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

[OPTIMIZER] Made layout simplification pass efficient for fused attention kernels (#790)

Browse Source
This commit is contained in:
Philippe Tillet
2022-10-21 16:52:15 -07:00
committed by GitHub
parent c4726333bf
commit bb0f9235d1
26 changed files with 683 additions and 229 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
test/Analysis/test-alias.mlir
View File

@@ -24,7 +24,7 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
%b_ = tt.load %b_ptr, %b_mask, %b_other {cache = 1 : i32, evict = 1 : i32, isOtherUnspecified = false, isVolatile = false} : tensor<32x128xf16, #BL>
// CHECK-NEXT: %6 -> %6
%b = triton_gpu.convert_layout %b_ : (tensor<32x128xf16, #BL>) -> tensor<32x128xf16, #B>
%c = tt.dot %a, %b, %prev_c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%c = tt.dot %a, %b, %prev_c {allowTF32 = true, transA = false, transB = false} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%next_a_ptr = tt.addptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.addptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>

6
test/Analysis/test-allocation.mlir
View File

@@ -29,7 +29,7 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
// CHECK-NEXT: offset = 8192, size = 8192
%b = triton_gpu.convert_layout %b_ : (tensor<32x128xf16, #BL>) -> tensor<32x128xf16, #B>
%c = tt.dot %a, %b, %prev_c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%c = tt.dot %a, %b, %prev_c {allowTF32 = true, transA = false, transB = false} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%next_a_ptr = tt.addptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.addptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>
@@ -59,11 +59,11 @@ func @reusable(%A : !tt.ptr<f16>) {
%a3_ = tt.load %a_ptr, %cst1, %cst2 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x32xf16, #AL>
// 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>
%c = tt.dot %a1, %a2, %c_init {allowTF32 = true, transA = false, transB = false} : 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: 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>
%c1 = tt.dot %a3, %a4, %c {allowTF32 = true, transA = false, transB = false} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
return
// CHECK-NEXT: size = 24576
}

2
test/Analysis/test-membar.mlir
View File

@@ -27,7 +27,7 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
%b_ = tt.load %b_ptr, %b_mask, %b_other {cache = 1 : i32, evict = 1 : i32, isOtherUnspecified = false, isVolatile = false} : tensor<32x128xf16, #BL>
%b = triton_gpu.convert_layout %b_ : (tensor<32x128xf16, #BL>) -> tensor<32x128xf16, #B>
// CHECK: Membar 13
%c = tt.dot %a, %b, %prev_c {allowTF32 = true} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%c = tt.dot %a, %b, %prev_c {allowTF32 = true, transA = false, transB = false} : tensor<128x32xf16, #A> * tensor<32x128xf16, #B> -> tensor<128x128xf32, #C>
%next_a_ptr = tt.addptr %a_ptr, %a_off : tensor<128x32x!tt.ptr<f16>, #AL>
%next_b_ptr = tt.addptr %b_ptr, %b_off : tensor<32x128x!tt.ptr<f16>, #BL>