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

[Triton-MLIR][BACKEND] adapt DotOp layout for FMADot (#872)

Browse Source
This commit is contained in:
Yan Chunwei
2022-11-14 16:56:30 +08:00
committed by GitHub
parent 516a241234
commit 1eedaf7bec
4 changed files with 482 additions and 253 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -820,18 +820,20 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#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}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#blocked}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#blocked}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
func @matmul_fmadot(%ptr:!tt.ptr<f32> {tt.divisibility = 16 : i32},
%a:tensor<32x16xf32, #shared>, %b:tensor<16x32xf32, #shared>) {
// We are going to completely depracate using shared layout for operands of dot
//%cst = arith.constant dense<0.000000e+00> : tensor<32x32xf32, #blocked>
//%28 = tt.dot %a, %b, %cst {allowTF32 = false, transA = false, transB = false} : tensor<32x16xf32, #shared> * tensor<16x32xf32, #shared> -> tensor<32x32xf32, #blocked>
//%30 = tt.splat %ptr : (!tt.ptr<f32>) -> tensor<32x1x!tt.ptr<f32>, #blocked>
//%36 = tt.broadcast %30 : (tensor<32x1x!tt.ptr<f32>, #blocked>) -> tensor<32x32x!tt.ptr<f32>, #blocked>
//tt.store %36, %28 : tensor<32x32xf32, #blocked>
%cst = arith.constant dense<0.000000e+00> : tensor<32x32xf32, #blocked>
// CHECK: llvm.intr.fmuladd
%a_mat = triton_gpu.convert_layout %a : (tensor<32x16xf32, #shared>) -> tensor<32x16xf32, #dot_operand_a>
%b_mat = triton_gpu.convert_layout %b : (tensor<16x32xf32, #shared>) -> tensor<16x32xf32, #dot_operand_b>
%28 = tt.dot %a_mat, %b_mat, %cst {allowTF32 = false, transA = false, transB = false} : tensor<32x16xf32, #dot_operand_a> * tensor<16x32xf32, #dot_operand_b> -> tensor<32x32xf32, #blocked>
%30 = tt.splat %ptr : (!tt.ptr<f32>) -> tensor<32x1x!tt.ptr<f32>, #blocked>
%36 = tt.broadcast %30 : (tensor<32x1x!tt.ptr<f32>, #blocked>) -> tensor<32x32x!tt.ptr<f32>, #blocked>
tt.store %36, %28 : tensor<32x32xf32, #blocked>
return
}
}
@@ -846,4 +848,4 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
%0 = "tt.atomic_rmw" (%arg0, %arg2, %arg1) {atomic_rmw_op = 5 : i32} : (tensor<256x!tt.ptr<f32>, #blocked0>, tensor<256xf32, #blocked0>, tensor<256xi1, #blocked0>) -> tensor<256xf32, #blocked0>
return
}
}
}