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

[Triton-IR] Added type inference and verifier for Triton-IR operations (#767)

Browse Source
This commit is contained in:
Philippe Tillet
2022-10-11 18:16:41 -07:00
committed by GitHub
parent b6e5a231e5
commit 623c99609f
27 changed files with 494 additions and 348 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -348,15 +348,17 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#block1 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [8], warpsPerCTA = [4], order = [0]}>
#block2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [4, 1], warpsPerCTA = [4, 1], order = [1, 0]}>
#block3 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 8], warpsPerCTA = [1, 4], order = [1, 0]}>
#slice2d1 = #triton_gpu.slice<{dim = 1, parent=#block2}>
#slice3d0 = #triton_gpu.slice<{dim = 0, parent=#block3}>
#AL = #triton_gpu.blocked<{sizePerThread = [1, 8], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 4, perPhase = 1, maxPhase = 4, order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: basic_insert_slice_async_v4
func @basic_insert_slice_async_v4(%arg0: !tt.ptr<f32> {tt.divisibility = 4 : i32}) {
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #block0>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<64xi32, #block1>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #block0>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<64xi32, #block1>) -> tensor<1x64xi32, #block3>
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #slice2d1>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<64xi32, #slice3d0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #slice2d1>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<64xi32, #slice3d0>) -> tensor<1x64xi32, #block3>
%broadcast_off0_scalar = tt.broadcast %off0 : (tensor<16x1xi32, #block2>) -> tensor<16x64xi32, #block2>
%cst_scalar = arith.constant 64 : i32
%cst = tt.splat %cst_scalar : (i32) -> tensor<16x64xi32, #block2>
@@ -387,15 +389,17 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#block1 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [8], warpsPerCTA = [4], order = [0]}>
#block2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [4, 1], warpsPerCTA = [4, 1], order = [1, 0]}>
#block3 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 8], warpsPerCTA = [1, 4], order = [1, 0]}>
#slice2d1 = #triton_gpu.slice<{dim = 1, parent=#block2}>
#slice3d0 = #triton_gpu.slice<{dim = 0, parent=#block3}>
#AL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 4, order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: basic_insert_slice_async_v1
func @basic_insert_slice_async_v1(%arg0: !tt.ptr<f32> {tt.divisibility = 4 : i32}) {
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #block0>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #block1>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #block0>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #block1>) -> tensor<1x32xi32, #block3>
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #slice2d1>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #slice3d0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #slice2d1>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #slice3d0>) -> tensor<1x32xi32, #block3>
%broadcast_off0_scalar = tt.broadcast %off0 : (tensor<16x1xi32, #block2>) -> tensor<16x32xi32, #block2>
%cst_scalar = arith.constant 32 : i32
%cst = tt.splat %cst_scalar : (i32) -> tensor<16x32xi32, #block2>
@@ -429,15 +433,17 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#block0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [8], warpsPerCTA = [4], order = [0]}>
#block2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [8, 1], warpsPerCTA = [4, 1], order = [1, 0]}>
#block3 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 8], warpsPerCTA = [1, 4], order = [1, 0]}>
#slice2d1 = #triton_gpu.slice<{dim = 1, parent=#block2}>
#slice3d0 = #triton_gpu.slice<{dim = 0, parent=#block3}>
#AL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 4, order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: basic_insert_slice_async_v1_multictas
func @basic_insert_slice_async_v1_multictas(%arg0: !tt.ptr<f32> {tt.divisibility = 4 : i32}) {
%off0_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #block0>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #block0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<32xi32, #block0>) -> tensor<32x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #block0>) -> tensor<1x32xi32, #block3>
%off0_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #slice2d1>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #slice3d0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<32xi32, #slice2d1>) -> tensor<32x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #slice3d0>) -> tensor<1x32xi32, #block3>
%broadcast_off0_scalar = tt.broadcast %off0 : (tensor<32x1xi32, #block2>) -> tensor<32x32xi32, #block2>
%cst_scalar = arith.constant 32 : i32
%cst = tt.splat %cst_scalar : (i32) -> tensor<32x32xi32, #block2>