Skip to content

[Coalesce]: Enhance the Intel coalescing pass to support while loops. #4290

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
merged 16 commits into from
May 26, 2025
Merged

[Coalesce]: Enhance the Intel coalescing pass to support while loops. #4290

Show file tree
Hide file tree
Changes from 12 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
4384ad1
Ensure block ptr is created with the same layout as the descriptor_lo…
etiotto May 9, 2025
f0ce91c
Remove naked print and unnecessary headers
etiotto May 9, 2025
04f1c1d
Merge remote-tracking branch 'origin/main' into etiotto.tensor_desc_t…
etiotto May 14, 2025
3543e6e
WIP: TensorDescToBlockPtr updates
etiotto May 14, 2025
38ef6c3
WIP: RemoveLAuoyutConversion improvement for tt.advance operation
etiotto May 15, 2025
e4d5d7d
Merge remote-tracking branch 'origin/main' into etiotto.tensor_desc_t…
etiotto May 15, 2025
9eb16ec
WIP: TensorDescToBlockPtr updates
etiotto May 15, 2025
7f9bbc9
WIP: TensorDescToBlockPtr updates
etiotto May 15, 2025
f6ed66a
WIP: TensorDescToBlockPtr updates
etiotto May 16, 2025
cb4bb2e
WIP: TensorDescToBlockPtr updates
etiotto May 20, 2025
f6ce50a
Merge branch 'main' into etiotto.tensor_desc_to_block_ptr.1
etiotto May 20, 2025
d1c3d0b
Enhance coalescing pass to support scf::While loop
etiotto May 23, 2025
1f4ec74
Merge branch 'main' into etiotto.coalesce.1
etiotto May 23, 2025
1cc7fd5
Merge branch 'main' into etiotto.coalesce.1
etiotto May 23, 2025
d4c2655
Remove unnecessary changes
etiotto May 23, 2025
65dc3c3
Address code review comments
etiotto May 26, 2025
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
22 changes: 13 additions & 9 deletions test/Triton/Intel/TensorDescToBlockPointer/basic.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,25 +1,27 @@
// RUN: triton-opt %s -triton-intel-tdesc-to-block-pointer | FileCheck %s

module {
module attributes {"ttg.num-warps" = 4 : i32} {
tt.func public @test_load(%arg0: !tt.ptr<f32>, %arg1: i32, %arg2: i32) {
%c1_i64 = arith.constant 1 : i64
%c64_i32 = arith.constant 64 : i32
%c8_i32 = arith.constant 8 : i32
%0 = arith.extsi %arg2 : i32 to i64
%desc = tt.make_tensor_descriptor %arg0, [%arg1, %arg2], [%0, %c1_i64] : <f32>, <tensor<16x128xf32>>
%load = tt.descriptor_load %desc[%c8_i32, %c64_i32] : !tt.tensordesc<tensor<16x128xf32>> -> tensor<16x128xf32>
%desc1 = tt.make_tensor_descriptor %arg0, [%arg1, %arg2], [%0, %c1_i64] : <f32>, <tensor<16x128xf32>>
%load1 = tt.descriptor_load %desc1[%c8_i32, %c64_i32] : !tt.tensordesc<tensor<16x128xf32>> -> tensor<16x128xf32>
tt.return
}
// CHECK: tt.func public @test_load([[PARAM_0:%.+]]: !tt.ptr<f32>, [[PARAM_1:%.+]]: i32, [[PARAM_2:%.+]]: i32) {
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK-NOT: tt.descriptor_load
// CHECK-DAG: [[CST_0_i32:%.+]] = arith.constant 0 : i32
// CHECK-DAG: [[CST_1_i64:%.+]] = arith.constant 1 : i64
// CHECK-DAG: [[CST_64_i32:%.+]] = arith.constant 64 : i32
// CHECK-DAG: [[CST_8_i32:%.+]] = arith.constant 8 : i32
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_8_i32]], [[CST_64_i32]]] {{.*}} : <tensor<16x128xf32>>
// CHECK: [[LOAD:%.+]] = tt.load [[TENSOR_PTR]] : !tt.ptr<tensor<16x128xf32>>
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_0_i32]], [[CST_0_i32]]] {{.*}} : <tensor<16x128xf32>>
// CHECK: [[TENSOR_PTR1:%.+]] = tt.advance [[TENSOR_PTR]], {{\[}}[[CST_8_i32]], [[CST_64_i32]]] : <tensor<16x128xf32>>
// CHECK: [[LOAD1:%.+]] = tt.load [[TENSOR_PTR1]] : !tt.ptr<tensor<16x128xf32>>
// CHECK: tt.return
// CHECK: }

Expand All @@ -29,21 +31,23 @@ module {
%c8_i32 = arith.constant 8 : i32
%cst = arith.constant dense<1.000000e+00> : tensor<16x128xf32>
%0 = arith.extsi %arg2 : i32 to i64
%desc = tt.make_tensor_descriptor %arg0, [%arg1, %arg2], [%0, %c1_i64] : <f32>, <tensor<16x128xf32>>
tt.descriptor_store %desc[%c8_i32, %c64_i32], %cst : !tt.tensordesc<tensor<16x128xf32>>, tensor<16x128xf32>
%desc1 = tt.make_tensor_descriptor %arg0, [%arg1, %arg2], [%0, %c1_i64] : <f32>, <tensor<16x128xf32>>
tt.descriptor_store %desc1[%c8_i32, %c64_i32], %cst : !tt.tensordesc<tensor<16x128xf32>>, tensor<16x128xf32>
tt.return
}
// CHECK: tt.func public @test_store([[PARAM_0:%.+]]: !tt.ptr<f32>, [[PARAM_1:%.+]]: i32, [[PARAM_2:%.+]]: i32) {
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK-NOT: tt.descriptor_store
// CHECK-DAG: [[CST_0_i32:%.+]] = arith.constant 0 : i32
// CHECK-DAG: [[CST_1_i64:%.+]] = arith.constant 1 : i64
// CHECK-DAG: [[CST_64_i32:%.+]] = arith.constant 64 : i32
// CHECK-DAG: [[CST_8_i32:%.+]] = arith.constant 8 : i32
// CHECK-DAG: [[CST:%.+]] = arith.constant dense<1.000000e+00> : tensor<16x128xf32>
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_8_i32]], [[CST_64_i32]]] {{.*}} : <tensor<16x128xf32>>
// CHECK: tt.store [[TENSOR_PTR]], [[CST]] : !tt.ptr<tensor<16x128xf32>>
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_0_i32]], [[CST_0_i32]]] {{.*}} : <tensor<16x128xf32>>
// CHECK: [[TENSOR_PTR1:%.+]] = tt.advance [[TENSOR_PTR]], {{\[}}[[CST_8_i32]], [[CST_64_i32]]] : <tensor<16x128xf32>>
// CHECK: tt.store [[TENSOR_PTR1]], [[CST]] : !tt.ptr<tensor<16x128xf32>>
// CHECK: tt.return
// CHECK: }
}
90 changes: 55 additions & 35 deletions test/Triton/Intel/TensorDescToBlockPointer/loop.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,15 +25,16 @@ module {
// CHECK-DAG: [[CST_1_i64:%.+]] = arith.constant 1 : i64
// CHECK-DAG: [[CST_8_i32:%.+]] = arith.constant 8 : i32
// CHECK-DAG: [[CST:%.+]] = arith.constant dense<0.000000e+00> : tensor<16x32xf16>
// CHECK-DAG: [[EXTSI_PARAM_2a:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} iter_args([[VAR_arg1:%.+]] = {{.*}}, [[VAR_arg2:%.+]] = [[CST]]) -> (!tt.tensordesc<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK-DAG: [[IDX_CAST_1:%.+]] = arith.index_cast [[IV]] : index to i32
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2b:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2b]]], {{\[}}[[EXTSI_PARAM_2a]], [[CST_1_i64]]], {{\[}}[[CST_8_i32]], [[IDX_CAST_1]]] {{.*}} : <tensor<16x32xf16>>
// CHECK: [[LOAD:%.+]] = tt.load [[TENSOR_PTR]] : !tt.ptr<tensor<16x32xf16>>
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK-DAG: [[CST_0_i32:%.+]] = arith.constant 0 : i32
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_0_i32]], [[CST_0_i32]]] {{.*}} : <tensor<16x32xf16>>
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} iter_args([[VAR_arg1:%.+]] = [[TENSOR_PTR]], [[VAR_arg2:%.+]] = [[CST]]) -> (!tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK: [[IDX_CAST:%.+]] = arith.index_cast [[IV]] : index to i32
// CHECK: [[TENSOR_PTR_1:%.+]] = tt.advance [[VAR_arg1]], {{\[}}[[CST_8_i32]], [[IDX_CAST]]] : <tensor<16x32xf16>>
// CHECK: [[LOAD:%.+]] = tt.load [[TENSOR_PTR_1]] : !tt.ptr<tensor<16x32xf16>>
// CHECK: [[ADD:%.+]] = arith.addf [[VAR_arg2]], [[LOAD]] : tensor<16x32xf16>
// CHECK: scf.yield {{.*}}, [[ADD]] : !tt.tensordesc<tensor<16x32xf16>>, tensor<16x32xf16>
// CHECK: scf.yield [[VAR_arg1]], [[ADD]] : !tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>
// CHECK: }
// CHECK: tt.return
// CHECK: }
Expand All @@ -60,12 +61,27 @@ module {
tt.return
}
// CHECK: tt.func public @load_in_loop2({{.*}}) {
// CHECK-NOT: tt.make_tensor_ptr
// CHECK-NOT: tt.load
// CHECK: tt.make_tensor_descriptor
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} -> (!tt.tensordesc<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK: tt.descriptor_load
// CHECK: tt.make_tensor_descriptor
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK-NOT: tt.descriptor_load
// CHECK-DAG: [[CST_1_i64:%.+]] = arith.constant 1 : i64
// CHECK-DAG: [[CST_8_i32:%.+]] = arith.constant 8 : i32
// CHECK-DAG: [[CST:%.+]] = arith.constant dense<0.000000e+00> : tensor<16x32xf16>
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK-DAG: [[CST_0_i32:%.+]] = arith.constant 0 : i32
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_0_i32]], [[CST_0_i32]]] {{.*}} : <tensor<16x32xf16>>
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} iter_args([[VAR_arg1:%.+]] = [[TENSOR_PTR]], [[VAR_arg2:%.+]] = [[CST]]) -> (!tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK: [[IDX_CAST:%.+]] = arith.index_cast [[IV]] : index to i32
// CHECK: [[TENSOR_PTR_1:%.+]] = tt.advance [[VAR_arg1]], {{\[}}[[CST_8_i32]], [[IDX_CAST]]] : <tensor<16x32xf16>>
// CHECK: [[LOAD:%.+]] = tt.load [[TENSOR_PTR_1]] : !tt.ptr<tensor<16x32xf16>>
// CHECK: [[ADD:%.+]] = arith.addf [[VAR_arg2]], [[LOAD]] : tensor<16x32xf16>
// CHECK-DAG: [[EXTSI_PARAM_1a:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2a:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK-DAG: [[CST_0_i32_1:%.+]] = arith.constant 0 : i32
// CHECK: [[TENSOR_PTR2:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_2a]], [[EXTSI_PARAM_1a]]], {{\[}}[[CST_1_i64]], [[EXTSI_PARAM_2]]], {{\[}}[[CST_0_i32_1]], [[CST_0_i32_1]]] {{.*}} : <tensor<16x32xf16>>
// CHECK: [[CMP:%.+]] = arith.cmpi eq, [[IDX_CAST]], [[CST_8_i32]] : i32
// CHECK: [[TENSOR_PTR3:%.+]] = arith.select [[CMP]], [[VAR_arg1]], [[TENSOR_PTR:%.+]] : !tt.ptr<tensor<16x32xf16>>
// CHECK: scf.yield [[TENSOR_PTR3]], [[ADD]] : !tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>
// CHECK: }
// CHECK: tt.return
// CHECK: }
Expand All @@ -87,10 +103,12 @@ module {
tt.return
}
// CHECK: tt.func public @load_uses_loop_result({{.*}}) {
// CHECK-NOT: tt.make_tensor_ptr
// CHECK-NOT: tt.load
// CHECK: tt.make_tensor_descriptor
// CHECK: tt.descriptor_load
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr {{.*}} : <tensor<16x32xf16>>
// CHECK: [[FOR_RES:%.+]] = scf.for [[IV:%.+]] = {{.*}} iter_args([[VAR_arg1:%.+]] = [[TENSOR_PTR]]) -> (!tt.ptr<tensor<16x32xf16>>)
// CHECK: [[TENSOR_PTR1:%.+]] = tt.advance [[FOR_RES]], {{.*}} : <tensor<16x32xf16>>
// CHECK: tt.load [[TENSOR_PTR1]] : !tt.ptr<tensor<16x32xf16>>
// CHECK: tt.return
// CHECK: }

Expand All @@ -115,18 +133,19 @@ module {
// CHECK: tt.func public @store_in_loop1([[PARAM_0:%.+]]: !tt.ptr<f16>, [[PARAM_1:%.+]]: i32, [[PARAM_2:%.+]]: i32) {
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK-NOT: tt.descriptor_store
// CHECK-DAG: [[CST_0_i32:%.+]] = arith.constant 0 : i32
// CHECK-DAG: [[CST_1_i64:%.+]] = arith.constant 1 : i64
// CHECK-DAG: [[CST_8_i32:%.+]] = arith.constant 8 : i32
// CHECK-DAG: [[CST:%.+]] = arith.constant dense<0.000000e+00> : tensor<16x32xf16>
// CHECK-DAG: [[EXTSI_PARAM_2a:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} iter_args([[VAR_arg1:%.+]] = {{.*}}, [[VAR_arg2:%.+]] = [[CST]]) -> (!tt.tensordesc<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK-DAG: [[IDX_CAST_1:%.+]] = arith.index_cast [[IV]] : index to i32
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2b:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2b]]], {{\[}}[[EXTSI_PARAM_2a]], [[CST_1_i64]]], {{\[}}[[CST_8_i32]], [[IDX_CAST_1]]] {{.*}} : <tensor<16x32xf16>>
// CHECK: tt.store [[TENSOR_PTR]], [[VAR_arg2]] : !tt.ptr<tensor<16x32xf16>>
// CHECK-DAG: [[EXTSI_PARAM_1:%.+]] = arith.extsi [[PARAM_1]] : i32 to i64
// CHECK-DAG: [[EXTSI_PARAM_2:%.+]] = arith.extsi [[PARAM_2]] : i32 to i64
// CHECK: [[TENSOR_PTR:%.+]] = tt.make_tensor_ptr [[PARAM_0]], {{\[}}[[EXTSI_PARAM_1]], [[EXTSI_PARAM_2]]], {{\[}}[[EXTSI_PARAM_2]], [[CST_1_i64]]], {{\[}}[[CST_0_i32]], [[CST_0_i32]]] {{.*}} : <tensor<16x32xf16>>
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} iter_args([[VAR_arg1:%.+]] = [[TENSOR_PTR]], [[VAR_arg2:%.+]] = [[CST]]) -> (!tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK: [[IDX_CAST_1:%.+]] = arith.index_cast [[IV]] : index to i32
// CHECK: [[TENSOR_PTR_1:%.+]] = tt.advance [[VAR_arg1]], {{\[}}[[CST_8_i32]], [[IDX_CAST]]] : <tensor<16x32xf16>>
// CHECK: tt.store [[TENSOR_PTR_1]], [[VAR_arg2]] : !tt.ptr<tensor<16x32xf16>>
// CHECK: [[ADD:%.+]] = arith.addf [[VAR_arg2]], [[CST]] : tensor<16x32xf16>
// CHECK: scf.yield {{.*}}, [[ADD]] : !tt.tensordesc<tensor<16x32xf16>>, tensor<16x32xf16>
// CHECK: scf.yield [[VAR_arg1]], [[ADD]] : !tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>
// CHECK: }
// CHECK: tt.return
// CHECK: }
Expand All @@ -153,12 +172,12 @@ module {
tt.return
}
// CHECK: tt.func public @store_in_loop2({{.*}}) {
// CHECK-NOT: tt.make_tensor_ptr
// CHECK-NOT: tt.store
// CHECK: tt.make_tensor_descriptor
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} -> (!tt.tensordesc<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK: tt.descriptor_store
// CHECK: tt.make_tensor_descriptor
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK-NOT: tt.descriptor_store
// CHECK: tt.make_tensor_ptr
// CHECK: [[FOR_RES:%.+]]:2 = scf.for [[IV:%.+]] = {{.*}} -> (!tt.ptr<tensor<16x32xf16>>, tensor<16x32xf16>) {
// CHECK: tt.advance
// CHECK: tt.store
// CHECK: }
// CHECK: tt.return
// CHECK: }
Expand All @@ -181,10 +200,11 @@ module {
tt.return
}
// CHECK: tt.func public @store_uses_loop_result({{.*}}) {
// CHECK-NOT: tt.make_tensor_ptr
// CHECK-NOT: tt.store
// CHECK: tt.make_tensor_descriptor
// CHECK: tt.descriptor_store
// CHECK-NOT: tt.make_tensor_descriptor
// CHECK-NOT: tt.descriptor_store
// CHECK: tt.make_tensor_ptr
// CHECK: tt.advance
// CHECK: tt.store
// CHECK: tt.return
// CHECK: }

Expand Down
5 changes: 4 additions & 1 deletion test/TritonIntelGPU/backward_combine_dpas_dot_layout.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@
#dot0 = #ttg.dot_op<{opIdx = 0, parent = #dpas, kWidth=1}>
#dot1 = #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth=2}>
module attributes {"ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32, "triton_intel_gpu.support_sg_2d_block"} {
// CHECK: matmul_kernel_with_block_pointers
tt.func public @matmul_kernel_with_block_pointers(%arg0: !tt.ptr<f16>, %arg1: !tt.ptr<f16>, %arg2: !tt.ptr<f16>, %arg3: i32, %arg4: i32, %arg5: i32, %arg6: i64, %arg7: i32, %arg8: i64) {
%c8_i32 = arith.constant 8 : i32
%c64_i32 = arith.constant 64 : i32
Expand Down Expand Up @@ -66,7 +67,6 @@ module attributes {"ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32,
}
%24 = arith.truncf %23#0 : tensor<64x256xf32, #dpas> to tensor<64x256xf16, #dpas>
%25 = ttg.convert_layout %24 : tensor<64x256xf16, #dpas> -> tensor<64x256xf16, #blocked1>

%27 = tt.make_tensor_ptr %arg2, [%15, %20], [%arg8, %c1_i64], [%14, %19] {order = array<i32: 1, 0>} : <tensor<64x256xf16, #blocked1>>
tt.store %27, %25 {boundaryCheck = array<i32: 0, 1>} : !tt.ptr<tensor<64x256xf16, #blocked1>>
tt.return
Expand All @@ -86,6 +86,7 @@ module attributes {"ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32,
#dot0 = #ttg.dot_op<{opIdx = 0, parent = #dpas, kWidth=1}>
#dot1 = #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth=2}>
module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32, "triton_intel_gpu.support_sg_2d_block"} {
// CHECK: matmul_kernel_with_block_pointers
tt.func public @matmul_kernel_with_block_pointers(%arg0: !tt.ptr<f16>, %arg1: !tt.ptr<f16>, %arg2: !tt.ptr<f16>, %arg3: i32, %arg4: i32, %arg5: i32, %arg6: i32, %arg7: i32, %arg8: i32) {
%c8_i32 = arith.constant 8 : i32
%c64_i32 = arith.constant 64 : i32
Expand Down Expand Up @@ -154,6 +155,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, "ttg.thr
#dot0 = #ttg.dot_op<{opIdx = 0, parent = #dpas, kWidth=1}>
#dot1 = #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth=2}>
module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32, "triton_intel_gpu.support_sg_2d_block"} {
// CHECK: matmul_kernel_with_block_pointers
tt.func public @matmul_kernel_with_block_pointers(%arg0: !tt.ptr<f16>, %arg1: !tt.ptr<f16>, %arg2: !tt.ptr<f16>, %arg3: i32, %arg4: i32, %arg5: i32, %arg6: i32, %arg7: i32, %arg8: i32, %arg13: !tt.ptr<f16>, %arg14: !tt.ptr<f32>) {
%c8_i32 = arith.constant 8 : i32
%c64_i32 = arith.constant 64 : i32
Expand Down Expand Up @@ -233,6 +235,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, "ttg.thr
#dot0 = #ttg.dot_op<{opIdx = 0, parent = #dpas, kWidth=1}>
#dot1 = #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth=2}>
module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, "ttg.threads-per-warp" = 16 : i32, "triton_intel_gpu.support_sg_2d_block"} {
// CHECK: matmul_kernel_with_block_pointers
tt.func public @matmul_kernel_with_block_pointers(%arg0: !tt.ptr<f16>, %arg1: !tt.ptr<f16>, %arg2: !tt.ptr<f16>, %arg3: i32, %arg4: i32, %arg5: i32, %arg6: i32, %arg7: i32, %arg8: i32) {
%c1_i64 = arith.constant 1 : i64
%c0_i32 = arith.constant 0 : i32
Expand Down
Loading