Skip to content

[LoadStoreOpToLLVM] Unify the 2D block IO lowering code for both regular pointer and block pointer #5500

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

Open
chengjunlu wants to merge 1 commit into
base: main
Choose a base branch
from
+287 −974
Open

[LoadStoreOpToLLVM] Unify the 2D block IO lowering code for both regular pointer and block pointer #5500

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
107 changes: 36 additions & 71 deletions test/TritonIntelGPU/blockptr_load.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,3 @@
// RUN: env TRITON_INTEL_PREDICATED=1 TRITON_INTEL_ONE_MATRIX_PER_LOAD_BT=0 triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s --implicit-check-not=llvm.inline_asm --check-prefixes=CHECK,LARGE-BLOCK-SIZE-TRANS-B
// RUN: env TRITON_INTEL_PREDICATED=1 TRITON_INTEL_ONE_MATRIX_PER_LOAD_BT=1 triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s --implicit-check-not=llvm.inline_asm --check-prefixes=CHECK,SMALL-BLOCK-SIZE-TRANS-B

#blocked = #ttg.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 16], warpsPerCTA = [2, 4], order = [1, 0]}>
Expand Down Expand Up @@ -71,7 +70,7 @@ module attributes {"ttg.num-warps" = 8 : i32, "ttg.threads-per-warp" = 16 : i32,
%B = tt.load %ptrB {boundaryCheck = array<i32: 0>, padding = 1 : i32, ttig.block_io = "row_major"} : !tt.ptr<tensor<32x64xf16, #dot1>>
%D = tt.dot %A, %B, %C, inputPrecision = tf32 : tensor<64x32xf16, #dot0> * tensor<32x64xf16, #dot1> -> tensor<64x64xf32, #dpas>
%ptrX = tt.make_tensor_ptr %arg2, [%arg3, %arg4], [%arg8, %c1_i64], [%c0_i32, %c0_i32] {order = array<i32: 1, 0>} : <tensor<64x64xf32, #dpas>>
// CHECK-NOT: triton_gen.2Dblockload {{.*}}
// CHECK-COUNT-4: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
%X = tt.load %ptrX {boundaryCheck = array<i32: 0, 1>, ttig.block_io = "column_major"} : !tt.ptr<tensor<64x64xf32, #dpas>>
%0 = arith.addf %D, %X : tensor<64x64xf32, #dpas>
tt.return
Expand Down Expand Up @@ -122,38 +121,30 @@ module attributes {"ttg.num-warps" = 1 : i32, "ttg.threads-per-warp" = 16 : i32,
// CHECK: %[[VAL_12:.*]] = llvm.insertvalue %[[VAL_3]], %[[VAL_11]][4] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[VAL_13:.*]] = llvm.insertvalue %[[VAL_6]], %[[VAL_12]][5] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[BLOCK_POINTER:.*]] = llvm.insertvalue %[[VAL_0]], %[[VAL_13]][6] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[SUB_GROUP_ID_RAW:.*]] = llvm.call spir_funccc @_Z16get_sub_group_id() {no_unwind, will_return} : () -> i32
// CHECK: %[[SUB_GROUP_ID_EXT:.*]] = llvm.zext %[[SUB_GROUP_ID_RAW]] : i32 to i64
// CHECK: %[[SUB_GROUP_ID:.*]] = llvm.trunc %[[SUB_GROUP_ID_EXT]] : i64 to i32
// CHECK: %[[VAL_18:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[VAL_19:.*]] = llvm.urem %[[SUB_GROUP_ID]], %[[VAL_18]] : i32
// CHECK: %[[VAL_20:.*]] = llvm.udiv %[[SUB_GROUP_ID]], %[[VAL_18]] : i32
// CHECK: %[[VAL_21:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[VAL_22:.*]] = llvm.urem %[[VAL_20]], %[[VAL_21]] : i32
// CHECK: %[[VAL_23:.*]] = llvm.udiv %[[VAL_20]], %[[VAL_21]] : i32
// CHECK: %[[VAL_24:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[outerDimWarpId:.*]] = llvm.urem %[[VAL_22]], %[[VAL_24]] : i32
// CHECK: %[[OFFSET_0:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][0] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[OFFSET_1:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][1] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[WIDTH_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][2] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[HEIGHT_i64:.*]] = llvm.extractvalue %[[VAL_11]][3] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[ROW_STRIDE_i64:.*]] = llvm.extractvalue %[[VAL_12]][4] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[HEIGHT_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][2] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[WIDTH_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][3] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[ROW_STRIDE_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][4] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[COL_STRIDE_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][5] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[BASE:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][6] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[ELEM_SIZE_IN_BYTES:.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK: %[[ROW_STRIDE_i32:.*]] = llvm.trunc %[[ROW_STRIDE_i64]] : i64 to i32
// CHECK: %[[HEIGHT_i32:.*]] = llvm.trunc %[[HEIGHT_i64]] : i64 to i32
// CHECK: %[[ROW_STRIDE_IN_BYTES:.*]] = llvm.mul %[[ROW_STRIDE_i32]], %[[ELEM_SIZE_IN_BYTES]] : i32
// CHECK: %[[WIDTH_i32:.*]] = llvm.trunc %[[WIDTH_i64]] : i64 to i32
// CHECK: %[[HEIGHT_i32:.*]] = llvm.trunc %[[HEIGHT_i64]] : i64 to i32
// CHECK: %[[ELEM_SIZE_IN_BYTES:.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK: %[[VAL_37:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[VAL_38:.*]] = llvm.mlir.constant(32 : i32) : i32
// CHECK: %[[VAL_39:.*]] = llvm.mul %[[outerDimWarpId]], %[[VAL_38]] : i32
// CHECK: %[[VAL_40:.*]] = llvm.add %[[VAL_39]], %[[VAL_37]] : i32
// CHECK: %[[VAL_41:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[offsetX_:.*]] = llvm.add %[[VAL_41]], %[[OFFSET_1]] : i32
// CHECK: %[[offsetY_:.*]] = llvm.add %[[VAL_40]], %[[OFFSET_0]] : i32
// CHECK: %[[ROW_STRIDE_IN_BYTES:.*]] = llvm.mul %[[ROW_STRIDE_i32]], %[[ELEM_SIZE_IN_BYTES]] : i32
// CHECK: %[[HEIGHT:.*]] = llvm.mul %[[HEIGHT_i32]], %[[ELEM_SIZE_IN_BYTES]] : i32
// CHECK: triton_gen.2Dblockload %[[BASE]], %[[HEIGHT]], %[[WIDTH_i32]], %[[ROW_STRIDE_IN_BYTES]], %[[offsetX_]], %[[offsetY_]] {elem_size_in_bits = 16, tile_width = 16, tile_height = 32, v_blocks = 2, transpose = false, vnni_transform = false, cache_control = Default}
// CHECK: %[[WIDTH:.*]] = llvm.mul %[[WIDTH_i32]], %[[ELEM_SIZE_IN_BYTES]] : i32
// CHECK: %[[SUB_GROUP_ID_RAW:.*]] = llvm.call spir_funccc @_Z16get_sub_group_id() {no_unwind, will_return} : () -> i32
// CHECK: %[[SUB_GROUP_ID_EXT:.*]] = llvm.zext %[[SUB_GROUP_ID_RAW]] : i32 to i64
// CHECK: %[[SUB_GROUP_ID:.*]] = llvm.trunc %[[SUB_GROUP_ID_EXT]] : i64 to i32
// CHECK: %[[ADJUSTED_HEIGHT:.*]] = llvm.mlir.constant(32 : i32) : i32
// CHECK: %[[OFFSET_Y:.*]] = llvm.mul {{.*}}, %[[ROW_STRIDE_IN_BYTES]] : i32
// CHECK: %[[ADJUSTED_BASE:.*]] = llvm.getelementptr %[[BASE]]{{\[}}%[[OFFSET_Y]]] : (!llvm.ptr<1>, i32) -> !llvm.ptr<1>, i8
// CHECK: %[[ADJUSTED_OFFSET_Y:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[NUM_VALUES_PER_PACK:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[OFFSET_X:.*]] = llvm.udiv {{.*}}, %[[NUM_VALUES_PER_PACK]] : i32
// CHECK: triton_gen.2Dblockload %[[ADJUSTED_BASE]], %[[WIDTH]], %[[ADJUSTED_HEIGHT]], %[[ROW_STRIDE_IN_BYTES]], %[[OFFSET_X]], %[[ADJUSTED_OFFSET_Y]] {elem_size_in_bits = 16, tile_width = 16, tile_height = 32, v_blocks = 2, transpose = false, vnni_transform = false, cache_control = Default}
%ptrA = tt.make_tensor_ptr %arg0, [%arg2, %arg4], [%arg5, %c1_i64], [%c0_i32, %c0_i32] {order = array<i32: 1, 0>} : <tensor<32x32xf16, #dot0>>
%A = tt.load %ptrA {boundaryCheck = array<i32: 1>, padding = 1 : i32, ttig.block_io = "row_major"} : !tt.ptr<tensor<32x32xf16, #dot0>>
%B = arith.constant dense<0.000000e+00> : tensor<32x32xf16, #dot1>
Expand Down Expand Up @@ -185,38 +176,29 @@ module attributes {"ttg.num-warps" = 1 : i32, "ttg.threads-per-warp" = 16 : i32,
// CHECK: %[[VAL_11:.*]] = llvm.insertvalue %[[VAL_3]], %[[VAL_10]][4] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[VAL_12:.*]] = llvm.insertvalue %[[VAL_5]], %[[VAL_11]][5] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[BLOCK_POINTER:.*]] = llvm.insertvalue %[[VAL_0]], %[[VAL_12]][6] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[SUB_GROUP_ID_RAW:.*]] = llvm.call spir_funccc @_Z16get_sub_group_id() {no_unwind, will_return} : () -> i32
// CHECK: %[[SUB_GROUP_ID_EXT:.*]] = llvm.zext %[[SUB_GROUP_ID_RAW]] : i32 to i64
// CHECK: %[[SUB_GROUP_ID:.*]] = llvm.trunc %[[SUB_GROUP_ID_EXT]] : i64 to i32
// CHECK: %[[VAL_17:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[VAL_18:.*]] = llvm.urem %[[SUB_GROUP_ID]], %[[VAL_17]] : i32
// CHECK: %[[VAL_19:.*]] = llvm.udiv %[[SUB_GROUP_ID]], %[[VAL_17]] : i32
// CHECK: %[[VAL_20:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[VAL_21:.*]] = llvm.urem %[[VAL_19]], %[[VAL_20]] : i32
// CHECK: %[[VAL_22:.*]] = llvm.udiv %[[VAL_19]], %[[VAL_20]] : i32
// CHECK: %[[VAL_23:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[outerDimWarpId:.*]] = llvm.urem %[[VAL_18]], %[[VAL_23]] : i32
// CHECK: %[[OFFSET_0:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][0] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[OFFSET_1:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][1] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[WIDTH_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][2] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[HEIGHT_i64:.*]] = llvm.extractvalue %[[VAL_10]][3] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[ROW_STRIDE_i64:.*]] = llvm.extractvalue %[[VAL_11]][4] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[HEIGHT_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][2] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[WIDTH_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][3] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[ROW_STRIDE_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][4] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[COL_STRIDE_i64:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][5] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[BASE:.*]] = llvm.extractvalue %[[BLOCK_POINTER]][6] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)>
// CHECK: %[[ROW_STRIDE_i32:.*]] = llvm.trunc %[[ROW_STRIDE_i64]] : i64 to i32
// CHECK: %[[HEIGHT_i32:.*]] = llvm.trunc %[[HEIGHT_i64]] : i64 to i32
// CHECK: %[[WIDTH_i32:.*]] = llvm.trunc %[[WIDTH_i64]] : i64 to i32
// CHECK: %[[ELEM_SIZE_IN_BYTES:.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK: %[[VAL_36:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[VAL_37:.*]] = llvm.mlir.constant(32 : i32) : i32
// CHECK: %[[VAL_38:.*]] = llvm.mul %[[outerDimWarpId]], %[[VAL_37]] : i32
// CHECK: %[[VAL_39:.*]] = llvm.add %[[VAL_38]], %[[VAL_36]] : i32
// CHECK: %[[VAL_40:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[offsetX_:.*]] = llvm.add %[[VAL_39]], %[[OFFSET_1]] : i32
// CHECK: %[[offsetY_:.*]] = llvm.add %[[VAL_40]], %[[OFFSET_0]] : i32
// CHECK: %[[ROW_STRIDE_i32:.*]] = llvm.trunc %[[ROW_STRIDE_i64]] : i64 to i32
// CHECK: %[[ROW_STRIDE_IN_BYTES:.*]] = llvm.mul %[[ROW_STRIDE_i32]], %[[ELEM_SIZE_IN_BYTES]] : i32
// CHECK: %[[HEIGHT:.*]] = llvm.mul %[[HEIGHT_i32]], %[[ELEM_SIZE_IN_BYTES]] : i32
// CHECK: triton_gen.2Dblockload %[[BASE]], %[[HEIGHT]], %[[WIDTH_i32]], %[[ROW_STRIDE_IN_BYTES]], %[[offsetX_]], %[[offsetY_]] {elem_size_in_bits = 16, tile_width = 16, tile_height = 32, v_blocks = 2, transpose = false, vnni_transform = true, cache_control = Default}
// CHECK: %[[WIDTH_i32:.*]] = llvm.trunc %[[WIDTH_i64]] : i64 to i32
// CHECK: %[[HEIGHT_i32:.*]] = llvm.trunc %[[HEIGHT_i64]] : i64 to i32
// CHECK: %[[SUB_GROUP_ID_RAW:.*]] = llvm.call spir_funccc @_Z16get_sub_group_id() {no_unwind, will_return} : () -> i32
// CHECK: %[[SUB_GROUP_ID_EXT:.*]] = llvm.zext %[[SUB_GROUP_ID_RAW]] : i32 to i64
// CHECK: %[[SUB_GROUP_ID:.*]] = llvm.trunc %[[SUB_GROUP_ID_EXT]] : i64 to i32
// CHECK: %[[OFFSET_X:.*]] = llvm.add %[[OFFSET_1]], {{.*}} : i32
// CHECK: %[[VAL_47:.*]] = llvm.add %[[OFFSET_0]], {{.*}} : i32
// CHECK: %[[ADJUSTED_WIDTH:.*]] = llvm.mlir.constant(64 : i32) : i32
// CHECK: %[[ADJUSTED_BASE:.*]] = llvm.getelementptr %[[BASE]]{{\[}}%[[OFFSET_X]]] : (!llvm.ptr<1>, i32) -> !llvm.ptr<1>, f16
// CHECK: %[[ADJUSTED_OFFSET_X:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[NUM_VALUES_PACKED:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[OFFSET_X:.*]] = llvm.udiv %[[ADJUSTED_OFFSET_X]], %[[NUM_VALUES_PACKED]] : i32
// CHECK: triton_gen.2Dblockload %[[ADJUSTED_BASE]], %[[ADJUSTED_WIDTH]], %[[HEIGHT_i32]], %[[ROW_STRIDE_IN_BYTES]], %[[OFFSET_X]], {{.*}} {elem_size_in_bits = 16, tile_width = 16, tile_height = 32, v_blocks = 2, transpose = false, vnni_transform = true, cache_control = Default}
%ptrB = tt.make_tensor_ptr %arg1, [%arg4, %arg3], [%arg7, %c1_i64], [%c0_i32, %c0_i32] {order = array<i32: 1, 0>} : <tensor<32x32xf16, #dot1>>
%B = tt.load %ptrB {boundaryCheck = array<i32: 0>, padding = 1 : i32, ttig.block_io = "row_major"} : !tt.ptr<tensor<32x32xf16, #dot1>>
%A = arith.constant dense<0.000000e+00> : tensor<32x32xf16, #dot0>
Expand All @@ -240,20 +222,7 @@ module attributes {"ttg.num-warps" = 1 : i32, "ttg.threads-per-warp" = 16 : i32,
%c32_i64 = arith.constant 32 : i64
%21 = tt.make_tensor_ptr %arg0, [%c64_i64, %c64_i64], [%c1_i64, %col_stride], [%c0_i32, %c0_i32] {order = array<i32: 0, 1>} : <tensor<64x32xf16, #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth = 2}>>>
// COM: One DPAS operand B per load instruction.
// SMALL-BLOCK-SIZE-TRANS-B-COUNT-8: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// COM: Two interleaved DPAS operand B per load instruction. Need to shuffle the loaded value to decompose the VNNI format DPAS operand B.
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_68:.*]] = triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 32, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_69:.*]] = llvm.shufflevector %[[VAL_68]], %[[VAL_68]] [0, 2, 4, 6, 8, 10, 12, 14] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_71:.*]] = llvm.shufflevector %[[VAL_68]], %[[VAL_68]] [1, 3, 5, 7, 9, 11, 13, 15] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_103:.*]] = triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 32, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_104:.*]] = llvm.shufflevector %[[VAL_103]], %[[VAL_103]] [0, 2, 4, 6, 8, 10, 12, 14] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_106:.*]] = llvm.shufflevector %[[VAL_103]], %[[VAL_103]] [1, 3, 5, 7, 9, 11, 13, 15] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_138:.*]] = triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 32, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_139:.*]] = llvm.shufflevector %[[VAL_138]], %[[VAL_138]] [0, 2, 4, 6, 8, 10, 12, 14] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_141:.*]] = llvm.shufflevector %[[VAL_138]], %[[VAL_138]] [1, 3, 5, 7, 9, 11, 13, 15] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_173:.*]] = triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 32, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_174:.*]] = llvm.shufflevector %[[VAL_173]], %[[VAL_173]] [0, 2, 4, 6, 8, 10, 12, 14] : vector<16xi32>
// LARGE-BLOCK-SIZE-TRANS-B: %[[VAL_176:.*]] = llvm.shufflevector %[[VAL_173]], %[[VAL_173]] [1, 3, 5, 7, 9, 11, 13, 15] : vector<16xi32>
// SMALL-BLOCK-SIZE-TRANS-B-COUNT-8: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
%45 = tt.load %21 {ttig.block_io = "column_major"} : !tt.ptr<tensor<64x32xf16, #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth = 2}>>>
tt.return
}
Expand All @@ -271,13 +240,9 @@ module attributes {"ttg.num-warps" = 1 : i32, "ttg.threads-per-warp" = 16 : i32,
%c0_i32 = arith.constant 0 : i32
%21 = tt.make_tensor_ptr %arg0, [%c64_i64, %c64_i64], [%c1_i64, %col_stride], [%c0_i32, %c0_i32] {order = array<i32: 0, 1>} : <tensor<64x16xf16, #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth = 2}>>>
// CHECK: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// CHECK: llvm.shufflevector {{.*}}, {{.*}} [0, 1, 2, 3, 4, 5, 6, 7] : vector<8xi32>
// CHECK: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// CHECK: llvm.shufflevector {{.*}}, {{.*}} [0, 1, 2, 3, 4, 5, 6, 7] : vector<8xi32>
// CHECK: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// CHECK: llvm.shufflevector {{.*}}, {{.*}} [0, 1, 2, 3, 4, 5, 6, 7] : vector<8xi32>
// CHECK: triton_gen.2Dblockload {{.*}} {elem_size_in_bits = 32, tile_width = 8, tile_height = 16, v_blocks = 1, transpose = true, vnni_transform = false, cache_control = Default}
// CHECK: llvm.shufflevector {{.*}}, {{.*}} [0, 1, 2, 3, 4, 5, 6, 7] : vector<8xi32>
%45 = tt.load %21 {ttig.block_io = "column_major"} : !tt.ptr<tensor<64x16xf16, #ttg.dot_op<{opIdx = 1, parent = #dpas, kWidth = 2}>>>
tt.return
}
Expand Down
Loading
Loading