Skip to content

[CI] Adding non exec WG-level cases running the whole compilation pipeline #1159

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Garra1980 merged 2 commits into from
Mar 19, 2026
Merged

[CI] Adding non exec WG-level cases running the whole compilation pipeline #1159

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1174700
[CI] Adding non exec WG-level cases that test the whole compilation p…
Garra1980 Mar 18, 2026
360929a
fix pre-commit
Garra1980 Mar 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
74 changes: 74 additions & 0 deletions test/Integration/Dialect/XeGPU/WG/non_exec_non_gemm_1.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,74 @@
// RUN: imex-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup"

module attributes {gpu.container_module} {
func.func @main(%arg0: memref<32x2x192xf32>, %arg1: memref<32x2x192xf32>, %arg2: memref<32x2x192xf32>) attributes {L2Mem = 0 : i64, gemm_tiles_x = dense<1> : vector<4xi64>, gemm_tiles_y = dense<[1, -1, 16, 24]> : vector<4xi64>, habana_runner.num_inputs = 2 : i64, habana_runner.tests = [{inputs = [dense<2.000000e+00> : tensor<32x2x192xf32>, dense<3.000000e+00> : tensor<32x2x192xf32>], outputs = [dense<2.000000e+00> : tensor<32x2x192xf32>]}], linear_block_size = array<i32: 1024, 1, 1>, linear_grid_size = array<i32: 16, 1, 1>, region_partition = 1 : i64, region_size = 16 : i64, syn.fusion_successful, syn.gemm_pipeline, syn.tensor_signature = (tensor<32x2x192xf32>, tensor<32x2x192xf32>) -> tensor<32x2x192xf32>, synFusionGenOps = 1 : i64, synFusionRequiredBeamSize = 1 : i64, synFusionTotalCost = 1.628800e+03 : f64} {
%c2 = arith.constant 2 : index
%cst = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} dense<true> : vector<2x2x6x32xi1>
%cst_0 = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 1, 3]>} dense<32> : vector<6xindex>
%cst_1 = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} dense<384> : vector<2xindex>
%cst_2 = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 2, 3]>} dense<192> : vector<2xindex>
%c16 = arith.constant 16 : index
%c768 = arith.constant 768 : index
%c1 = arith.constant 1 : index
gpu.launch_func @main_kernel::@main_kernel blocks in (%c16, %c1, %c1) threads in (%c768, %c1, %c1) args(%arg0 : memref<32x2x192xf32>, %arg1 : memref<32x2x192xf32>, %arg2 : memref<32x2x192xf32>)
return
}
gpu.module @main_kernel {
gpu.func @main_kernel(%arg0: memref<32x2x192xf32>, %arg1: memref<32x2x192xf32>, %arg2: memref<32x2x192xf32>) kernel attributes {known_block_size = array<i32: 768, 1, 1>, known_grid_size = array<i32: 16, 1, 1>} {
%block_id_x = gpu.block_id x
%block_id_y = gpu.block_id y
%block_id_z = gpu.block_id z
%thread_id_x = gpu.thread_id x
%thread_id_y = gpu.thread_id y
%thread_id_z = gpu.thread_id z
%grid_dim_x = gpu.grid_dim x
%grid_dim_y = gpu.grid_dim y
%grid_dim_z = gpu.grid_dim z
%block_dim_x = gpu.block_dim x
%block_dim_y = gpu.block_dim y
%block_dim_z = gpu.block_dim z
%c16 = arith.constant 16 : index
%c2 = arith.constant 2 : index
%cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 1, 3]>} dense<32> : vector<6xindex>
%cst_0 = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} dense<384> : vector<2xindex>
%cst_1 = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 2, 3]>} dense<192> : vector<2xindex>
%cst_2 = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} dense<true> : vector<2x2x6x32xi1>
%block_id_x_3 = gpu.block_id x
%c4 = arith.constant 4 : index
%0 = arith.shrsi %block_id_x_3, %c4 : index
%1 = arith.muli %0, %c16 : index
%2 = arith.subi %block_id_x_3, %1 : index
%3 = arith.muli %2, %c2 overflow<nsw> : index
%4 = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} : vector<2xindex>
%5 = vector.broadcast %3 {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} : index to vector<2xindex>
%6 = arith.addi %5, %4 {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} : vector<2xindex>
%7 = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 2, 3]>} : vector<2xindex>
%8 = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 1, 3]>} : vector<6xindex>
%9 = arith.muli %8, %cst {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 1, 3]>} : vector<6xindex>
%10 = vector.shape_cast %9 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>} : vector<6xindex> to vector<1x1x6x1xindex>
%11 = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>, dims = [0, 1, 2]>} : vector<32xindex>
%12 = vector.broadcast %10 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<1x1x6x1xindex> to vector<2x2x6x32xindex>
%13 = vector.broadcast %11 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<32xindex> to vector<2x2x6x32xindex>
%14 = arith.addi %12, %13 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<2x2x6x32xindex>
%15 = arith.muli %6, %cst_0 {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} : vector<2xindex>
%16 = vector.shape_cast %15 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>} : vector<2xindex> to vector<2x1x1x1xindex>
%17 = vector.broadcast %16 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<2x1x1x1xindex> to vector<2x2x6x32xindex>
%18 = arith.muli %7, %cst_1 {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [0, 2, 3]>} : vector<2xindex>
%19 = vector.shape_cast %18 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 1]>} : vector<2xindex> to vector<1x2x1x1xindex>
%20 = vector.broadcast %19 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<1x2x1x1xindex> to vector<2x2x6x32xindex>
%21 = arith.addi %17, %20 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<2x2x6x32xindex>
%22 = arith.addi %21, %14 {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<2x2x6x32xindex>
%intptr = memref.extract_aligned_pointer_as_index %arg0 : memref<32x2x192xf32> -> index
%23 = arith.index_cast %intptr : index to i64
%24 = xegpu.load %23[%22], %cst_2 {layout = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : i64, vector<2x2x6x32xindex>, vector<2x2x6x32xi1> -> vector<2x2x6x32xf32>
%intptr_4 = memref.extract_aligned_pointer_as_index %arg1 : memref<32x2x192xf32> -> index
%25 = arith.index_cast %intptr_4 : index to i64
%26 = xegpu.load %25[%22], %cst_2 {layout = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : i64, vector<2x2x6x32xindex>, vector<2x2x6x32xi1> -> vector<2x2x6x32xf32>
%27 = arith.minimumf %24, %26 fastmath<nnan> {layout_result_0 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<2x2x6x32xf32>
%intptr_5 = memref.extract_aligned_pointer_as_index %arg2 : memref<32x2x192xf32> -> index
%28 = arith.index_cast %intptr_5 : index to i64
xegpu.store %27, %28[%22], %cst_2 {layout = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>, layout_operand_2 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>, layout_operand_3 = #xegpu.layout<sg_layout = [2, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : vector<2x2x6x32xf32>, i64, vector<2x2x6x32xindex>, vector<2x2x6x32xi1>
gpu.return
}
}
}
54 changes: 54 additions & 0 deletions test/Integration/Dialect/XeGPU/WG/non_exec_non_gemm_2.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,54 @@
// RUN: imex-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup"

module attributes {gpu.container_module} {
func.func @unaligned_entry(%arg0: memref<50xf32>, %arg1: memref<50xf32>, %arg2: memref<50xf32>) attributes {L2Mem = 1024 : i64, gemm_tiles_x = dense<1> : vector<4xi64>, gemm_tiles_y = dense<[1, -1, 1, 2]> : vector<4xi64>, habana_runner.tests = [{inputs = [-2.000000e+00 : f32, 2.000000e+00 : f32], outputs = [1.000000e+00 : f32], tolerance = [1.000000e-02]}], linear_block_size = array<i32: 1024, 1, 1>, linear_grid_size = array<i32: 1, 1, 1>, region_partition = 1 : i64, region_size = 1 : i64, syn.fusion_successful, syn.gemm_pipeline, syn.tensor_signature = (tensor<50xf32>, tensor<50xf32>) -> tensor<50xf32>, synFusionGenOps = 1 : i64, synFusionRequiredBeamSize = 1 : i64, synFusionTotalCost = 144.66249999999999 : f64} {
%cst = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} dense<50> : vector<2x32xindex>
%cst_0 = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} dense<0.000000e+00> : vector<2x32xf32>
%cst_1 = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 1], sg_data = [1, 1]>, dims = [1]>} dense<32> : vector<2xindex>
%c1 = arith.constant 1 : index
%c64 = arith.constant 64 : index
gpu.launch_func @unaligned::@unaligned blocks in (%c1, %c1, %c1) threads in (%c64, %c1, %c1) args(%arg0 : memref<50xf32>, %arg1 : memref<50xf32>, %arg2 : memref<50xf32>)
return
}
gpu.module @unaligned {
gpu.func @unaligned(%arg0: memref<50xf32>, %arg1: memref<50xf32>, %arg2: memref<50xf32>) kernel attributes {known_block_size = array<i32: 64, 1, 1>, known_grid_size = array<i32: 1, 1, 1>} {
%block_id_x = gpu.block_id x
%block_id_y = gpu.block_id y
%block_id_z = gpu.block_id z
%thread_id_x = gpu.thread_id x
%thread_id_y = gpu.thread_id y
%thread_id_z = gpu.thread_id z
%grid_dim_x = gpu.grid_dim x
%grid_dim_y = gpu.grid_dim y
%grid_dim_z = gpu.grid_dim z
%block_dim_x = gpu.block_dim x
%block_dim_y = gpu.block_dim y
%block_dim_z = gpu.block_dim z
%cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 1], sg_data = [1, 1]>, dims = [1]>} dense<32> : vector<2xindex>
%cst_0 = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} dense<50> : vector<2x32xindex>
%cst_1 = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} dense<0.000000e+00> : vector<2x32xf32>
%0 = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 1], sg_data = [1, 1]>, dims = [1]>} : vector<2xindex>
%1 = arith.muli %0, %cst {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 1], sg_data = [1, 1]>, dims = [1]>} : vector<2xindex>
%2 = vector.shape_cast %1 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 1]>} : vector<2xindex> to vector<2x1xindex>
%3 = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>, dims = [0]>} : vector<32xindex>
%4 = vector.broadcast %2 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x1xindex> to vector<2x32xindex>
%5 = vector.broadcast %3 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<32xindex> to vector<2x32xindex>
%6 = arith.addi %4, %5 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xindex>
%7 = arith.cmpi slt, %6, %cst_0 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xindex>
%intptr = memref.extract_aligned_pointer_as_index %arg0 : memref<50xf32> -> index
%8 = arith.index_cast %intptr : index to i64
%9 = xegpu.load %8[%6], %7 {layout = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
%10 = arith.select %7, %9, %cst_1 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xi1>, vector<2x32xf32>
%intptr_2 = memref.extract_aligned_pointer_as_index %arg1 : memref<50xf32> -> index
%11 = arith.index_cast %intptr_2 : index to i64
%12 = xegpu.load %11[%6], %7 {layout = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
%13 = arith.select %7, %12, %cst_1 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xi1>, vector<2x32xf32>
%14 = arith.addf %10, %13 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xf32>
%15 = math.cos %14 {layout_result_0 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xf32>
%intptr_3 = memref.extract_aligned_pointer_as_index %arg2 : memref<50xf32> -> index
%16 = arith.index_cast %intptr_3 : index to i64
xegpu.store %15, %16[%6], %7 {layout = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>, layout_operand_2 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>, layout_operand_3 = #xegpu.layout<sg_layout = [2, 1], sg_data = [1, 32]>} : vector<2x32xf32>, i64, vector<2x32xindex>, vector<2x32xi1>
gpu.return
}
}
}
Loading
Loading