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

[BACKEND] add dot conversion (mma version=2) (#672)

Browse Source 
LLVM Conversion for Dot op.

Due to the lack of `convert_layout`, currently, the dot only supports
the following combination of operands

- `$a` in shared layout
- `$b` in shared layout
- `$c` in MMA layout(but only Splat-like, leaving the generic cases to
`convert_layout`)

This PR focus on `mma.16816` related logic support, leaving the other
cases to the following PR.

Co-authored-by: Philippe Tillet <phil@openai.com>
This commit is contained in:
Yan Chunwei
2022-09-23 11:43:54 +08:00
committed by GitHub
parent 23f424c660
commit 922155f1d2
7 changed files with 1033 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
test/Analysis/test-allocation.mlir
View File

@@ -22,9 +22,11 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
scf.for %iv = %lb to %ub step %step iter_args(%a_ptr = %a_ptr_init, %b_ptr = %b_ptr_init, %prev_c = %c_init) -> (tensor<128x32x!tt.ptr<f16>, #AL>, tensor<32x128x!tt.ptr<f16>, #BL>, tensor<128x128xf32, #C>) {
%a_ = tt.load %a_ptr, %a_mask, %a_other {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x32xf16, #AL>
// CHECK: offset = 0, size = 8192
// CHECK: scratch offset = 8192, size = 0
// CHECK-NEXT: offset = 0, size = 8192
%a = triton_gpu.convert_layout %a_ : (tensor<128x32xf16, #AL>) -> tensor<128x32xf16, #A>
%b_ = tt.load %b_ptr, %b_mask, %b_other {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<32x128xf16, #BL>
// CHECK-NEXT: scratch offset = 16384, size = 0
// CHECK-NEXT: offset = 8192, size = 8192
%b = triton_gpu.convert_layout %b_ : (tensor<32x128xf16, #BL>) -> tensor<32x128xf16, #B>
@@ -50,16 +52,20 @@ func @reusable(%A : !tt.ptr<f16>) {
%a_ptr = tt.broadcast %A : (!tt.ptr<f16>) -> tensor<128x32x!tt.ptr<f16>, #AL>
%b_ptr = tt.broadcast %A : (!tt.ptr<f16>) -> tensor<32x128x!tt.ptr<f16>, #AL>
%a1_ = tt.load %a_ptr, %cst1, %cst2 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x32xf16, #AL>
// CHECK: offset = 0, size = 8192
// CHECK: scratch offset = 8192, size = 0
// CHECK-NEXT: offset = 0, size = 8192
%a1 = triton_gpu.convert_layout %a1_ : (tensor<128x32xf16, #AL>) -> tensor<128x32xf16, #A>
%a2_ = tt.load %b_ptr, %cst3, %cst4 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<32x128xf16, #AL>
// CHECK-NEXT: scratch offset = 16384, size = 0
// CHECK-NEXT: offset = 8192, size = 8192
%a2 = triton_gpu.convert_layout %a2_ : (tensor<32x128xf16, #AL>) -> tensor<32x128xf16, #A>
%a3_ = tt.load %a_ptr, %cst1, %cst2 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x32xf16, #AL>
// CHECK-NEXT: scratch offset = 24576, size = 0
// CHECK-NEXT: offset = 16384, size = 8192
%a3 = triton_gpu.convert_layout %a3_ : (tensor<128x32xf16, #AL>) -> tensor<128x32xf16, #A>
%c = tt.dot %a1, %a2, %c_init {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%a4_ = tt.load %b_ptr, %cst3, %cst4 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<32x128xf16, #AL>
// CHECK-NEXT: scratch offset = 8192, size = 0
// CHECK-NEXT: offset = 0, size = 8192
%a4 = triton_gpu.convert_layout %a4_ : (tensor<32x128xf16, #AL>) -> tensor<32x128xf16, #A>
%c1 = tt.dot %a3, %a4, %c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>