Skip to content

[AMD] Enable InThreadTranspose pass for RDNA3 / RDNA3.5 (gfx110x/115x) #10390

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

Merged
antiagainst merged 1 commit into from
May 27, 2026
Merged

[AMD] Enable InThreadTranspose pass for RDNA3 / RDNA3.5 (gfx110x/115x) #10390

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
23 changes: 23 additions & 0 deletions test/TritonGPU/amd/in-thread-transpose.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -546,3 +546,26 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 8 : i32, ttg.targ
tt.return
}
}

// -----
// RDNA3.5 (wave32) WMMA equivalent of inThreadTranspose_simple.

// CHECK-DAG: [[$TRANSPOSABLE:#blocked[0-9]*]] = #ttg.blocked<{sizePerThread = [4, 8], threadsPerWarp = [2, 16], warpsPerCTA = [2, 1], order = [1, 0]}>
// CHECK-DAG: [[$LINEAR:#linear[0-9]*]] = #ttg.linear<
// CHECK-LABEL: inThreadTranspose_wmma
// CHECK: tt.load {{.*}} : tensor<32x128x!tt.ptr<f16>, [[$TRANSPOSABLE]]>
// CHECK: amdg.in_thread_transpose {{.*}} -> tensor<32x128xf16, [[$LINEAR]]>
// CHECK: ttg.local_load {{.*}}-> tensor<32x128xf16, #ttg.dot_op<{opIdx = 1, parent = {{.*}}, kWidth = 16}>>
#blocked_wmma = #ttg.blocked<{sizePerThread = [1, 8], threadsPerWarp = [4, 8], warpsPerCTA = [2, 1], order = [1, 0]}>
#mma_wmma = #ttg.amd_wmma<{version = 1, isTranspose = true, ctaLayout = {warp = [[1, 0]]}}>
#shared_wmma = #ttg.swizzled_shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [1, 0]}>
#smem_wmma = #ttg.shared_memory
module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 2 : i32, ttg.target = "hip:gfx1151", "ttg.threads-per-warp" = 32 : i32} {
tt.func public @inThreadTranspose_wmma(%arg0: !tt.ptr<f16> {tt.divisibility = 16 : i32, tt.pointer_range = 32 : i32}) {
%0 = tt.splat %arg0 : !tt.ptr<f16> -> tensor<32x128x!tt.ptr<f16>, #blocked_wmma>
%1 = tt.load %0 : tensor<32x128x!tt.ptr<f16>, #blocked_wmma>
%2 = ttg.local_alloc %1 : (tensor<32x128xf16, #blocked_wmma>) -> !ttg.memdesc<32x128xf16, #shared_wmma, #smem_wmma>
%3 = ttg.local_load %2 : !ttg.memdesc<32x128xf16, #shared_wmma, #smem_wmma> -> tensor<32x128xf16, #ttg.dot_op<{opIdx = 1, parent = #mma_wmma, kWidth = 16}>>
tt.return
}
}
2 changes: 1 addition & 1 deletion third_party/amd/backend/compiler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ def is_pingpong_schedule_enabled(arch, use_async_copy):


def is_in_thread_transpose_enabled(arch):
return (arch == "gfx942" or "gfx120" in arch) \
return (arch == "gfx942" or "gfx110" in arch or "gfx115" in arch or "gfx120" in arch) \
if knobs.amd.use_in_thread_transpose is None else knobs.amd.use_in_thread_transpose


Expand Down
Loading