| # Copyright 2025 The IREE Authors |
| # |
| # Licensed under the Apache License v2.0 with LLVM Exceptions. |
| # See https://llvm.org/LICENSE.txt for license information. |
| # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| |
| from iree.compiler import ir |
| from iree.compiler.dialects import iree_codegen |
| from iree.compiler.dialects import affine |
| from iree.compiler.ir import AffineMap, AffineDimExpr |
| |
| |
| def run(fn): |
| with ir.Context(), ir.Location.unknown(): |
| module = ir.Module.create() |
| with ir.InsertionPoint(module.body): |
| print("\nTEST:", fn.__name__) |
| fn() |
| return fn |
| |
| |
| @run |
| def root_op(): |
| module_str = """ |
| module { |
| func.func @matmul(%arg0: tensor<4x4xf32>, %arg1: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %cst = arith.constant 0.000000e+00 : f32 |
| %0 = tensor.empty() : tensor<4x4xf32> |
| %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| %2 = linalg.matmul ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x4xf32>) outs(%1 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %2 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| assert input_module is not None, "Failed to parse input MLIR module" |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 0 |
| |
| module_str = """ |
| module { |
| func.func @matmul(%arg0: tensor<4x4xf32>, %arg1: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %cst = arith.constant 0.000000e+00 : f32 |
| %0 = tensor.empty() : tensor<4x4xf32> |
| %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| %2 = linalg.matmul { root_op } ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x4xf32>) outs(%1 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %2 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| assert input_module is not None, "Failed to parse input MLIR module" |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 1 |
| assert root_op_list[0].name == "linalg.matmul" |
| |
| module_str = """ |
| module { |
| func.func @matmul(%arg0: tensor<4x4xf32>, %arg1: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %cst = arith.constant 0.000000e+00 : f32 |
| %0 = tensor.empty() : tensor<4x4xf32> |
| %1 = linalg.fill { root_op } ins(%cst : f32) outs(%0 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| %2 = linalg.matmul { root_op } ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x4xf32>) outs(%1 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %2 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| assert input_module is not None, "Failed to parse input MLIR module" |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 2 |
| assert root_op_list[0].name == "linalg.fill" |
| assert root_op_list[1].name == "linalg.matmul" |
| |
| |
| @run |
| def attention_op_detail(): |
| dim_exprs = [affine.AffineDimExpr.get(i) for i in range(5)] |
| |
| q_map = affine.AffineMap.get(5, 0, [dim_exprs[0], dim_exprs[1], dim_exprs[2]]) |
| k_map = affine.AffineMap.get(5, 0, [dim_exprs[0], dim_exprs[3], dim_exprs[2]]) |
| v_map = affine.AffineMap.get(5, 0, [dim_exprs[0], dim_exprs[3], dim_exprs[4]]) |
| o_map = affine.AffineMap.get(5, 0, [dim_exprs[0], dim_exprs[1], dim_exprs[4]]) |
| |
| result = iree_codegen.get_attention_op_detail(q_map, k_map, v_map, o_map) |
| |
| assert result.domain_rank == 5 |
| assert result.batch_dims == [0] |
| assert result.m_dims == [1] |
| assert result.k1_dims == [2] |
| assert result.k2_dims == [3] |
| assert result.n_dims == [4] |
| |
| dim_exprs = [affine.AffineDimExpr.get(i) for i in range(4)] |
| |
| # Input affine maps that do not follow the expected pattern for an attention operation. |
| q_map = affine.AffineMap.get(4, 0, [dim_exprs[0], dim_exprs[1]]) |
| k_map = affine.AffineMap.get(4, 0, [dim_exprs[0], dim_exprs[2]]) |
| v_map = affine.AffineMap.get(4, 0, [dim_exprs[0], dim_exprs[3]]) |
| o_map = affine.AffineMap.get(4, 0, [dim_exprs[0], dim_exprs[1]]) |
| |
| result = iree_codegen.get_attention_op_detail(q_map, k_map, v_map, o_map) |
| assert result.domain_rank == 4 |
| assert result.batch_dims == [0] |
| assert result.m_dims == [1] |
| assert result.k1_dims == [] |
| assert result.k2_dims == [2] |
| assert result.n_dims == [3] |
| |
| |
| @run |
| def test_isa_attention_op(): |
| module_str = """ |
| module { |
| func.func @attention_20x4096x64x4096x64( |
| %q : tensor<20x4096x64xf16>, |
| %k : tensor<20x4096x64xf16>, |
| %v : tensor<20x4096x64xf16>, |
| %scale : f16, |
| %output : tensor<20x4096x64xf16> |
| ) -> tensor<20x4096x64xf16> { |
| %result = iree_linalg_ext.attention { root_op, |
| indexing_maps = [ |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2)>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d3, d2)>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d3, d4)>, |
| affine_map<(d0, d1, d2, d3, d4) -> ()>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d4)> |
| ] |
| } ins(%q, %k, %v, %scale : tensor<20x4096x64xf16>, tensor<20x4096x64xf16>, tensor<20x4096x64xf16>, f16) |
| outs(%output : tensor<20x4096x64xf16>) { |
| ^bb0(%score: f32): |
| iree_linalg_ext.yield %score : f32 |
| } -> tensor<20x4096x64xf16> |
| return %result : tensor<20x4096x64xf16> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| assert input_module is not None, "Failed to parse input MLIR module" |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 1 |
| assert root_op_list[0].name == "iree_linalg_ext.attention" |
| assert iree_codegen.isa_attention_op(root_op_list[0]) |
| |
| |
| @run |
| def test_igemm_conv_details(): |
| # Test 1: conv_2d_nhwc_hwcf. |
| module_str = """ |
| module { |
| func.func @conv_2d_nhwc_hwcf(%arg0: tensor<1x16x16x4xf32>, %arg1: tensor<3x3x4x16xf32>, %arg2: tensor<1x14x14x16xf32>) -> tensor<1x14x14x16xf32> { |
| %0 = linalg.conv_2d_nhwc_hwcf { root_op, dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64> } |
| ins(%arg0, %arg1 : tensor<1x16x16x4xf32>, tensor<3x3x4x16xf32>) |
| outs(%arg2 : tensor<1x14x14x16xf32>) -> tensor<1x14x14x16xf32> |
| return %0 : tensor<1x14x14x16xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| |
| details = iree_codegen.get_igemm_generic_conv_details(root_op_list[0]) |
| assert details is not None, "IGEMM details should be valid for NHWC_HWCF conv" |
| assert details.igemm_loop_bounds == [1, 14, 14, 16, 36] |
| |
| assert len(details.igemm_contraction_maps) == 3 |
| maps = [map_attr.value for map_attr in details.igemm_contraction_maps] |
| d0, d1, d2, d3, d4 = [AffineDimExpr.get(i) for i in range(5)] |
| # For channel-last (NHWC): input (N,H,W,K), filter (K,OC), output (N,H,W,OC). |
| assert maps[0] == AffineMap.get( |
| 5, 0, [d0, d1, d2, d4] |
| ), f"Input map mismatch: {maps[0]}" |
| assert maps[1] == AffineMap.get(5, 0, [d4, d3]), f"Filter map mismatch: {maps[1]}" |
| assert maps[2] == AffineMap.get( |
| 5, 0, [d0, d1, d2, d3] |
| ), f"Output map mismatch: {maps[2]}" |
| iter_types = [str(attr) for attr in details.igemm_loop_iterators] |
| assert iter_types == [ |
| '"parallel"', |
| '"parallel"', |
| '"parallel"', |
| '"parallel"', |
| '"reduction"', |
| ] |
| assert details.im2col_output_perm == [0, 1, 2, 3] |
| assert details.filter_reassoc_indices == [[0, 1, 2], [3]] |
| assert not details.is_output_channel_first |
| assert details.conv_to_igemm_dim_map == {0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 4, 6: 4} |
| |
| # Test 2: conv_2d_nhwc_fhwc. |
| module_str = """ |
| module { |
| func.func @conv_2d_nhwc_fhwc(%arg0: tensor<1x16x16x4xf32>, %arg1: tensor<16x3x3x4xf32>, %arg2: tensor<1x14x14x16xf32>) -> tensor<1x14x14x16xf32> { |
| %0 = linalg.conv_2d_nhwc_fhwc { root_op, dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64> } |
| ins(%arg0, %arg1 : tensor<1x16x16x4xf32>, tensor<16x3x3x4xf32>) |
| outs(%arg2 : tensor<1x14x14x16xf32>) -> tensor<1x14x14x16xf32> |
| return %0 : tensor<1x14x14x16xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| |
| details = iree_codegen.get_igemm_generic_conv_details(root_op_list[0]) |
| assert details is not None, "IGEMM details should be valid for NHWC_FHWC conv" |
| assert details.igemm_loop_bounds == [1, 14, 14, 16, 36] |
| assert len(details.igemm_contraction_maps) == 3 |
| maps = [map_attr.value for map_attr in details.igemm_contraction_maps] |
| # Verify expected affine maps (NHWC_FHWC layout). |
| d0, d1, d2, d3, d4 = [AffineDimExpr.get(i) for i in range(5)] |
| assert maps[0] == AffineMap.get( |
| 5, 0, [d0, d1, d2, d4] |
| ), f"Input map mismatch: {maps[0]}" |
| assert maps[1] == AffineMap.get(5, 0, [d3, d4]), f"Filter map mismatch: {maps[1]}" |
| assert maps[2] == AffineMap.get( |
| 5, 0, [d0, d1, d2, d3] |
| ), f"Output map mismatch: {maps[2]}" |
| iter_types = [str(attr) for attr in details.igemm_loop_iterators] |
| assert iter_types == [ |
| '"parallel"', |
| '"parallel"', |
| '"parallel"', |
| '"parallel"', |
| '"reduction"', |
| ] |
| assert details.im2col_output_perm == [0, 1, 2, 3] |
| assert details.filter_reassoc_indices == [[0], [1, 2, 3]] |
| assert not details.is_output_channel_first |
| assert details.conv_to_igemm_dim_map == {0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 4, 6: 4} |
| |
| # Test 3: conv_2d_nchw_fchw. |
| module_str = """ |
| module { |
| func.func @conv_2d_nchw_fchw(%arg0: tensor<1x4x16x16xf32>, %arg1: tensor<16x4x3x3xf32>, %arg2: tensor<1x16x14x14xf32>) -> tensor<1x16x14x14xf32> { |
| %0 = linalg.conv_2d_nchw_fchw { root_op, dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64> } |
| ins(%arg0, %arg1 : tensor<1x4x16x16xf32>, tensor<16x4x3x3xf32>) |
| outs(%arg2 : tensor<1x16x14x14xf32>) -> tensor<1x16x14x14xf32> |
| return %0 : tensor<1x16x14x14xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| |
| details = iree_codegen.get_igemm_generic_conv_details(root_op_list[0]) |
| assert details is not None, "IGEMM details should be valid for NCHW conv" |
| assert details.igemm_loop_bounds == [1, 16, 14, 14, 36] |
| assert len(details.igemm_contraction_maps) == 3 |
| maps = [map_attr.value for map_attr in details.igemm_contraction_maps] |
| # Verify expected affine maps for NCHW with loop dims [N, OC, H, W, K]. |
| # Note: operands are swapped - filter first, then input. |
| d0, d1, d2, d3, d4 = [AffineDimExpr.get(i) for i in range(5)] |
| assert maps[0] == AffineMap.get(5, 0, [d1, d4]), f"Filter map mismatch: {maps[0]}" |
| assert maps[1] == AffineMap.get( |
| 5, 0, [d0, d2, d3, d4] |
| ), f"Input map mismatch: {maps[1]}" |
| assert maps[2] == AffineMap.get( |
| 5, 0, [d0, d1, d2, d3] |
| ), f"Output map mismatch: {maps[2]}" |
| iter_types = [str(attr) for attr in details.igemm_loop_iterators] |
| assert iter_types == [ |
| '"parallel"', |
| '"parallel"', |
| '"parallel"', |
| '"parallel"', |
| '"reduction"', |
| ] |
| assert details.im2col_output_perm == [0, 1, 2, 3] |
| assert details.filter_reassoc_indices == [[0], [1, 2, 3]] |
| assert details.is_output_channel_first |
| assert details.conv_to_igemm_dim_map == {0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 4, 6: 4} |
| |
| # Test 4: linalg.generic with convolution pattern (weight backward). |
| module_str = """ |
| module { |
| func.func @conv_generic_weight_backward(%arg0: tensor<16x98x64x96xf32>, %arg1: tensor<16x96x64x96xf32>, %arg2: tensor<96x3x96xf32>) -> tensor<96x3x96xf32> { |
| %0 = linalg.generic { |
| indexing_maps = [ |
| affine_map<(d0, d1, d2, d3, d4, d5) -> (d3, d1 + d4, d5, d2)>, |
| affine_map<(d0, d1, d2, d3, d4, d5) -> (d3, d4, d5, d0)>, |
| affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2)> |
| ], |
| iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction", "reduction"] |
| } ins(%arg0, %arg1 : tensor<16x98x64x96xf32>, tensor<16x96x64x96xf32>) outs(%arg2 : tensor<96x3x96xf32>) attrs = {root_op} { |
| ^bb0(%in: f32, %in_1: f32, %out: f32): |
| %mul = arith.mulf %in, %in_1 : f32 |
| %add = arith.addf %out, %mul : f32 |
| linalg.yield %add : f32 |
| } -> tensor<96x3x96xf32> |
| return %0 : tensor<96x3x96xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| |
| details = iree_codegen.get_igemm_generic_conv_details(root_op_list[0]) |
| assert ( |
| details is not None |
| ), "IGEMM details should be valid for generic 1D conv weight backward" |
| assert details.igemm_loop_bounds == [96, 3, 96, 98304] |
| assert len(details.igemm_contraction_maps) == 3 |
| maps = [map_attr.value for map_attr in details.igemm_contraction_maps] |
| d0, d1, d2, d3 = [AffineDimExpr.get(i) for i in range(4)] |
| assert maps[0] == AffineMap.get(4, 0, [d3, d0]), f"Map 0 mismatch: {maps[0]}" |
| assert maps[1] == AffineMap.get(4, 0, [d1, d3, d2]), f"Map 1 mismatch: {maps[1]}" |
| assert maps[2] == AffineMap.get(4, 0, [d0, d1, d2]), f"Map 2 mismatch: {maps[2]}" |
| iter_types = [str(attr) for attr in details.igemm_loop_iterators] |
| assert iter_types == ['"parallel"', '"parallel"', '"parallel"', '"reduction"'] |
| assert details.im2col_output_perm == [1, 2, 0] |
| assert details.filter_reassoc_indices == [[0, 1, 2], [3]] |
| assert details.is_output_channel_first |
| assert details.conv_to_igemm_dim_map == {0: 0, 1: 1, 2: 2, 3: 3, 4: 3, 5: 3} |
| |
| # Test with a non-conv operation. |
| module_str = """ |
| module { |
| func.func @matmul(%arg0: tensor<4x4xf32>, %arg1: tensor<4x4xf32>, %arg2: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %0 = linalg.matmul { root_op } ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x4xf32>) outs(%arg2 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %0 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| matmul_op = root_op_list[0] |
| |
| details = iree_codegen.get_igemm_generic_conv_details(matmul_op) |
| assert details is None, "IGEMM details should be None for non-conv operation" |
| |
| |
| @run |
| def test_isa_scaled_contraction_op(): |
| # Test 1: Regular matmul is not a scaled contraction. |
| module_str = """ |
| module { |
| func.func @matmul(%arg0: tensor<4x4xf32>, %arg1: tensor<4x4xf32>, %arg2: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %0 = linalg.matmul { root_op } ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x4xf32>) outs(%arg2 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %0 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| assert input_module is not None, "Failed to parse input MLIR module" |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 1 |
| matmul_op = root_op_list[0] |
| |
| assert not iree_codegen.isa_scaled_contraction_op( |
| matmul_op |
| ), "Regular matmul should not be a scaled contraction" |
| |
| # Test 2: Fill op is not a scaled contraction. |
| module_str = """ |
| module { |
| func.func @fill(%arg0: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %cst = arith.constant 0.000000e+00 : f32 |
| %0 = linalg.fill { root_op } ins(%cst : f32) outs(%arg0 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %0 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 1 |
| fill_op = root_op_list[0] |
| |
| assert not iree_codegen.isa_scaled_contraction_op( |
| fill_op |
| ), "Fill op should not be a scaled contraction" |
| |
| # Test 3: Scaled matmul as linalg.generic should be detected. |
| # Pattern: linalg.generic with 5 indexing maps (lhs, rhs, lhs_scale, rhs_scale, output), |
| # and 4 iterator types (2 parallel for M,N; 2 reduction for Ko,Kb). |
| # Uses f4E2M1FN for operands and f8E8M0FNU for scales (matching real scaled matmul pattern). |
| module_str = """ |
| module { |
| func.func @scaled_matmul(%lhs: tensor<16x4x32xf4E2M1FN>, %rhs: tensor<16x4x32xf4E2M1FN>, |
| %lhs_scales: tensor<16x4xf8E8M0FNU>, %rhs_scales: tensor<16x4xf8E8M0FNU>, |
| %out: tensor<16x16xf32>) -> tensor<16x16xf32> { |
| %result = linalg.generic { |
| indexing_maps = [ |
| affine_map<(d0, d1, d2, d3) -> (d0, d2, d3)>, |
| affine_map<(d0, d1, d2, d3) -> (d1, d2, d3)>, |
| affine_map<(d0, d1, d2, d3) -> (d0, d2)>, |
| affine_map<(d0, d1, d2, d3) -> (d1, d2)>, |
| affine_map<(d0, d1, d2, d3) -> (d0, d1)> |
| ], |
| iterator_types = ["parallel", "parallel", "reduction", "reduction"], |
| root_op |
| } ins(%lhs, %rhs, %lhs_scales, %rhs_scales : tensor<16x4x32xf4E2M1FN>, tensor<16x4x32xf4E2M1FN>, tensor<16x4xf8E8M0FNU>, tensor<16x4xf8E8M0FNU>) |
| outs(%out : tensor<16x16xf32>) { |
| ^bb0(%a: f4E2M1FN, %b: f4E2M1FN, %a_scale: f8E8M0FNU, %b_scale: f8E8M0FNU, %acc: f32): |
| %a_scaled = arith.scaling_extf %a, %a_scale : f4E2M1FN, f8E8M0FNU to f32 |
| %b_scaled = arith.scaling_extf %b, %b_scale : f4E2M1FN, f8E8M0FNU to f32 |
| %prod = arith.mulf %a_scaled, %b_scaled : f32 |
| %sum = arith.addf %acc, %prod : f32 |
| linalg.yield %sum : f32 |
| } -> tensor<16x16xf32> |
| return %result : tensor<16x16xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 1, "Should have one root op" |
| |
| scaled_generic_op = root_op_list[0] |
| is_scaled = iree_codegen.isa_scaled_contraction_op(scaled_generic_op) |
| assert ( |
| is_scaled |
| ), "linalg.generic with scaled matmul pattern should be detected as scaled contraction" |
| |
| dims = iree_codegen.infer_scaled_contraction_dimensions(scaled_generic_op) |
| assert dims is not None, "Should be able to infer dimensions for scaled contraction" |
| |
| assert dims.m == [0], f"Got {dims.m}" |
| assert dims.n == [1], f"Got {dims.n}" |
| assert dims.k == [2], f"Got {dims.k}" |
| assert dims.kB == [3], f"Got {dims.kB}" |
| assert dims.batch == [], f"Got {dims.batch}" |
| |
| |
| @run |
| def test_infer_scaled_contraction_dimensions(): |
| # Test 1: Verify dimension inference on a scaled matmul operation. |
| module_str = """ |
| module { |
| func.func @scaled_matmul(%lhs: tensor<16x4x32xf4E2M1FN>, %rhs: tensor<16x4x32xf4E2M1FN>, |
| %lhs_scales: tensor<16x4xf8E8M0FNU>, %rhs_scales: tensor<16x4xf8E8M0FNU>, |
| %out: tensor<16x16xf32>) -> tensor<16x16xf32> { |
| %result = linalg.generic { |
| indexing_maps = [ |
| affine_map<(d0, d1, d2, d3) -> (d0, d2, d3)>, |
| affine_map<(d0, d1, d2, d3) -> (d1, d2, d3)>, |
| affine_map<(d0, d1, d2, d3) -> (d0, d2)>, |
| affine_map<(d0, d1, d2, d3) -> (d1, d2)>, |
| affine_map<(d0, d1, d2, d3) -> (d0, d1)> |
| ], |
| iterator_types = ["parallel", "parallel", "reduction", "reduction"], |
| root_op |
| } ins(%lhs, %rhs, %lhs_scales, %rhs_scales : tensor<16x4x32xf4E2M1FN>, tensor<16x4x32xf4E2M1FN>, tensor<16x4xf8E8M0FNU>, tensor<16x4xf8E8M0FNU>) |
| outs(%out : tensor<16x16xf32>) { |
| ^bb0(%a: f4E2M1FN, %b: f4E2M1FN, %a_scale: f8E8M0FNU, %b_scale: f8E8M0FNU, %acc: f32): |
| %a_scaled = arith.scaling_extf %a, %a_scale : f4E2M1FN, f8E8M0FNU to f32 |
| %b_scaled = arith.scaling_extf %b, %b_scale : f4E2M1FN, f8E8M0FNU to f32 |
| %prod = arith.mulf %a_scaled, %b_scaled : f32 |
| %sum = arith.addf %acc, %prod : f32 |
| linalg.yield %sum : f32 |
| } -> tensor<16x16xf32> |
| return %result : tensor<16x16xf32> |
| } |
| } |
| """ |
| input_module = ir.Module.parse(module_str) |
| root_op_list = iree_codegen.get_tuner_root_ops(input_module) |
| assert len(root_op_list) == 1, "Should have exactly one root op" |
| scaled_op = root_op_list[0] |
| |
| assert iree_codegen.isa_scaled_contraction_op( |
| scaled_op |
| ), "Operation should be recognized as scaled contraction" |
| |
| dims = iree_codegen.infer_scaled_contraction_dimensions(scaled_op) |
| assert dims is not None, "Should successfully infer dimensions" |
| assert dims.m == [0], f"Got {dims.m}" |
| assert dims.n == [1], f"Got {dims.n}" |
| assert dims.k == [2], f"Got {dims.k}" |
| assert dims.kB == [3], f"Got {dims.kB}" |
| assert dims.batch == [], f"Got {dims.batch}" |
| |
| # Test 2: Non-scaled contraction should return None. |
| module_str_regular = """ |
| module { |
| func.func @regular_matmul(%arg0: tensor<4x4xf32>, %arg1: tensor<4x4xf32>, %arg2: tensor<4x4xf32>) -> tensor<4x4xf32> { |
| %0 = linalg.matmul { root_op } ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x4xf32>) outs(%arg2 : tensor<4x4xf32>) -> tensor<4x4xf32> |
| return %0 : tensor<4x4xf32> |
| } |
| } |
| """ |
| input_module_regular = ir.Module.parse(module_str_regular) |
| regular_ops = iree_codegen.get_tuner_root_ops(input_module_regular) |
| assert len(regular_ops) == 1 |
| regular_matmul = regular_ops[0] |
| |
| # Regular matmul should not have scaled contraction dimensions. |
| # Check if all dimensions are empty (indicating it's not a scaled contraction). |
| dims_regular = iree_codegen.infer_scaled_contraction_dimensions(regular_matmul) |
| if dims_regular is not None: |
| all_empty = ( |
| len(dims_regular.m) == 0 |
| and len(dims_regular.n) == 0 |
| and len(dims_regular.k) == 0 |
| and len(dims_regular.kB) == 0 |
| and len(dims_regular.batch) == 0 |
| ) |
| assert ( |
| all_empty or dims_regular is None |
| ), "Regular matmul should not have valid scaled contraction dimensions" |
| |
| # Test 3: Batched scaled matmul. |
| module_str_batched = """ |
| module { |
| func.func @batched_scaled_matmul(%lhs: tensor<8x16x4x32xf4E2M1FN>, %rhs: tensor<8x16x4x32xf4E2M1FN>, |
| %lhs_scales: tensor<8x16x4xf8E8M0FNU>, %rhs_scales: tensor<8x16x4xf8E8M0FNU>, |
| %out: tensor<8x16x16xf32>) -> tensor<8x16x16xf32> { |
| %result = linalg.generic { |
| indexing_maps = [ |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d3, d4)>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d2, d3, d4)>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d3)>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d2, d3)>, |
| affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2)> |
| ], |
| iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"], |
| root_op |
| } ins(%lhs, %rhs, %lhs_scales, %rhs_scales : tensor<8x16x4x32xf4E2M1FN>, tensor<8x16x4x32xf4E2M1FN>, tensor<8x16x4xf8E8M0FNU>, tensor<8x16x4xf8E8M0FNU>) |
| outs(%out : tensor<8x16x16xf32>) { |
| ^bb0(%a: f4E2M1FN, %b: f4E2M1FN, %a_scale: f8E8M0FNU, %b_scale: f8E8M0FNU, %acc: f32): |
| %a_scaled = arith.scaling_extf %a, %a_scale : f4E2M1FN, f8E8M0FNU to f32 |
| %b_scaled = arith.scaling_extf %b, %b_scale : f4E2M1FN, f8E8M0FNU to f32 |
| %prod = arith.mulf %a_scaled, %b_scaled : f32 |
| %sum = arith.addf %acc, %prod : f32 |
| linalg.yield %sum : f32 |
| } -> tensor<8x16x16xf32> |
| return %result : tensor<8x16x16xf32> |
| } |
| } |
| """ |
| input_module_batched = ir.Module.parse(module_str_batched) |
| batched_ops = iree_codegen.get_tuner_root_ops(input_module_batched) |
| assert len(batched_ops) == 1, "Batched op should be found" |
| batched_op = batched_ops[0] |
| assert iree_codegen.isa_scaled_contraction_op( |
| batched_op |
| ), "Batched scaled matmul should be recognized" |
| |
| dims_batched = iree_codegen.infer_scaled_contraction_dimensions(batched_op) |
| assert ( |
| dims_batched is not None |
| ), "Batch dimension must be present in batched scaled matmul" |
| assert dims_batched.batch == [0], f"Got {dims_batched.batch}" |
| assert dims_batched.m == [1], f"Got {dims_batched.m}" |
| assert dims_batched.n == [2], f"Got {dims_batched.n}" |
| assert dims_batched.k == [3], f"Got {dims_batched.k}" |
| assert dims_batched.kB == [4], f"Got {dims_batched.kB}" |
