="A")
B = te.placeholder((n,), "int8x4", name="B")
C = te.compute(A.shape, lambda i: tvm.tir.call_pure_extern("int32", "__dp4a", A[i], B[i]), name="C")
s = te.create_schedule(C.op)
bx, tx = s[C].split(C.op.axis[0], factor=64)
s[C].bind(bx, te.thread_axis("blockIdx.x"))
s[C].bind(tx, te.thread_axis("threadIdx.x"))
mod = tvm.build(s, [A, B, C], tgt, name="dp4aTest")
```
Issue: apache#16627
tqchen
approved these changes
May 8, 2024
tqchen
requested changes
May 8, 2024
src/target/source/codegen_webgpu.cc
Outdated
| if (op->args.size() != 3) { | ||
| LOG(FATAL) << "__dp4a can only accept 2 parameters (now: " << op->args.size() - 1 << ")"; | ||
| } else { | ||
| os << "dot4I8Packed(" << PrintExpr(op->args[1]) << ", " << PrintExpr(op->args[2]) << ")"; |
There was a problem hiding this comment.
given we are calling pure_extern about we use dot4I8Packed directly in code? instead of __dp4a. If there is a builtin high-level intrinsic function tir.dp4a, we can do the intrinsic translation to this dot4I8Packed in https://github.com/apache/tvm/blob/main/src/target/source/intrin_rule_webgpu.cc
There was a problem hiding this comment.
Done. Thanks for your suggestion!
Jiawei-Shao
changed the title
__dp4a(int8x4, int8x4) as a pure extern methoddot4I8Packed(int8x4, int8x4) as a pure extern method
tqchen
reviewed
May 9, 2024
| @@ -113,6 +113,8 @@ TVM_REGISTER_OP("tir.trunc") | |||
| // extra dispatch | |||
| TVM_REGISTER_OP("tir.erf").set_attr<FLowerIntrinsic>("webgpu.FLowerIntrinsic", DispatchFastErf); | |||
|
|
|||
| TVM_REGISTER_OP("tir.dot4I8Packed").set_attr<FLowerIntrinsic>("webgpu.FLowerIntrinsic", DispatchPureExtern<Direct>); | |||
There was a problem hiding this comment.
sorry i was not being clear, for tir, it is better to have a common name dp4a (as this intrinsic shared across backends)
There was a problem hiding this comment.
we can add tir.dp4a intrinsic, and use it to lower to various places
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch adds the support of
dot4I8Packed(int8x4, int8x4)as a pure extern method of WebGPU target. In the generated WGSL shader,int8x4will be translated intou32, anddot4I8Packed(int8x4, int8x4)will be translated into the WGSL built-in functiondot4I8Packed(u32, u32).Here is an example to use
dot4I8Packedin WebGPU target:Issue: #16627