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

[Triton-MLIR][BACKEND] add convert_layout[shared->dot_op] converstion to adapt DotOperand layout (#786)

Browse Source 
This PR helps to

1. Adapt the existing DotOp conversion to the design of the new
DotOperand layout,
2. Making the DotOp conversion work with both shared-layout inputs case
and dotoperand-layout inputs case for further upstream switch.
This commit is contained in:
Yan Chunwei
2022-10-24 11:40:13 +08:00
committed by GitHub
parent 3aa8296b06
commit 877844de4f
4 changed files with 241 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -714,3 +714,27 @@ module attributes {"triton_gpu.num-warps" = 1 : i32} {
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]}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mma}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mma}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
func @matmul_kernel_dot_operand_layout(%ptr:!tt.ptr<f32> {tt.divisibility = 16 : i32},
%a:tensor<128x32xf16, #shared>, %b:tensor<32x256xf16, #shared>) {
%cst = arith.constant dense<0.000000e+00> : tensor<128x256xf32, #mma>
// CHECK: ldmatrix.sync.aligned.m8n8.x4.shared.b16
%a_mat = triton_gpu.convert_layout %a : (tensor<128x32xf16, #shared>) -> tensor<128x32xf16, #dot_operand_a>
%b_mat = triton_gpu.convert_layout %b : (tensor<32x256xf16, #shared>) -> tensor<32x256xf16, #dot_operand_b>
%28 = tt.dot %a_mat, %b_mat, %cst {allowTF32 = true, transA = false, transB = false} : tensor<128x32xf16, #dot_operand_a> * tensor<32x256xf16, #dot_operand_b> -> tensor<128x256xf32, #mma>
%38 = triton_gpu.convert_layout %28 : (tensor<128x256xf32, #mma>) -> tensor<128x256xf32, #blocked>
%30 = tt.splat %ptr : (!tt.ptr<f32>) -> tensor<128x1x!tt.ptr<f32>, #blocked>
%36 = tt.broadcast %30 : (tensor<128x1x!tt.ptr<f32>, #blocked>) -> tensor<128x256x!tt.ptr<f32>, #blocked>
tt.store %36, %38 : tensor<128x256xf32, #blocked>
return
}
}