Skip to content

Commit

Permalink
Support init_tile with nD memrefs in wg to sg pass
Browse files Browse the repository at this point in the history
  • Loading branch information
@nbpatel
nbpatel committed Nov 14, 2024
1 parent fd40cd5 commit f797378
Show file tree
Hide file tree
Showing 12 changed files with 146 additions and 5 deletions.
19 changes: 14 additions & 5 deletions lib/Dialect/XeTile/Transforms/WgToSg.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -178,15 +178,15 @@ class WGToSGInitTileOpPattern : public XeOneToNConversion<xetile::InitTileOp> {
if (it != opSgLayoutMap.end()){
assert((opSgLayoutMap[op->getResult(0)] == std::array<int, 2>{0, 1}));
calculateGlobalOffsets(globalOffsetsY, wgTileShape[0], sgTileShape[0],
sgLayout[0], sgDataDimYConst, sgIdX, offsets[0]);
sgLayout[0], sgDataDimYConst, sgIdX, offsets[offsets.size() - 2]);
calculateGlobalOffsets(globalOffsetsX, wgTileShape[1], sgTileShape[1],
sgLayout[1], sgDataDimXConst, sgIdY, offsets[1]);
sgLayout[1], sgDataDimXConst, sgIdY, offsets[offsets.size() - 1]);
}
else {
calculateGlobalOffsets(globalOffsetsY, wgTileShape[0], sgTileShape[0],
sgLayout[0], sgDataDimYConst, sgIdY, offsets[0]);
sgLayout[0], sgDataDimYConst, sgIdY, offsets[offsets.size() - 2]);
calculateGlobalOffsets(globalOffsetsX, wgTileShape[1], sgTileShape[1],
sgLayout[1], sgDataDimXConst, sgIdX, offsets[1]);
sgLayout[1], sgDataDimXConst, sgIdX, offsets[offsets.size() - 1]);
}
// TODO: check for how to broadcast
for (auto y : globalOffsetsY) {
Expand All @@ -197,9 +197,16 @@ class WGToSGInitTileOpPattern : public XeOneToNConversion<xetile::InitTileOp> {

mlir::SmallVector<mlir::Value> newInitTileOps;
llvm::SmallVector<mlir::Type> newResultTypes;
llvm::SmallVector<mlir::OpFoldResult> newOffsets;
for (size_t j = 0; j < offsets.size() - 2; ++j) {
newOffsets.push_back(offsets[j]);
}
for (size_t i = 0; i < offsetPermutations.size(); i++) {
newOffsets.push_back(offsetPermutations[i][0]);
newOffsets.push_back(offsetPermutations[i][1]);
auto newInitTileOp = rewriter.create<xetile::InitTileOp>(
loc, newTileTy, source, offsetPermutations[i]);
loc, newTileTy, source, newOffsets);
newOffsets.clear();
newInitTileOps.push_back(newInitTileOp);
newResultTypes.push_back(newTileTy);
}
Expand Down Expand Up @@ -988,6 +995,8 @@ class XeTileWgToSgPass

target.addDynamicallyLegalOp<mlir::scf::ForOp>(
[&](mlir::scf::ForOp op) -> bool {
if(op.getInitArgs().empty())
return true;
for (auto arg : op.getInitArgs()) {
auto tileTy = mlir::dyn_cast<xetile::TileType>(arg.getType());
if (!tileTy)
Expand Down
0 ...XeTile/Transforms/wg_to_sg_broadcast.mlir → ...t/XeTile/Transforms/WgToSg/broadcast.mlir
View file
File renamed without changes.
0 ...eTile/Transforms/wg_to_sg_btranspose.mlir → .../XeTile/Transforms/WgToSg/btranspose.mlir
View file
File renamed without changes.
0 ...e/Transforms/wg_to_sg_convert_layout.mlir → ...ile/Transforms/WgToSg/convert_layout.mlir
View file
File renamed without changes.
0 ...t/XeTile/Transforms/wg_to_sg_1k_gemm.mlir → ...ect/XeTile/Transforms/WgToSg/gemm_1k.mlir
View file
File renamed without changes.
0 ...t/XeTile/Transforms/wg_to_sg_4k_gemm.mlir → ...ect/XeTile/Transforms/WgToSg/gemm_4k.mlir
View file
File renamed without changes.
64 changes: 64 additions & 0 deletions test/Dialect/XeTile/Transforms/WgToSg/gemm_batch.mlir
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,64 @@
// RUN: imex-opt --split-input-file --xetile-wg-to-sg --cse %s -verify-diagnostics | FileCheck %s

module attributes {gpu.container_module} {
func.func @tiles_b_2_entry(%arg0: memref<4x3x2x128x96xf16>, %arg1: memref<4x3x2x64x96xf16>, %arg2: memref<4x3x2x128x64xf32>) attributes {gemm_tiles_b = 2 : i64, gemm_tiles_x = dense<[2, 1, 1, 2]> : vector<4xi64>, gemm_tiles_y = dense<[2, 1, 1, 1]> : vector<4xi64>, habana_runner.num_inputs = 2 : i64, habana_runner.tests = [{inputs = [dense<1.000000e+00> : tensor<4x3x2x128x96xf16>, dense<1.000000e+00> : tensor<4x3x2x64x96xf16>], outputs = [dense<9.600000e+01> : tensor<4x3x2x128x64xf32>]}], physical_nd_range = dense<[8, 1]> : vector<2xi64>, region_partition = 0 : i64, region_size = 1 : i64, syn.fusion_successful, syn.tensor_signature = (tensor<4x3x2x128x96xf16>, tensor<4x3x2x64x96xf16>) -> tensor<4x3x2x128x64xf32>, synFusionGenOps = 9 : i64, synFusionRequiredBeamSize = 1 : i64, synFusionTotalCost = 1000021695.96 : f64} {
%c8 = arith.constant 8 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
gpu.launch_func @tiles_b_2::@tiles_b_2 blocks in (%c8, %c1, %c1) threads in (%c2, %c1, %c1) args(%arg0 : memref<4x3x2x128x96xf16>, %arg1 : memref<4x3x2x64x96xf16>, %arg2 : memref<4x3x2x128x64xf32>)
return
}
gpu.module @tiles_b_2 attributes {spirv.target_env = #spirv.target_env<#spirv.vce<v1.0, [Addresses, Bfloat16ConversionINTEL, BFloat16TypeKHR, Float16Buffer, Int64, Int16, Int8, Kernel, Linkage, Vector16, GenericPointer, Groups, Float16, Float64, AtomicFloat32AddEXT, ExpectAssumeKHR, VectorAnyINTEL, VectorComputeINTEL], [SPV_EXT_shader_atomic_float_add, SPV_KHR_bfloat16, SPV_KHR_expect_assume, SPV_INTEL_bfloat16_conversion, SPV_INTEL_vector_compute]>, api=OpenCL, #spirv.resource_limits<>>} {
gpu.func @tiles_b_2(%arg0: memref<4x3x2x128x96xf16>, %arg1: memref<4x3x2x64x96xf16>, %arg2: memref<4x3x2x128x64xf32>) kernel attributes {VectorComputeFunctionINTEL, known_block_size = array<i32: 2, 1, 1>, known_grid_size = array<i32: 8, 1, 1>, spirv.entry_point_abi = #spirv.entry_point_abi<>} {
%c96 = arith.constant 96 : index
%c1 = arith.constant 1 : index
%c64 = arith.constant 64 : index
%c32 = arith.constant 32 : index
%c0 = arith.constant 0 : index
%cst = arith.constant {map = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} dense<0.000000e+00> : vector<64x32xf32>
%c3 = arith.constant 3 : index
%c6 = arith.constant 6 : index
%c12 = arith.constant 12 : index
%c8 = arith.constant 8 : index
%c4 = arith.constant 4 : index
%c2 = arith.constant 2 : index
%block_id_x = gpu.block_id x
%0 = arith.remsi %block_id_x, %c4 : index
%1 = arith.divsi %0, %c2 : index
%2 = arith.remsi %0, %c2 : index
%3 = arith.muli %block_id_x, %c2 : index
%4 = arith.divsi %3, %c8 : index
%5 = arith.muli %4, %c12 : index
%6 = arith.muli %1, %c64 : index
%7 = arith.muli %2, %c32 : index
scf.for %arg3 = %c0 to %c12 step %c1 {
%8 = arith.addi %5, %arg3 : index
%9 = arith.divsi %8, %c6 : index
%10 = arith.remsi %9, %c4 : index
%11 = arith.divsi %8, %c2 : index
%12 = arith.remsi %11, %c3 : index
%13 = arith.remsi %8, %c2 : index
//CHECK: %[[INITTILE:.*]] = xetile.init_tile {{%.*}}[{{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}] : memref<4x3x2x128x96xf16> -> !xetile.tile<32x32xf16>
//CHECK: %[[INITTILE:.*]] = xetile.init_tile {{%.*}}[{{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}] : memref<4x3x2x64x96xf16> -> !xetile.tile<32x32xf16>
%14 = xetile.init_tile %arg0[%10, %12, %13, %6, %c0] : memref<4x3x2x128x96xf16> -> !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%15 = xetile.init_tile %arg1[%10, %12, %13, %7, %c0] : memref<4x3x2x64x96xf16> -> !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%16:3 = scf.for %arg4 = %c0 to %c96 step %c32 iter_args(%arg5 = %cst, %arg6 = %14, %arg7 = %15) -> (vector<64x32xf32>, !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>, !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>) {
%18 = xetile.update_tile_offset %arg7, [%c0, %c32] : !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%19 = xetile.update_tile_offset %arg6, [%c0, %c32] : !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%20 = xetile.load_tile %arg6 : !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>> -> vector<64x32xf16>
%21 = xetile.load_tile %arg7 : !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>> -> vector<32x32xf16>
%22 = vector.transpose %21, [1, 0] {map = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} : vector<32x32xf16> to vector<32x32xf16>
xegpu.compile_hint
%23 = xetile.tile_mma %20, %22, %cst {wg_map_a = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>, wg_map_b = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>, wg_map_c = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} : vector<64x32xf16>, vector<32x32xf16>, vector<64x32xf32> -> vector<64x32xf32>
xegpu.compile_hint
%24 = arith.addf %arg5, %23 {map = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} : vector<64x32xf32>
scf.yield %24, %19, %18 : vector<64x32xf32>, !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>, !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
}
%17 = xetile.init_tile %arg2[%10, %12, %13, %6, %7] : memref<4x3x2x128x64xf32> -> !xetile.tile<64x32xf32, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
xetile.store_tile %16#0, %17 : vector<64x32xf32>, !xetile.tile<64x32xf32, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
}
gpu.return
}
}
}

68 changes: 68 additions & 0 deletions test/Dialect/XeTile/Transforms/WgToSg/gemm_batch_oob.mlir
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
// RUN: imex-opt --split-input-file --xetile-wg-to-sg --cse %s -verify-diagnostics | FileCheck %s

module attributes {gpu.container_module} {
func.func @tiles_b_4_oob_entry(%arg0: memref<2x3x3x128x96xf16>, %arg1: memref<2x3x3x64x96xf16>, %arg2: memref<2x3x3x128x64xf32>) attributes {gemm_tiles_b = 4 : i64, gemm_tiles_x = dense<[2, 1, 1, 2]> : vector<4xi64>, gemm_tiles_y = dense<[1, 1, 2, 1]> : vector<4xi64>, habana_runner.num_inputs = 2 : i64, habana_runner.tests = [{inputs = [dense<1.000000e+00> : tensor<2x3x3x128x96xf16>, dense<1.000000e+00> : tensor<2x3x3x64x96xf16>], outputs = [dense<9.600000e+01> : tensor<2x3x3x128x64xf32>]}], physical_nd_range = dense<[8, 2]> : vector<2xi64>, region_partition = 0 : i64, region_size = 2 : i64, syn.fusion_successful, syn.tensor_signature = (tensor<2x3x3x128x96xf16>, tensor<2x3x3x64x96xf16>) -> tensor<2x3x3x128x64xf32>, synFusionGenOps = 9 : i64, synFusionRequiredBeamSize = 1 : i64, synFusionTotalCost = 1000016310.36 : f64} {
%c8 = arith.constant 8 : index
%c2 = arith.constant 2 : index
%c1 = arith.constant 1 : index
gpu.launch_func @tiles_b_4_oob::@tiles_b_4_oob blocks in (%c8, %c2, %c1) threads in (%c2, %c1, %c1) args(%arg0 : memref<2x3x3x128x96xf16>, %arg1 : memref<2x3x3x64x96xf16>, %arg2 : memref<2x3x3x128x64xf32>)
return
}
gpu.module @tiles_b_4_oob attributes {spirv.target_env = #spirv.target_env<#spirv.vce<v1.0, [Addresses, Bfloat16ConversionINTEL, BFloat16TypeKHR, Float16Buffer, Int64, Int16, Int8, Kernel, Linkage, Vector16, GenericPointer, Groups, Float16, Float64, AtomicFloat32AddEXT, ExpectAssumeKHR, VectorAnyINTEL, VectorComputeINTEL], [SPV_EXT_shader_atomic_float_add, SPV_KHR_bfloat16, SPV_KHR_expect_assume, SPV_INTEL_bfloat16_conversion, SPV_INTEL_vector_compute]>, api=OpenCL, #spirv.resource_limits<>>} {
gpu.func @tiles_b_4_oob(%arg0: memref<2x3x3x128x96xf16>, %arg1: memref<2x3x3x64x96xf16>, %arg2: memref<2x3x3x128x64xf32>) kernel attributes {VectorComputeFunctionINTEL, known_block_size = array<i32: 2, 1, 1>, known_grid_size = array<i32: 8, 2, 1>, spirv.entry_point_abi = #spirv.entry_point_abi<>} {
%c96 = arith.constant 96 : index
%c1 = arith.constant 1 : index
%c64 = arith.constant 64 : index
%c32 = arith.constant 32 : index
%c0 = arith.constant 0 : index
%cst = arith.constant {map = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} dense<0.000000e+00> : vector<64x32xf32>
%c3 = arith.constant 3 : index
%c9 = arith.constant 9 : index
%c18 = arith.constant 18 : index
%c5 = arith.constant 5 : index
%c8 = arith.constant 8 : index
%c4 = arith.constant 4 : index
%c2 = arith.constant 2 : index
%block_id_x = gpu.block_id x
%block_id_y = gpu.block_id y
%0 = arith.remsi %block_id_x, %c2 : index
%1 = arith.remsi %block_id_y, %c2 : index
%2 = arith.muli %block_id_x, %c4 : index
%3 = arith.divsi %2, %c8 : index
%4 = arith.muli %3, %c5 : index
%5 = arith.subi %c18, %4 : index
%6 = arith.cmpi sgt, %5, %c5 : index
%7 = arith.select %6, %c5, %5 : index
%8 = arith.muli %0, %c64 : index
%9 = arith.muli %1, %c32 : index
scf.for %arg3 = %c0 to %7 step %c1 {
%10 = arith.addi %4, %arg3 : index
%11 = arith.divsi %10, %c9 : index
%12 = arith.remsi %11, %c2 : index
%13 = arith.divsi %10, %c3 : index
%14 = arith.remsi %13, %c3 : index
%15 = arith.remsi %10, %c3 : index
//CHECK: %[[INITTILE:.*]] = xetile.init_tile {{%.*}}[{{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}] : memref<2x3x3x128x96xf16> -> !xetile.tile<32x32xf16>
//CHECK: %[[INITTILE:.*]] = xetile.init_tile {{%.*}}[{{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}, {{%.*}}] : memref<2x3x3x64x96xf16> -> !xetile.tile<32x32xf16>
%16 = xetile.init_tile %arg0[%12, %14, %15, %8, %c0] : memref<2x3x3x128x96xf16> -> !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%17 = xetile.init_tile %arg1[%12, %14, %15, %9, %c0] : memref<2x3x3x64x96xf16> -> !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%18:3 = scf.for %arg4 = %c0 to %c96 step %c32 iter_args(%arg5 = %cst, %arg6 = %16, %arg7 = %17) -> (vector<64x32xf32>, !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>, !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>) {
%20 = xetile.update_tile_offset %arg7, [%c0, %c32] : !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%21 = xetile.update_tile_offset %arg6, [%c0, %c32] : !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
%22 = xetile.load_tile %arg6 : !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>> -> vector<64x32xf16>
%23 = xetile.load_tile %arg7 : !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>> -> vector<32x32xf16>
%24 = vector.transpose %23, [1, 0] {map = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} : vector<32x32xf16> to vector<32x32xf16>
xegpu.compile_hint
%25 = xetile.tile_mma %22, %24, %cst {wg_map_a = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>, wg_map_b = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>, wg_map_c = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} : vector<64x32xf16>, vector<32x32xf16>, vector<64x32xf32> -> vector<64x32xf32>
xegpu.compile_hint
%26 = arith.addf %arg5, %25 {map = #xetile.wg_map<sg_layout = [2, 1], sg_data = [32, 32]>} : vector<64x32xf32>
scf.yield %26, %21, %20 : vector<64x32xf32>, !xetile.tile<64x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>, !xetile.tile<32x32xf16, #xetile.tile_attr<wg_map = <sg_layout = [1, 2], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
}
%19 = xetile.init_tile %arg2[%12, %14, %15, %8, %9] : memref<2x3x3x128x64xf32> -> !xetile.tile<64x32xf32, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
xetile.store_tile %18#0, %19 : vector<64x32xf32>, !xetile.tile<64x32xf32, #xetile.tile_attr<wg_map = <sg_layout = [2, 1], sg_data = [32, 32]>, inner_blocks = [], memory_space = 0 : i32, scattered = false>>
}
gpu.return
}
}
}

0 ...Tile/Transforms/wg_to_sg_gemm_postop.mlir → ...XeTile/Transforms/WgToSg/gemm_postop.mlir
View file
File renamed without changes.
0 .../XeTile/Transforms/wg_to_sg_prefetch.mlir → ...ct/XeTile/Transforms/WgToSg/prefetch.mlir
View file
File renamed without changes.
0 ...Tile/Transforms/wg_to_sg_round_robin.mlir → ...XeTile/Transforms/WgToSg/round_robin.mlir
View file
File renamed without changes.
0 ...Dialect/XeTile/Transforms/unit_tests.mlir → .../XeTile/Transforms/WgToSg/unit_tests.mlir
View file
File renamed without changes.

0 comments on commit f797378

Please sign in to comment.