@@ -79,25 +79,28 @@ module attributes {"ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32,
79
79
%51 = arith.muli %arg7 , %c32_i32 : i32
80
80
%52 = tt.splat %51 : i32 -> tensor <32 x256 xi32 , #blocked1 >
81
81
// COM: There are 3 stages in loop pipelining, the first 2 prefetching stages are before the loop and the last one is inside the loop.
82
- // CHECK: %[[LOAD_MASK:.*]] = arith.cmpi slt, {{.*}}
83
- // CHECK: %[[LOOP_MASK:.*]] = tt.splat %[[LOAD_MASK]] : i1 -> tensor<64x32xi1, #[[$BLOCK_0]]>
84
- // CHECK: triton_intel_gpu.prefetch {{.*}}, %[[LOOP_MASK]] {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
85
- // CHECK: triton_intel_gpu.prefetch {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
86
- // CHECK: triton_intel_gpu.prefetch {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
87
- // CHECK: triton_intel_gpu.prefetch {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
88
- // CHECK: scf.for %[[VAL_92:.*]] = {{.*}} to {{.*}} step {{.*}} iter_args(%[[VAL_93:.*]] = {{.*}}, %[[VAL_94:.*]] = {{.*}}, %[[VAL_95:.*]] = {{.*}}, %[[VAL_96:.*]] = {{.*}}, %[[VAL_97:.*]] = {{.*}}) -> (tensor<64x256xf32, #[[$DPAS]]>, tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>, tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>) : i32 {
89
- // CHECK: %[[LOAD_MASK:.*]] = arith.cmpi slt, {{.*}}
82
+ // CHECK: %[[LOAD_MASK:.*]] = arith.cmpi slt, {{.*}} : tensor<1x32xi32, #[[$BLOCK_0]]>
83
+ // CHECK: %[[LOAD_MASK_2D:.*]] = tt.broadcast %[[LOAD_MASK]] : tensor<1x32xi1, #[[$BLOCK_0]]> -> tensor<64x32xi1, #[[$BLOCK_0]]>
84
+ // CHECK: %[[LOOP_MASK:.*]] = tt.splat {{.*}} : i1 -> tensor<64x32xi1, #[[$BLOCK_0]]>
85
+ // CHECK: %[[PREFETCH_MASK:.*]] = arith.andi %[[LOOP_MASK]], %[[LOAD_MASK_2D]] : tensor<64x32xi1, #[[$BLOCK_0]]>
86
+ // CHECK: triton_intel_gpu.prefetch {{.*}}, %[[PREFETCH_MASK]] {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
87
+ // CHECK: %[[LOAD_MASK_2:.*]] = arith.cmpi slt, {{.*}} : tensor<32x1xi32, #[[$BLOCK_1]]>
88
+ // CHECK: %[[LOAD_MASK_2D_2:.*]] = tt.broadcast %[[LOAD_MASK_2]] : tensor<32x1xi1, #[[$BLOCK_1]]> -> tensor<32x256xi1, #[[$BLOCK_1]]>
89
+ // CHECK: %[[LOOP_MASK:.*]] = tt.splat {{.*}} : i1 -> tensor<32x256xi1, #[[$BLOCK_1]]>
90
+ // CHECK: %[[PREFETCH_MASK:.*]] = arith.andi %[[LOOP_MASK]], %[[LOAD_MASK_2D_2]] : tensor<32x256xi1, #[[$BLOCK_1]]>
91
+ // CHECK: triton_intel_gpu.prefetch {{.*}}, %[[PREFETCH_MASK]] {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
92
+ // CHECK: triton_intel_gpu.prefetch {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
93
+ // CHECK: triton_intel_gpu.prefetch {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
94
+ // CHECK: scf.for %[[VAL_92:.*]] = {{.*}} to {{.*}} step {{.*}} iter_args(%[[VAL_93:.*]] = {{.*}}, %[[VAL_94:.*]] = {{.*}}, %[[VAL_95:.*]] = {{.*}}, %[[VAL_96:.*]] = {{.*}}, %[[VAL_97:.*]] = {{.*}}, %[[VAL_98:.*]] = {{.*}}, %[[VAL_99:.*]] = {{.*}}, %[[VAL_100:.*]] = {{.*}}, %[[VAL_101:.*]] = {{.*}}) -> (tensor<64x256xf32, #[[$DPAS]]>, tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>, tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<64x32xi1, #[[$BLOCK_0]]>, tensor<64x32xi1, #[[$BLOCK_0]]>, tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>, tensor<32x256xi1, #[[$BLOCK_1]]>, tensor<32x256xi1, #[[$BLOCK_1]]>) : i32 {
90
95
// CHECK: %[[VAL_106:.*]] = tt.addptr %[[VAL_94]], {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<64x32xi32, #[[$BLOCK_0]]>
91
96
// CHECK: %[[VAL_107:.*]] = tt.addptr %[[VAL_95]], {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>, tensor<32x256xi32, #[[$BLOCK_1]]>
92
- // CHECK: %[[LOOP_MASK:.*]] = tt.splat %[[LOAD_MASK]] : i1 -> tensor<64x32xi1, #[[$BLOCK_0]]>
93
- // CHECK: triton_intel_gpu.prefetch %[[VAL_106]], %[[LOOP_MASK]] {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
97
+ // CHECK: triton_intel_gpu.prefetch %[[VAL_106]], {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
94
98
// CHECK: triton_intel_gpu.prefetch %[[VAL_107]], {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
95
- // CHECK: %[[VAL_116:.*]] = tt.load %[[VAL_96]], {{.*}}, {{.*}} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
96
- // CHECK: %[[VAL_120:.*]] = tt.load %[[VAL_97 ]], {{.*}}, {{.*}} : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
99
+ // CHECK: %[[VAL_116:.*]] = tt.load %[[VAL_96]], {{.*}}, {{.*}} {triton_intel_gpu.block_io = "row_major"} : tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>
100
+ // CHECK: %[[VAL_120:.*]] = tt.load %[[VAL_99 ]], {{.*}}, {{.*}} {triton_intel_gpu.block_io = "row_major" } : tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
97
101
// CHECK: %[[VAL_121:.*]] = ttg.convert_layout %[[VAL_116]] : tensor<64x32xf16, #[[$BLOCK_0]]> -> tensor<64x32xf16, #{{.*}}<{opIdx = 0, parent = #[[$DPAS]], kWidth = 1}>>
98
102
// CHECK: %[[VAL_122:.*]] = ttg.convert_layout %[[VAL_120]] : tensor<32x256xf16, #[[$BLOCK_1]]> -> tensor<32x256xf16, #{{.*}}<{opIdx = 1, parent = #[[$DPAS]], kWidth = 2}>>
99
103
// CHECK: %[[VAL_123:.*]] = tt.dot %[[VAL_121]], %[[VAL_122]], %[[VAL_93]], inputPrecision = tf32 : tensor<64x32xf16, #{{.*}}<{opIdx = 0, parent = #[[$DPAS]], kWidth = 1}>> * tensor<32x256xf16, #{{.*}}<{opIdx = 1, parent = #[[$DPAS]], kWidth = 2}>> -> tensor<64x256xf32, #[[$DPAS]]>
100
- // CHECK: scf.yield %[[VAL_123]], %[[VAL_106]], %[[VAL_107]], %[[VAL_94]], %[[VAL_95]] : tensor<64x256xf32, #[[$DPAS]]>, tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>, tensor<64x32x!tt.ptr<f16>, #[[$BLOCK_0]]>, tensor<32x256x!tt.ptr<f16>, #[[$BLOCK_1]]>
101
104
%53:3 = scf.for %arg9 = %c0_i32 to %50 step %c1_i32 iter_args (%arg10 = %cst_2 , %arg11 = %38 , %arg12 = %48 ) -> (tensor <64 x256 xf32 , #dpas >, tensor <64 x32 x!tt.ptr <f16 >, #blocked >, tensor <32 x256 x!tt.ptr <f16 >, #blocked1 >) : i32 {
102
105
%72 = arith.muli %arg9 , %c32_i32 : i32
103
106
%73 = arith.subi %arg5 , %72 : i32
0 commit comments