aboutsummaryrefslogtreecommitdiff
path: root/mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir
diff options
context:
space:
mode:
Diffstat (limited to 'mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir')
-rw-r--r--mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir139
1 files changed, 67 insertions, 72 deletions
diff --git a/mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir b/mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir
index d67bdb4..628a485 100644
--- a/mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir
+++ b/mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-rr.mlir
@@ -2,122 +2,117 @@
gpu.module @test_round_robin_assignment {
// CHECK-LABEL: create_nd_tdesc
- // CHECK-SAME: %[[ARG_0:.*]]: memref<24x32xf32>
- gpu.func @create_nd_tdesc(%src: memref<24x32xf32>) {
- // CHECK-COUNT-12: xegpu.create_nd_tdesc %[[ARG_0]][%{{.*}}, %{{.*}}] : memref<24x32xf32>
- // CHECK-SAME: -> !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
+ // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>
+ gpu.func @create_nd_tdesc(%src: memref<256x128xf32>) {
+ // CHECK-COUNT-4: xegpu.create_nd_tdesc %[[ARG_0]][%{{.*}}, %{{.*}}] : memref<256x128xf32>
+ // CHECK-SAME: -> !xegpu.tensor_desc<16x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
// CHECK-NOT: xegpu.create_nd_tdesc
- %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32>
- -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<256x128xf32>
+ -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
gpu.return
}
// CHECK-LABEL: load_nd_tdesc
- // CHECK-SAME: %[[ARG_0:.*]]: memref<24x32xf32>
- gpu.func @load_nd_tdesc(%src: memref<24x32xf32>) {
- %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32>
- -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- // CHECK-COUNT-12: xegpu.load_nd %{{.*}}
- // CHECK-SAME-COUNT-12: : !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
- // CHECK-SAME-COUNT-12: -> vector<2x2xf32>
+ // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>
+ gpu.func @load_nd_tdesc(%src: memref<256x128xf32>) {
+ %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<256x128xf32>
+ -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
+ // CHECK-COUNT-4: xegpu.load_nd %{{.*}}
+ // CHECK-SAME-COUNT-4: : !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
+ // CHECK-SAME-COUNT-4: -> vector<16x16xf32>
// CHECK-NOT: xegpu.load_nd
%load = xegpu.load_nd %tdesc
- : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- -> vector<24x32xf32>
+ : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
+ -> vector<256x128xf32>
gpu.return
}
// CHECK-LABEL: store_nd
- // CHECK-SAME: %[[ARG_0:.*]]: memref<24x32xf32>
- gpu.func @store_nd(%src: memref<24x32xf32>) {
- %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32>
- -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- // CHECK-COUNT-12: xegpu.store_nd %{{.*}}, %{{.*}}
- // CHECK-SAME-COUNT-12: : vector<2x2xf32>, !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
+ // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>
+ gpu.func @store_nd(%src: memref<256x128xf32>) {
+ %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<256x128xf32>
+ -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
+ // CHECK-COUNT-4: xegpu.store_nd %{{.*}}, %{{.*}}
+ // CHECK-SAME-COUNT-4: : vector<16x16xf32>, !xegpu.tensor_desc<16x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
// CHECK-NOT : xegpu.store_nd
%load = xegpu.load_nd %tdesc
- : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- -> vector<24x32xf32>
+ : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
+ -> vector<256x128xf32>
xegpu.store_nd %load, %tdesc
- : vector<24x32xf32>, !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ : vector<256x128xf32>, !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
gpu.return
}
// CHECK-LABEL: update_nd
- // CHECK-SAME: %[[ARG_0:.*]]: memref<24x32xf32>
- gpu.func @update_nd(%src: memref<24x32xf32>){
- %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32>
- -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- // CHECK-COUNT-12: xegpu.update_nd_offset %{{.*}}, [0, 16]
- // CHECK-SAME-COUNT-12: : !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
+ // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>
+ gpu.func @update_nd(%src: memref<256x128xf32>){
+ %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<256x128xf32>
+ -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
+ // CHECK-COUNT-4: xegpu.update_nd_offset %{{.*}}, [0, 16]
+ // CHECK-SAME-COUNT-4: : !xegpu.tensor_desc<16x16xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>>
// CHECK-NOT: xegpu.update_nd_offset
%update = xegpu.update_nd_offset %tdesc, [0, 16]
- : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
gpu.return
}
// CHECK-LABEL: dpas
- // CHECK-SAME: (%[[ARG_0:.*]]: memref<8x8xf32>, %[[ARG_1:.*]]: memref<8x8xf32>, %[[ARG_2:.*]]: memref<8x8xf32>)
- gpu.func @dpas(%a: memref<8x8xf32>, %b: memref<8x8xf32>, %c: memref<8x8xf32>) {
- // CHECK-COUNT-4: xegpu.create_nd_tdesc %[[ARG_0]][%{{.*}}, %{{.*}}] : memref<8x8xf32>
- // CHECK-SAME-COUNT-4: -> !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
+ // CHECK-SAME: (%[[ARG_0:.*]]: memref<256x128xf16>, %[[ARG_1:.*]]: memref<128x256xf16>)
+ gpu.func @dpas(%a: memref<256x128xf16>, %b: memref<128x256xf16>) {
+ // CHECK-COUNT-4: xegpu.create_nd_tdesc %[[ARG_0]][%{{.*}}, %{{.*}}] : memref<256x128xf16>
+ // CHECK-SAME-COUNT-4: -> !xegpu.tensor_desc<16x16xf16, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
// CHECK-NOT: xegpu.create_nd_tdesc
- // CHECK-COUNT-4: xegpu.create_nd_tdesc %[[ARG_1]][%{{.*}}, %{{.*}}] : memref<8x8xf32>
- // CHECK-SAME-COUNT-4: -> !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
- // CHECK-NOT: xegpu.create_nd_tdesc
- // CHECK-COUNT-4: xegpu.create_nd_tdesc %{{.*}}[%{{.*}}, %{{.*}}] : memref<8x8xf32>
- // CHECK-SAME-COUNT-4: -> !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
+ // CHECK-COUNT-4: xegpu.create_nd_tdesc %[[ARG_1]][%{{.*}}, %{{.*}}] : memref<128x256xf16>
+ // CHECK-SAME-COUNT-4: -> !xegpu.tensor_desc<16x16xf16, #xegpu.layout<lane_layout = [4, 8], lane_data = [1, 1]>>
// CHECK-NOT: xegpu.create_nd_tdesc
// CHECK-COUNT-16: xegpu.dpas %{{.*}}, %{{.*}}
- // CHECK-SAME-COUNT-16: {layout = #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>}
- // CHECK-SAME-COUNT-16: : vector<2x2xf32>, vector<2x2xf32> -> vector<2x2xf32>
+ // CHECK-SAME-COUNT-16: {layout = #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>}
+ // CHECK-SAME-COUNT-16: : vector<16x16xf16>, vector<16x16xf16> -> vector<16x16xf32>
// CHECK-NOT: xegpu.dpas
- %tdesc_a = xegpu.create_nd_tdesc %a[0, 0] : memref<8x8xf32>
- -> !xegpu.tensor_desc<8x8xf32, #xegpu.layout<sg_layout = [2, 2], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ %tdesc_a = xegpu.create_nd_tdesc %a[0, 0] : memref<256x128xf16>
+ -> !xegpu.tensor_desc<256x128xf16, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
%load_a = xegpu.load_nd %tdesc_a
- : !xegpu.tensor_desc<8x8xf32, #xegpu.layout<sg_layout = [2, 2], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- -> vector<8x8xf32>
- %tdesc_b = xegpu.create_nd_tdesc %b[0, 0] : memref<8x8xf32>
- -> !xegpu.tensor_desc<8x8xf32, #xegpu.layout<sg_layout = [2, 2], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ : !xegpu.tensor_desc<256x128xf16, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
+ -> vector<256x128xf16>
+ %tdesc_b = xegpu.create_nd_tdesc %b[0, 0] : memref<128x256xf16>
+ -> !xegpu.tensor_desc<128x256xf16, #xegpu.layout<sg_layout = [4, 8], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [2, 1]>>
%load_b = xegpu.load_nd %tdesc_b
- : !xegpu.tensor_desc<8x8xf32, #xegpu.layout<sg_layout = [2, 2], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
- -> vector<8x8xf32>
- %tdesc_c = xegpu.create_nd_tdesc %c[0, 0] : memref<8x8xf32>
- -> !xegpu.tensor_desc<8x8xf32, #xegpu.layout<sg_layout = [2, 2], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ : !xegpu.tensor_desc<128x256xf16, #xegpu.layout<sg_layout = [4, 8], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [2, 1]>>
+ -> vector<128x256xf16>
%dpas = xegpu.dpas %load_a, %load_b
- {layout_result_0 = #xegpu.layout<sg_layout = [2, 2], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>}
- : vector<8x8xf32>, vector<8x8xf32> -> vector<8x8xf32>
+ {layout_result_0 = #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>}
+ : vector<256x128xf16>, vector<128x256xf16> -> vector<256x256xf32>
gpu.return
}
// CHECK-LABEL: prefetch_nd_tdesc
- // CHECK-SAME: %[[ARG_0:.*]]: memref<24x32xf32>
- gpu.func @prefetch_nd_tdesc(%src: memref<24x32xf32>) {
- // CHECK-COUNT-12: xegpu.prefetch_nd %{{.*}}
- // CHECK-SAME-COUNT-12 : !xegpu.tensor_desc<2x2xf32, #xegpu.layout<lane_layout = [2, 2], lane_data = [1, 1]>>
+ // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>
+ gpu.func @prefetch_nd_tdesc(%src: memref<256x128xf32>) {
+ // CHECK-COUNT-4: xegpu.prefetch_nd %{{.*}}
+ // CHECK-SAME-COUNT-4: !xegpu.tensor_desc<256x128xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>
// CHECK-NOT: xegpu.prefetch_nd
- %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32>
- -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<256x128xf32>
+ -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
xegpu.prefetch_nd %tdesc
- : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2], lane_layout = [2, 2], lane_data = [1, 1]>>
+ : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>>
gpu.return
}
// CHECK-LABEL: broadcast
- // CHECK-SAME: %[[ARG_0:.*]]: memref<24x1xf32>
- gpu.func @broadcast(%src: memref<24x1xf32>) {
- %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x1xf32>
- -> !xegpu.tensor_desc<24x1xf32, #xegpu.layout<sg_layout = [4, 1], sg_data = [2, 1], lane_layout = [2, 1], lane_data = [1, 1]>>
+ // CHECK-SAME: %[[ARG_0:.*]]: memref<128x1xf32>
+ gpu.func @broadcast(%src: memref<128x1xf32>) {
+ %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<128x1xf32>
+ -> !xegpu.tensor_desc<128x1xf32, #xegpu.layout<sg_layout = [4, 1], sg_data = [16, 1], lane_layout = [8, 1], lane_data = [1, 1]>>
%load = xegpu.load_nd %tdesc
- : !xegpu.tensor_desc<24x1xf32, #xegpu.layout<sg_layout = [4, 1], sg_data = [2, 1], lane_layout = [2, 1], lane_data = [1, 1]>>
- -> vector<24x1xf32>
- // CHECK-COUNT-3: vector.broadcast {{.*}}
- // CHECK-SAME-COUNT-3: {layout_result_0 = #xegpu.layout<lane_layout = [2, 1], lane_data = [1, 1]>}
- // CHECK-SAME-COUNT-3: : vector<2x1xf32> to vector<2x4xf32>
+ : !xegpu.tensor_desc<128x1xf32, #xegpu.layout<sg_layout = [4, 1], sg_data = [16, 1], lane_layout = [8, 1], lane_data = [1, 1]>>
+ -> vector<128x1xf32>
+ // CHECK-COUNT-2: vector.broadcast {{.*}}
+ // CHECK-SAME-COUNT-2: {layout_result_0 = #xegpu.layout<lane_layout = [8, 1], lane_data = [1, 1]>}
+ // CHECK-SAME-COUNT-2: : vector<16x1xf32> to vector<16x32xf32>
// CHECK-NOT: vector.broadcast
%broadcast = vector.broadcast %load
- {layout_result_0 = #xegpu.layout<sg_layout = [4, 1], sg_data = [2, 4], lane_layout = [2, 1], lane_data = [1, 1]>}
- : vector<24x1xf32> to vector<24x8xf32>
+ {layout_result_0 = #xegpu.layout<sg_layout = [4, 1], sg_data = [16, 32], lane_layout = [8, 1], lane_data = [1, 1]>}
+ : vector<128x1xf32> to vector<128x64xf32>
gpu.return
}
generated by cgit v1.1 at 2025-08-06 09:58:20 +0000