Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[BACKEND] Don't use MMAV3 when K dim is smaller than native size #4700

Merged
merged 1 commit into from
Sep 11, 2024
Merged

[BACKEND] Don't use MMAV3 when K dim is smaller than native size #4700

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions lib/Analysis/Utility.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -488,6 +488,11 @@ bool supportMMA(triton::DotOp op, int version) {
if (triton::tools::getBoolEnv("DISABLE_MMA_V3"))
return false;
auto retType = op.getType();
RankedTensorType typeA = op.getA().getType();
int k = typeA.getShape().back();
// If k size is smaller than the native mma size, we cannot use MMA.
if (k < 256 / aElemTy.getIntOrFloatBitWidth())
return false;
auto retShapePerCTA = getShapePerCTA(retType);
auto rank = retShapePerCTA.size();
auto mod = op->getParentOfType<ModuleOp>();
Expand Down
17 changes: 17 additions & 0 deletions test/TritonGPU/accelerate-matmul.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,3 +142,20 @@ module attributes {"triton_gpu.target" = "cuda:90", "triton_gpu.num-ctas" = 1 :
tt.return
}
}


// -----

// Verify that we use mmav2 when the k dim is too small for mmav3.
// CHECK: #triton_gpu.nvidia_mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [8, 4], instrShape = [16, 8]}>
#blocked = #triton_gpu.blocked<{sizePerThread = [4, 4], threadsPerWarp = [1, 32], warpsPerCTA = [32, 1], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
module attributes {"triton_gpu.target" = "cuda:90", "triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 32 : i32, "triton_gpu.threads-per-warp" = 32 : i32} {
tt.func @small_k_size(
%a: tensor<128x16xf8E5M2, #triton_gpu.dot_op<{opIdx = 0, parent = #blocked}>>,
%b: tensor<16x128xf8E5M2, #triton_gpu.dot_op<{opIdx = 1, parent = #blocked}>>)
-> tensor<128x128xf32, #blocked> {
%zero_f32 = arith.constant dense<0.000000e+00> : tensor<128x128xf32, #blocked>
%result = tt.dot %a, %b, %zero_f32 : tensor<128x16xf8E5M2, #triton_gpu.dot_op<{opIdx = 0, parent = #blocked}>> * tensor<16x128xf8E5M2, #triton_gpu.dot_op<{opIdx = 1, parent = #blocked}>> -> tensor<128x128xf32, #blocked>
tt.return %result : tensor<128x128xf32, #blocked>
}
}
Loading