Converting local CPU target to support executable-create2.

It already did for the inline HAL so this is mostly just removing the pipeline layouts from the full HAL implementation.
iree-org · Aug 20, 2024 · a69a93b · a69a93b
1 parent 7121ece
commit a69a93b
Show file tree

Hide file tree

Showing 33 changed files with 123 additions and 929 deletions.
diff --git a/compiler/plugins/target/LLVMCPU/LLVMCPUTarget.cpp b/compiler/plugins/target/LLVMCPU/LLVMCPUTarget.cpp
@@ -398,10 +398,11 @@ class LLVMCPUTargetBackend final : public TargetBackend {
 
       // Specify the constant and binding information used to validate
       // dispatches.
-      // TODO(#18154): pack per-binding information bitfields.
-      dispatchAttrs.constantCount = exportOp.getLayout().getPushConstants();
-      dispatchAttrs.bindingCount =
-          exportOp.getLayout().getSetLayout(0).getBindings().size();
+      if (auto layoutAttr = exportOp.getLayout()) {
+        dispatchAttrs.constantCount = layoutAttr.getPushConstants();
+        dispatchAttrs.bindingCount =
+            layoutAttr.getSetLayout(0).getBindings().size();
+      }
 
       LibraryBuilder::SourceLocation sourceLocation;
       if (options.debugLevel >= 1) {

diff --git a/compiler/plugins/target/LLVMCPU/LibraryBuilder.cpp b/compiler/plugins/target/LLVMCPU/LibraryBuilder.cpp
@@ -112,7 +112,9 @@ makeDispatchFunctionType(llvm::LLVMContext &context) {
 // %struct.iree_hal_executable_dispatch_attrs_v0_t = type {
 //   i16,
 //   i8,
-//   i8
+//   i8,
+//   i32,
+//   i64[8]
 // }
 static llvm::StructType *makeDispatchAttrsType(llvm::LLVMContext &context) {
   if (auto *existingType = llvm::StructType::getTypeByName(
@@ -121,12 +123,20 @@ static llvm::StructType *makeDispatchAttrsType(llvm::LLVMContext &context) {
   }
   auto *i8Type = llvm::IntegerType::getInt8Ty(context);
   auto *i16Type = llvm::IntegerType::getInt16Ty(context);
+  auto *i32Type = llvm::IntegerType::getInt32Ty(context);
+  auto *i64Type = llvm::IntegerType::getInt64Ty(context);
   auto *type =
       llvm::StructType::create(context,
                                {
-                                   i16Type,
-                                   i8Type,
-                                   i8Type,
+                                   i16Type, i8Type, i8Type, i32Type,
+                                   i64Type, // [0]
+                                   i64Type, // [1]
+                                   i64Type, // [2]
+                                   i64Type, // [3]
+                                   i64Type, // [4]
+                                   i64Type, // [5]
+                                   i64Type, // [6]
+                                   i64Type, // [7]
                                },
                                "iree_hal_executable_dispatch_attrs_v0_t",
                                /*isPacked=*/false);
@@ -490,6 +500,7 @@ LibraryBuilder::buildLibraryV0ExportTable(std::string libraryName) {
   auto *i8Type = llvm::IntegerType::getInt8Ty(context);
   auto *i16Type = llvm::IntegerType::getInt16Ty(context);
   auto *i32Type = llvm::IntegerType::getInt32Ty(context);
+  auto *i64Type = llvm::IntegerType::getInt64Ty(context);
 
   // iree_hal_executable_export_table_v0_t::ptrs
   SmallVector<llvm::Constant *> exportPtrValues;
@@ -520,6 +531,24 @@ LibraryBuilder::buildLibraryV0ExportTable(std::string libraryName) {
               llvm::ConstantInt::get(i8Type, dispatch.attrs.constantCount),
               // binding_count=
               llvm::ConstantInt::get(i8Type, dispatch.attrs.bindingCount),
+              // reserved_0=
+              llvm::ConstantInt::get(i32Type, 0),
+              // reserved_1[0]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[1]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[2]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[3]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[4]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[5]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[6]=
+              llvm::ConstantInt::get(i64Type, 0),
+              // reserved_1[7]=
+              llvm::ConstantInt::get(i64Type, 0),
           }));
     }
     exportAttrs = createArrayConstant(libraryName + "_attrs", dispatchAttrsType,

diff --git a/compiler/plugins/target/LLVMCPU/test/materialize_homogeneous_encodings.mlir b/compiler/plugins/target/LLVMCPU/test/materialize_homogeneous_encodings.mlir
@@ -6,7 +6,7 @@
 #map2 = affine_map<(d0, d1, d2) -> (d2, d1)>
 #map3 = affine_map<(d0, d1, d2) -> (d0, d1)>
 #encoding = #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3], round_dims_to = array<i64: 16, 16, 16>>
-#device_target_llvm_cpu = #hal.device.target<"llvm-cpu", [#executable_target_embedded_elf_x86_64_]> : !hal.device
+#device_target_llvm_cpu = #hal.device.target<"local", [#executable_target_embedded_elf_x86_64_]> : !hal.device
 module attributes {hal.device.targets = [#device_target_llvm_cpu]} {
   util.func public @lhs_encoding(%arg0: tensor<?x?xf32>) -> tensor<?x?xf32> {
     %3 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #encoding>

diff --git a/compiler/plugins/target/LLVMCPU/test/smoketest_embedded.mlir b/compiler/plugins/target/LLVMCPU/test/smoketest_embedded.mlir
@@ -3,7 +3,7 @@
 
 module attributes {
   hal.device.targets = [
-    #hal.device.target<"llvm-cpu", [
+    #hal.device.target<"local", [
       #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64", {
         native_vector_size = 16 : index
       }>

diff --git a/compiler/plugins/target/LLVMCPU/test/smoketest_system.mlir b/compiler/plugins/target/LLVMCPU/test/smoketest_system.mlir
@@ -5,7 +5,7 @@
 
 module attributes {
   hal.device.targets = [
-    #hal.device.target<"llvm-cpu", [
+    #hal.device.target<"local", [
       #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64", {
         native_vector_size = 16 : index
       }>

diff --git a/compiler/plugins/target/VMVX/VMVXTarget.cpp b/compiler/plugins/target/VMVX/VMVXTarget.cpp
@@ -132,7 +132,6 @@ class VMVXTargetBackend final : public TargetBackend {
 
       // Specify the constant and binding information used to validate
       // dispatches.
-      // TODO(#18154): pack per-binding information bitfields.
       if (auto layoutAttr = exportOp.getLayout()) {
         int64_t constantCount = layoutAttr.getPushConstants();
         if (constantCount > 0) {

diff --git a/compiler/plugins/target/VulkanSPIRV/test/materialize_homogeneous_encodings.mlir b/compiler/plugins/target/VulkanSPIRV/test/materialize_homogeneous_encodings.mlir
@@ -26,7 +26,7 @@ module attributes {hal.device.targets = [#device_target_vulkan]} {
 #map3 = affine_map<(d0, d1, d2) -> (d0, d1)>
 #encoding = #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3], round_dims_to = array<i64: 16, 16, 16>>
 #executable_target_embedded_elf_x86_64_ = #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64", {target_triple = "x86_64-none-elf", cpu_features = "+avx512f"}>
-#device_target_llvm_cpu = #hal.device.target<"llvm-cpu", [#executable_target_embedded_elf_x86_64_]> : !hal.device
+#device_target_llvm_cpu = #hal.device.target<"local", [#executable_target_embedded_elf_x86_64_]> : !hal.device
 #executable_target_vulkan_spirv_fb = #hal.executable.target<"vulkan-spirv", "vulkan-spirv-fb">
 #device_target_vulkan = #hal.device.target<"vulkan", [#executable_target_vulkan_spirv_fb]> : !hal.device
 module attributes {hal.device.targets = [#hal.device.select<[#device_target_vulkan, #device_target_llvm_cpu]> : !hal.device]} {

diff --git a/compiler/src/iree/compiler/Dialect/HAL/IR/HALAttrs.td b/compiler/src/iree/compiler/Dialect/HAL/IR/HALAttrs.td
@@ -744,7 +744,7 @@ def HAL_DeviceTargetAttr : AttrDef<HAL_Dialect, "DeviceTarget", [
 
     Example:
     ```mlir
-    #hal.device.target<"llvm-cpu", {
+    #hal.device.target<"local", {
       device_configuration = ...
     }, [
       #hal.executable.target<"llvm-cpu", "embedded-elf-arm_32">,

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Target/Devices/LocalDevice.cpp b/compiler/src/iree/compiler/Dialect/HAL/Target/Devices/LocalDevice.cpp
@@ -36,11 +36,14 @@ LocalDevice::LocalDevice(const LocalDevice::Options options)
 IREE::HAL::DeviceTargetAttr LocalDevice::getDefaultDeviceTarget(
     MLIRContext *context, const TargetRegistry &targetRegistry) const {
   Builder b(context);
-  SmallVector<NamedAttribute> configItems;
 
-  // TODO(benvanik): flags for common queries.
+  SmallVector<NamedAttribute> deviceConfigAttrs;
+  deviceConfigAttrs.emplace_back(b.getStringAttr("executable_create_2"),
+                                 b.getUnitAttr());
+  auto deviceConfigAttr = b.getDictionaryAttr(deviceConfigAttrs);
 
-  auto configAttr = b.getDictionaryAttr(configItems);
+  SmallVector<NamedAttribute> executableConfigAttrs;
+  auto executableConfigAttr = b.getDictionaryAttr(executableConfigAttrs);
 
   SmallVector<IREE::HAL::ExecutableTargetAttr> executableTargetAttrs;
   for (auto backendName : options.defaultTargetBackends) {
@@ -50,23 +53,27 @@ IREE::HAL::DeviceTargetAttr LocalDevice::getDefaultDeviceTarget(
                    << "\n";
       return {};
     }
-    targetBackend->getDefaultExecutableTargets(context, "local", configAttr,
-                                               executableTargetAttrs);
+    targetBackend->getDefaultExecutableTargets(
+        context, "local", executableConfigAttr, executableTargetAttrs);
   }
 
   return IREE::HAL::DeviceTargetAttr::get(context, b.getStringAttr("local"),
-                                          configAttr, executableTargetAttrs);
+                                          deviceConfigAttr,
+                                          executableTargetAttrs);
 }
 
 std::optional<IREE::HAL::DeviceTargetAttr>
 LocalDevice::getHostDeviceTarget(MLIRContext *context,
                                  const TargetRegistry &targetRegistry) const {
   Builder b(context);
-  SmallVector<NamedAttribute> configItems;
 
-  // TODO(benvanik): flags for overrides or ask LLVM for info about the host.
+  SmallVector<NamedAttribute> deviceConfigAttrs;
+  deviceConfigAttrs.emplace_back(b.getStringAttr("executable_create_2"),
+                                 b.getUnitAttr());
+  auto deviceConfigAttr = b.getDictionaryAttr(deviceConfigAttrs);
 
-  auto configAttr = b.getDictionaryAttr(configItems);
+  SmallVector<NamedAttribute> executableConfigAttrs;
+  auto executableConfigAttr = b.getDictionaryAttr(executableConfigAttrs);
 
   SmallVector<IREE::HAL::ExecutableTargetAttr> executableTargetAttrs;
   for (auto backendName : options.defaultHostBackends) {
@@ -76,12 +83,13 @@ LocalDevice::getHostDeviceTarget(MLIRContext *context,
                    << "\n";
       return std::nullopt;
     }
-    targetBackend->getHostExecutableTargets(context, "local", configAttr,
-                                            executableTargetAttrs);
+    targetBackend->getHostExecutableTargets(
+        context, "local", executableConfigAttr, executableTargetAttrs);
   }
 
   return IREE::HAL::DeviceTargetAttr::get(context, b.getStringAttr("local"),
-                                          configAttr, executableTargetAttrs);
+                                          deviceConfigAttr,
+                                          executableTargetAttrs);
 }
 
 Value LocalDevice::buildDeviceTargetMatch(

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/dump_executable_benchmarks.mlir b/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/dump_executable_benchmarks.mlir
@@ -6,7 +6,7 @@
 // Ensure devices are copied and made available:
 #executable_target_embedded_elf_x86_64 = #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64">
 // CHECK: util.global private @device
-util.global private @device = #hal.device.target<"llvm-cpu", [
+util.global private @device = #hal.device.target<"local", [
   #executable_target_embedded_elf_x86_64
 ]> : !hal.device
 
@@ -174,10 +174,10 @@ util.func public @main(%dynamic_arg: i32) -> !stream.timepoint attributes {
 
 #executable_target_embedded_elf_aarch64 = #hal.executable.target<"llvm-cpu", "embedded-elf-aarch64">
 #executable_target_embedded_elf_x86_64 = #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64">
-util.global private @device_a = #hal.device.target<"llvm-cpu", [
+util.global private @device_a = #hal.device.target<"local", [
   #executable_target_embedded_elf_aarch64
 ]> : !hal.device
-util.global private @device_b = #hal.device.target<"llvm-cpu", [
+util.global private @device_b = #hal.device.target<"local", [
   #executable_target_embedded_elf_x86_64
 ]> : !hal.device
 

diff --git a/...r/src/iree/compiler/Dialect/HAL/Transforms/test/materialize_dispatch_instrumentation.mlir b/...r/src/iree/compiler/Dialect/HAL/Transforms/test/materialize_dispatch_instrumentation.mlir
@@ -1,7 +1,7 @@
 // RUN: iree-opt --split-input-file --pass-pipeline='builtin.module(iree-hal-materialize-dispatch-instrumentation{buffer-size=64mib})' %s | FileCheck %s
 
 module attributes {hal.device.targets = [
-  #hal.device.target<"llvm-cpu", [
+  #hal.device.target<"local", [
     #hal.executable.target<"llvm-cpu", "embedded-elf-arm_64">,
     #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64">
   ]> : !hal.device

diff --git a/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/verify_devices.mlir b/compiler/src/iree/compiler/Dialect/HAL/Transforms/test/verify_devices.mlir
@@ -53,7 +53,7 @@ module @module {
   util.global private @optional = #hal.device.fallback<@device> : !hal.device
   util.global private @ordinal = #hal.device.ordinal<0> : !hal.device
   util.global private @selected = #hal.device.select<[
-    #hal.device.target<"llvm-cpu"> : !hal.device,
+    #hal.device.target<"local"> : !hal.device,
     #hal.device.target<"vmvx"> : !hal.device
   ]> : !hal.device
   util.func private @func() -> () attributes {

diff --git a/docs/website/docs/community/blog/posts/microkernels.md b/docs/website/docs/community/blog/posts/microkernels.md
@@ -338,7 +338,7 @@ This then goes to the LLVM x86 backend, which produces x86 assembly.
 [...]
 // -----// IR Dump After Inliner (inline) //----- //
 #executable_target_embedded_elf_x86_64_ = #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64", {cpu = "znver4", cpu_features = "+mmx,+popcnt,+sse,+sse2,+sse3,+ssse3,+sse4.1,+sse4.2,+avx,+avx2,+sse4a,+fma,+avx512f,+bmi,+bmi2,+aes,+pclmul,+avx512vl,+avx512bw,+avx512dq,+avx512cd,+avx512vbmi,+avx512ifma,+avx512vpopcntdq,+avx512vbmi2,+gfni,+vpclmulqdq,+avx512vnni,+avx512bitalg,+avx512bf16,+adx,+clflushopt,+clwb,+clzero,+cx16,+cx8,+crc32,+f16c,+fsgsbase,+fxsr,+invpcid,+lzcnt,+movbe,+mwaitx,+pku,+prfchw,+rdpid,+rdpru,+rdrnd,+rdseed,+sahf,+sha,+shstk,+vaes,+wbnoinvd,+x87,+xsave,+xsavec,+xsaveopt,+xsaves,+evex512", data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", native_vector_size = 64 : index, target_triple = "x86_64-unknown-unknown-eabi-elf", ukernels = "all"}>
-#device_target_llvm_cpu = #hal.device.target<"llvm-cpu", {executable_targets = [#executable_target_embedded_elf_x86_64_]}> : !hal.device
+#device_target_llvm_cpu = #hal.device.target<"local", {executable_targets = [#executable_target_embedded_elf_x86_64_]}> : !hal.device
 module attributes {hal.device.targets = [#device_target_llvm_cpu]} {
   func.func @matmul_dynamic(%arg0: !hal.buffer_view, %arg1: !hal.buffer_view, %arg2: !hal.buffer_view) -> !hal.buffer_view attributes {iree.abi.stub, iree.reflection = {iree.abi.declaration = "sync func @matmul_dynamic(%input0: tensor<?x?xf32>, %input1: tensor<?x?xf32>, %input2: tensor<?x?xf32>) -> (%output0: tensor<?x?xf32>)"}} {
     %0 = hal.buffer_view.dim<%arg0 : !hal.buffer_view>[0] : index
@@ -367,7 +367,7 @@ module attributes {hal.device.targets = [#device_target_llvm_cpu]} {
 // -----// IR Dump After CSE (cse) //----- //
 #executable_target_embedded_elf_x86_64_ = #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64", {cpu = "znver4", cpu_features = "+mmx,+popcnt,+sse,+sse2,+sse3,+ssse3,+sse4.1,+sse4.2,+avx,+avx2,+sse4a,+fma,+avx512f,+bmi,+bmi2,+aes,+pclmul,+avx512vl,+avx512bw,+avx512dq,+avx512cd,+avx512vbmi,+avx512ifma,+avx512vpopcntdq,+avx512vbmi2,+gfni,+vpclmulqdq,+avx512vnni,+avx512bitalg,+avx512bf16,+adx,+clflushopt,+clwb,+clzero,+cx16,+cx8,+crc32,+f16c,+fsgsbase,+fxsr,+invpcid,+lzcnt,+movbe,+mwaitx,+pku,+prfchw,+rdpid,+rdpru,+rdrnd,+rdseed,+sahf,+sha,+shstk,+vaes,+wbnoinvd,+x87,+xsave,+xsavec,+xsaveopt,+xsaves,+evex512", data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", native_vector_size = 64 : index, target_triple = "x86_64-unknown-unknown-eabi-elf", ukernels = "all"}>
 #map = affine_map<()[s0] -> (s0 ceildiv 16)>
-#device_target_llvm_cpu = #hal.device.target<"llvm-cpu", {executable_targets = [#executable_target_embedded_elf_x86_64_]}> : !hal.device
+#device_target_llvm_cpu = #hal.device.target<"local", {executable_targets = [#executable_target_embedded_elf_x86_64_]}> : !hal.device
 module attributes {hal.device.targets = [#device_target_llvm_cpu]} {
   func.func @matmul_dynamic(%arg0: !hal.buffer_view, %arg1: !hal.buffer_view, %arg2: !hal.buffer_view) -> !hal.buffer_view attributes {iree.abi.stub, iree.reflection = {iree.abi.declaration = "sync func @matmul_dynamic(%input0: tensor<?x?xf32>, %input1: tensor<?x?xf32>, %input2: tensor<?x?xf32>) -> (%output0: tensor<?x?xf32>)"}} {
     %cst = arith.constant 0.000000e+00 : f32

diff --git a/runtime/src/iree/hal/drivers/local_sync/sync_device.c b/runtime/src/iree/hal/drivers/local_sync/sync_device.c
@@ -17,7 +17,6 @@
 #include "iree/hal/local/executable_environment.h"
 #include "iree/hal/local/inline_command_buffer.h"
 #include "iree/hal/local/local_executable_cache.h"
-#include "iree/hal/local/local_pipeline_layout.h"
 #include "iree/hal/utils/deferred_command_buffer.h"
 #include "iree/hal/utils/file_transfer.h"
 #include "iree/hal/utils/memory_file.h"
@@ -247,17 +246,6 @@ static iree_status_t iree_hal_sync_device_create_command_buffer(
   }
 }
 
-static iree_status_t iree_hal_sync_device_create_descriptor_set_layout(
-    iree_hal_device_t* base_device,
-    iree_hal_descriptor_set_layout_flags_t flags,
-    iree_host_size_t binding_count,
-    const iree_hal_descriptor_set_layout_binding_t* bindings,
-    iree_hal_descriptor_set_layout_t** out_descriptor_set_layout) {
-  return iree_hal_local_descriptor_set_layout_create(
-      flags, binding_count, bindings,
-      iree_hal_device_host_allocator(base_device), out_descriptor_set_layout);
-}
-
 static iree_status_t iree_hal_sync_device_create_event(
     iree_hal_device_t* base_device, iree_hal_queue_affinity_t queue_affinity,
     iree_hal_event_flags_t flags, iree_hal_event_t** out_event) {
@@ -290,16 +278,6 @@ static iree_status_t iree_hal_sync_device_import_file(
       iree_hal_device_host_allocator(base_device), out_file);
 }
 
-static iree_status_t iree_hal_sync_device_create_pipeline_layout(
-    iree_hal_device_t* base_device, iree_host_size_t push_constants,
-    iree_host_size_t set_layout_count,
-    iree_hal_descriptor_set_layout_t* const* set_layouts,
-    iree_hal_pipeline_layout_t** out_pipeline_layout) {
-  return iree_hal_local_pipeline_layout_create(
-      push_constants, set_layout_count, set_layouts,
-      iree_hal_device_host_allocator(base_device), out_pipeline_layout);
-}
-
 static iree_status_t iree_hal_sync_device_create_semaphore(
     iree_hal_device_t* base_device, uint64_t initial_value,
     iree_hal_semaphore_flags_t flags, iree_hal_semaphore_t** out_semaphore) {
@@ -540,12 +518,9 @@ static const iree_hal_device_vtable_t iree_hal_sync_device_vtable = {
     .query_i64 = iree_hal_sync_device_query_i64,
     .create_channel = iree_hal_sync_device_create_channel,
     .create_command_buffer = iree_hal_sync_device_create_command_buffer,
-    .create_descriptor_set_layout =
-        iree_hal_sync_device_create_descriptor_set_layout,
     .create_event = iree_hal_sync_device_create_event,
     .create_executable_cache = iree_hal_sync_device_create_executable_cache,
     .import_file = iree_hal_sync_device_import_file,
-    .create_pipeline_layout = iree_hal_sync_device_create_pipeline_layout,
     .create_semaphore = iree_hal_sync_device_create_semaphore,
     .query_semaphore_compatibility =
         iree_hal_sync_device_query_semaphore_compatibility,