gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[Triton-MLIR][Backend] Minor fix for allocation and backend in handling tt.ptr tensors (#878)

Browse Source
This commit is contained in:
goostavz
2022-11-15 18:08:07 +08:00
committed by GitHub
parent a22ff39017
commit 37f5846280
3 changed files with 37 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -742,6 +742,7 @@ module attributes {"triton_gpu.num-warps" = 1 : i32} {
}
// -----
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [1], order = [0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [1, 1], order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 1 : i32} {
@@ -753,8 +754,8 @@ module attributes {"triton_gpu.num-warps" = 1 : i32} {
}
}
// -----
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [1], order = [0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [1, 1], order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 1 : i32} {
@@ -766,8 +767,25 @@ module attributes {"triton_gpu.num-warps" = 1 : i32} {
}
}
// -----
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [1], order = [0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [1], order = [0]}>
module attributes {"triton_gpu.num-warps" = 1 : i32} {
// CHECK-LABEL: convert_blocked_to_blocked_ptr
func @convert_blocked_to_blocked_ptr(%src:tensor<32x!tt.ptr<f32>, #blocked0>) {
// CHECK: llvm.ptrtoint
// CHECK: llvm.store
// CHECK: nvvm.barrier0
// CHECK: llvm.inttoptr
// CHECK-COUNT-4: llvm.insertvalue
%cvt = triton_gpu.convert_layout %src : (tensor<32x!tt.ptr<f32>, #blocked0>) -> tensor<32x!tt.ptr<f32>, #blocked1>
return
}
}
// -----
#blocked = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [2, 16], warpsPerCTA = [1, 4], order = [1, 0]}>
#shared = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [1, 0]}>
#mma = #triton_gpu.mma<{version = 2, warpsPerCTA = [2, 2]}>
@@ -839,6 +857,7 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
}
// -----
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: atomic_add_f32