[tuner]: retire dataclass GPUPipelineOptions

Signed-off-by: Bangtian Liu <[email protected]>

tuner/tuner/candidate_gen_test.py

@@ -55,7 +55,9 @@ def test_apply_params_mmt(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[8, 8, 8],
         subgroup_m_count=16,
         subgroup_n_count=16,
-        gpu_pipeline_options=common.GpuPipelineOptions(prefetch_shared_memory=True),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(
+            prefetch_shared_memory=True
+        ),
         waves_per_eu=8,
     )
 
@@ -109,8 +111,10 @@ def test_apply_params_conv(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[464, 320, 16],
         subgroup_m_count=1,
         subgroup_n_count=4,
-        gpu_pipeline_options=common.GpuPipelineOptions(
-            reorder_workgroups_strategy=common.ReorderWorkgroupsStrategy.TRANSPOSE
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(
+            reorder_workgroups_strategy=iree_gpu.ReorderWorkgroupsStrategyAttr.get(
+                iree_gpu.ReorderWorkgroupsStrategy.Transpose
+            )
         ),
         waves_per_eu=2,
     )
@@ -131,7 +135,6 @@ def test_apply_params_conv(tuner_ctx: common.TunerContext) -> None:
 
     assert modified
     modified = remove_comments(modified)
-
     assert embeddable
     assert (
         "intrinsic = #iree_gpu.mma_layout<MFMA_F32_16x16x16_F16>, subgroup_m_count = 1, subgroup_n_count = 4"
@@ -143,7 +146,7 @@ def test_apply_params_conv(tuner_ctx: common.TunerContext) -> None:
     )
     assert "tile_sizes = [[1, 1, 464, 320, 1, 1, 16]]" in modified
     assert (
-        "gpu_pipeline_options = #iree_gpu.pipeline_options<reorder_workgroups_strategy = Transpose>"
+        "gpu_pipeline_options = #iree_gpu.pipeline_options<reorder_workgroups_strategy = <Transpose>>"
         in modified
     )
     assert '{llvm_func_attrs = {"amdgpu-waves-per-eu" = "2"}' in modified
@@ -175,7 +178,7 @@ def test_apply_params_contract(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[480, 384, 32],
         subgroup_m_count=1,
         subgroup_n_count=4,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=2,
     )
 
@@ -224,7 +227,7 @@ def test_apply_params_batch_matmul(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[416, 320, 128],
         subgroup_m_count=2,
         subgroup_n_count=2,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=2,
     )
 
@@ -276,7 +279,7 @@ def test_apply_params_batch_mmt_float(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[128, 64, 128],
         subgroup_m_count=2,
         subgroup_n_count=2,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=2,
     )
 
@@ -326,7 +329,7 @@ def test_apply_params_batch_mmt_int(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[128, 64, 128],
         subgroup_m_count=2,
         subgroup_n_count=2,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=4,
     )
 
@@ -399,7 +402,7 @@ def test_apply_params_broadcast_rhs_mmt(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[128, 64, 128],
         subgroup_m_count=2,
         subgroup_n_count=2,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=4,
     )
 

tuner/tuner/common.py

@@ -114,35 +114,6 @@ def __str__(self) -> str:
         return self.name.title()
 
 
-@dataclass
-class GpuPipelineOptions:
-    """Represents the `iree_gpu.pipeline_options` attribute"""
-
-    prefetch_shared_memory: Optional[bool] = None
-    no_reduce_shared_memory_bank_conflicts: Optional[bool] = None
-    reorder_workgroups_strategy: Optional[ReorderWorkgroupsStrategy] = None
-
-    def all_default(self) -> bool:
-        return all(x is None for x in astuple(self))
-
-    def __str__(self) -> str:
-        options: list[str] = []
-        if self.prefetch_shared_memory is not None:
-            options.append(
-                f"prefetch_shared_memory = {str(self.prefetch_shared_memory).lower()}"
-            )
-        if self.no_reduce_shared_memory_bank_conflicts is not None:
-            options.append(
-                f"no_reduce_shared_memory_bank_conflicts = {str(self.no_reduce_shared_memory_bank_conflicts).lower()}"
-            )
-        if self.reorder_workgroups_strategy is not None:
-            options.append(
-                f"reorder_workgroups_strategy = {self.reorder_workgroups_strategy}"
-            )
-
-        return f"#iree_gpu.pipeline_options<{', '.join(options)}>"
-
-
 @dataclass
 class Configuration:
     subgroup_size: int
@@ -151,14 +122,21 @@ class Configuration:
     tile_sizes: list[int]
     subgroup_m_count: int
     subgroup_n_count: int
-    gpu_pipeline_options: GpuPipelineOptions
+    gpu_pipeline_options: iree_gpu.PipelineOptionsAttr
     waves_per_eu: int
 
 
 def get_pipeline_config(configuration: Configuration) -> str:
     extra_config = ""
-    if not configuration.gpu_pipeline_options.all_default():
-        extra_config += f", gpu_pipeline_options = {configuration.gpu_pipeline_options}"
+    pipeline_options = configuration.gpu_pipeline_options
+    if (
+        pipeline_options.prefetch_shared_memory is not None
+        or pipeline_options.no_reduce_shared_memory_bank_conflicts is not None
+        or pipeline_options.use_igemm_convolution is not None
+        or pipeline_options.reorder_workgroups_strategy is not None
+    ):
+        extra_config += f", gpu_pipeline_options = {pipeline_options}"
+
     if configuration.waves_per_eu != 2:
         extra_config += f', llvm_func_attrs = {{"amdgpu-waves-per-eu" = "{configuration.waves_per_eu}"}}'
     return extra_config

tuner/tuner/common_test.py

@@ -47,27 +47,29 @@ def test_get_shaped_type_to_str(tuner_ctx: common.TunerContext) -> None:
     assert str(common.ShapedType([-1, 2, 3], tuner_ctx.type.f16)) == "?x2x3xf16"
 
 
-def test_gpu_pipeline_options() -> None:
-    options = common.GpuPipelineOptions()
-    assert options.all_default()
+def test_gpu_pipeline_options(tuner_ctx: common.TunerContext) -> None:
+    options = iree_gpu.PipelineOptionsAttr.get()
     assert str(options) == "#iree_gpu.pipeline_options<>"
 
-    options.prefetch_shared_memory = True
-    assert not options.all_default()
+    options = iree_gpu.PipelineOptionsAttr.get(prefetch_shared_memory=True)
     assert str(options) == "#iree_gpu.pipeline_options<prefetch_shared_memory = true>"
 
-    options.no_reduce_shared_memory_bank_conflicts = False
+    options = iree_gpu.PipelineOptionsAttr.get(
+        prefetch_shared_memory=True, no_reduce_shared_memory_bank_conflicts=False
+    )
     assert (
         str(options)
         == "#iree_gpu.pipeline_options<prefetch_shared_memory = true, no_reduce_shared_memory_bank_conflicts = false>"
     )
 
-    options = common.GpuPipelineOptions()
-    options.reorder_workgroups_strategy = common.ReorderWorkgroupsStrategy.TRANSPOSE
-    assert not options.all_default()
+    options = iree_gpu.PipelineOptionsAttr.get(
+        reorder_workgroups_strategy=iree_gpu.ReorderWorkgroupsStrategyAttr.get(
+            iree_gpu.ReorderWorkgroupsStrategy.Transpose
+        )
+    )
     assert (
         str(options)
-        == "#iree_gpu.pipeline_options<reorder_workgroups_strategy = Transpose>"
+        == "#iree_gpu.pipeline_options<reorder_workgroups_strategy = <Transpose>>"
     )
 
 
@@ -81,7 +83,7 @@ def test_get_pipeline_config(mlir_ctx: ir.Context) -> None:
         tile_sizes=[4, 8, 16],
         subgroup_m_count=1,
         subgroup_n_count=1,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=2,
     )
     config1_str: str = common.get_pipeline_config(config)
@@ -91,7 +93,9 @@ def test_get_pipeline_config(mlir_ctx: ir.Context) -> None:
     config2_str: str = common.get_pipeline_config(config)
     assert config2_str == ', llvm_func_attrs = {"amdgpu-waves-per-eu" = "4"}'
 
-    config.gpu_pipeline_options.prefetch_shared_memory = True
+    config.gpu_pipeline_options = iree_gpu.PipelineOptionsAttr.get(
+        prefetch_shared_memory=True
+    )
     config3_str = common.get_pipeline_config(config)
     assert (
         config3_str

tuner/tuner/dispatch_constraints.py

@@ -236,7 +236,7 @@ def generate_solutions(
             [lookup(m), lookup(n), lookup(k)],
             lookup(sg_m_cnt),
             lookup(sg_n_cnt),
-            GpuPipelineOptions(),
+            iree_gpu.PipelineOptionsAttr.get(),
             lookup(waves_per_eu),
         )
         solver.add(z3.simplify(z3.Not(z3.And(list(x == model[x] for x in all_vars)))))

tuner/tuner/dispatch_parser_test.py

@@ -49,7 +49,7 @@ def test_get_mmt_tile_sizes(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[128, 320, 32],
         subgroup_m_count=0,
         subgroup_n_count=0,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=0,
     )
     assert dispatch_parser.get_mmt_tile_sizes(config) == [128, 320, 32]
@@ -65,7 +65,7 @@ def test_get_conv_tile_sizes(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[464, 320, 16],
         subgroup_m_count=1,
         subgroup_n_count=4,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=1,
     )
     assert dispatch_parser.ConvParser().get_conv_tile_sizes(config) == [
@@ -89,7 +89,7 @@ def test_get_contract_tile_sizes(tuner_ctx: common.TunerContext) -> None:
         tile_sizes=[4, 8, 16],
         subgroup_m_count=1,
         subgroup_n_count=1,
-        gpu_pipeline_options=common.GpuPipelineOptions(),
+        gpu_pipeline_options=iree_gpu.PipelineOptionsAttr.get(),
         waves_per_eu=2,
     )
     assert dispatch_parser.get_contract_tile_sizes(config, "mnk") == [4, 8, 16]