csullivan commented on code in PR #11878: URL: https://github.com/apache/tvm/pull/11878#discussion_r931581417
########## src/relay/transforms/annotate_texture_storage.cc: ########## @@ -0,0 +1,509 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * \file annotate_texture_storage.cc + * \brief Collection of target specific relay passes which + * storage scope related information. + * + * - CollectStorageInfo returns a mapping from relay expr + * to a list of output storage scopes for each output. + * These scopes are used during memory planning as well + * as downstream when doing codegen and in the graph runtime when doing runtime dataspace + * allocations. + * + * - AnnotateMemoryScope calls *target.CollectStorageInfo for all target been represented + * in the graph and rewrites graph modifying or inserting of VirtualDevice with required + * memory_scop collected from the CollectStorageInfo + */ + +#include <tvm/relay/attrs/nn.h> +#include <tvm/relay/expr.h> +#include <tvm/relay/expr_functor.h> +#include <tvm/relay/transform.h> +#include <tvm/tir/expr.h> + +#include <memory> +#include <unordered_map> + +#include "../transforms/device_aware_visitors.h" + +namespace tvm { +namespace relay { +namespace { + +/** + * @brief Analyzes the graph and returns mapping of expressions vs desired memory scope + */ +class StorageInfo : private transform::DeviceAwareExprVisitor { + public: + StorageInfo() : transform::DeviceAwareExprVisitor(Optional<IRModule>()) {} + + static Map<Expr, Array<String>> GetStorageMap(const Expr& expr) { + StorageInfo storage_info; + storage_info.VisitExpr(expr); + storage_info.LegalizeProducerStorage(); + Map<Expr, Array<String>> storage_map; + for (auto& kv : storage_info.storage_scope_) { + std::vector<String> storage_scopes; + std::copy(kv.second.begin(), kv.second.end(), std::back_inserter(storage_scopes)); + storage_map.Set(GetRef<Expr>(kv.first), Array<String>{storage_scopes}); + } + + // Filling the input arguments by "global" scope to handle PlanDevice algo which propagates + // virtual devices from outputs to inputs. At the same time outputs must be unconstrained + // to avoid useless device_copy + for (const auto& cs : storage_info.consumer_storage_scopes_) { + // we have record in consumers that mean that potentially consumer + // dealt with textures anyhow, it's safe to mark this expr as global scope + // even without verification of the consumer's outputs scope + if (storage_info.CanConsumeTextures(cs.second) && + storage_map.find(GetRef<Expr>(cs.first)) == storage_map.end()) { + storage_map.Set(GetRef<Expr>(cs.first), Array<String>{"global"}); + } + } + + // initial algo assumes mapping of outputs of the expr that is not enough, need to update + // VirtualDevice for function variables to get proper codegen. Adding vars to storage_map + for (const auto& a : storage_info.args_to_vars_) { + if (storage_map.count(a.first)) { + for (const auto& v : a.second) { + storage_map.Set(v, storage_map[a.first]); + } + } + } + return storage_map; + } + + private: + void Visit(const Expr& expr) { + // Pre-order traversal to enable upward propagation + // of consumer storage scopes to producers when desirable. + if (const auto* fn = expr.as<FunctionNode>()) { + this->VisitExpr(fn->body); + for (const auto& param : fn->params) { + this->VisitExpr(param); + } + } else { + this->VisitExpr(expr); + } + } + + void VisitExpr_(const VarNode* vn) final { ApplyConsumerScopeToInputs(vn); } + + void VisitExpr_(const ConstantNode* cn) final { ApplyConsumerScopeToInputs(cn); } + + void DeviceAwareVisitExpr_(const CallNode* call) final { + // Check the contents of this primitive function + if (const auto* fn = call->op.as<FunctionNode>()) { + if (fn->HasNonzeroAttr(attr::kPrimitive)) { + primitive_supports_texture_ = false; + Visit(call->op); + if (primitive_supports_texture_) { + if (call->checked_type().as<TensorTypeNode>()) { + std::string scope = "global.texture"; + if (const auto* ttype = call->checked_type().as<TensorTypeNode>()) { + if (ttype->shape.size() == 5) { + scope = Scope(ttype->shape, GetVirtualDevice(GetRef<Expr>(call))); + } + } + storage_scope_[call].push_back(scope); + } else { + const auto* tuple_type = call->type_as<TupleTypeNode>(); + ICHECK(tuple_type); + // TODO(csullivan): Add support for mixed output storage scope. + // In current adreno storage planner all outputs of a + // primitive function are assumed to be of the same storage + // type. This should be easy to extend in the future. + for (size_t i = 0; i < tuple_type->fields.size(); i++) { + storage_scope_[call].push_back("global.texture"); + } + } + for (size_t i = 0; i < fn->params.size(); i++) { + args_to_vars_[call->args[i]].push_back(fn->params[i]); + } + } + // Add consumer storage scope information for call arguments + for (auto& arg : call->args) { + if (storage_scope_.count(call)) { + ICHECK(!HasMixedStorageOutputs(call)) + << "Mixed output storage scopes are not currently supported"; + consumer_storage_scopes_[arg.operator->()].push_back("global.texture"); + } else { + consumer_storage_scopes_[arg.operator->()].push_back("global"); + } + } + } + } + + primitive_supports_texture_ = SupportsTextureStorage(call); + + for (auto& arg : call->args) { + Visit(arg); + } + // We have all callees filled into storage_scope_ if they support textures + // We need to verify if this call expects texture and if it does not, remove from + // storage_scope_ since initially storage_scope_ is filled only based on knowledge + // that function able to work with textures, but not necessary that this texture is + // expected by function callee + for (auto& arg : call->args) { + if (consumer_storage_scopes_.count(arg.operator->()) && + GetConsumerScope(consumer_storage_scopes_[arg.operator->()]) != "global.texture") { + storage_scope_.erase(arg.operator->()); + if (const auto* cn = arg.as<CallNode>()) { + if (const auto* fn = cn->op.as<FunctionNode>()) { + storage_scope_.erase(fn->body.operator->()); + } + } + } + } + } + + std::string Scope(Array<PrimExpr> shape, const VirtualDevice& vd) { + if (vd != VirtualDevice::FullyUnconstrained()) { + std::map<int, std::string> diffs; + int limit = + vd->target->GetAttr<Integer>("texture_spatial_limit").value_or(Integer(16384))->value; + int a0 = shape[0].as<IntImmNode>()->value; + int a1 = shape[1].as<IntImmNode>()->value; + int a2 = shape[2].as<IntImmNode>()->value; + int a3 = shape[3].as<IntImmNode>()->value; + + int d3l = a0 * a1 * a2; + int d3r = a3; + int diff3 = d3l > d3r ? d3l - d3r : d3r - d3l; + if (d3l < limit && d3r < limit) diffs[diff3] = ""; + + int d2l = a0 * a1; + int d2r = a2 * a3; + int diff2 = d2l > d2r ? d2l - d2r : d2r - d2l; + if (d2l < limit && d2r < limit) diffs[diff2] = "nhwc"; + + int d1l = a0; + int d1r = a1 * a2 * a3; + int diff1 = d1l > d1r ? d1l - d1r : d1r - d1l; + if (d1l < limit && d1r < limit) diffs[diff1] = "weight"; + if (!diffs.empty()) { + std::string scope = "global.texture"; + if (!diffs.begin()->second.empty()) { + scope += ("-" + diffs.begin()->second); + } + return scope; + } + } + return "global"; + } + + void ApplyConsumerScopeToInputs(const ExprNode* expr) { + std::string scope; + auto consumer_scopes_it = consumer_storage_scopes_.find(expr); + if (consumer_scopes_it != consumer_storage_scopes_.end()) { + std::string consumer_scope = GetConsumerScope(consumer_scopes_it->second); + ICHECK(!storage_scope_.count(expr)) + << "Already propagated consumer scopes to input: " << GetRef<Expr>(expr); + + bool expr_is_rgba_vectorizable = false; + if (const auto* ttype = expr->checked_type().as<TensorTypeNode>()) { + if (ttype->shape.size() == 5) { + scope = Scope(ttype->shape, GetVirtualDevice(GetRef<Expr>(expr))); + if (scope != "global") { + auto inner_dim = ttype->shape.back().as<IntImmNode>(); + if (inner_dim && inner_dim->value == 4) { + expr_is_rgba_vectorizable = true; + } + } + } + } + + // Only propagate texture scope from consumers to input expr if + // the input shape of the input expr is rgba vectorizable. + if (consumer_scope.find("global.texture") != std::string::npos) { + if (expr_is_rgba_vectorizable) { + storage_scope_[expr].push_back(scope); + } + } else { + storage_scope_[expr].push_back(consumer_scope); + } + } + } + + void LegalizeProducerStorage() { + for (auto& kv : consumer_storage_scopes_) { + const ExprNode* producer = kv.first; + std::string legal_scope = GetConsumerScope(kv.second); + if (storage_scope_.count(producer)) { + ICHECK(!HasMixedStorageOutputs(producer)) + << "Mixed output storage scopes are not currently supported"; + if (storage_scope_[producer][0].find(legal_scope) == std::string::npos) { + for (size_t i = 0; i < storage_scope_[producer].size(); i++) { + // Only support uniform storage scope across all outputs for now + storage_scope_[producer][i] = legal_scope; + } + } + } + } + } + + std::string GetConsumerScope(const std::vector<std::string>& consumer_scopes) const { + if (!consumer_scopes.size()) { + return "global"; + } + std::string texture_tag = "global.texture"; + for (auto& consumer_scope : consumer_scopes) { + if (consumer_scope.find(texture_tag) == std::string::npos) { + return "global"; + } + } + return texture_tag; + } + + bool CanConsumeTextures(const std::vector<std::string>& consumer_scopes) const { + std::string texture_tag = "global.texture"; + for (auto& consumer_scope : consumer_scopes) { + if (consumer_scope.find(texture_tag) == 0) { + return true; + } + } + return false; + } + + bool HasMixedStorageOutputs(const ExprNode* expr) { + if (storage_scope_.count(expr)) { + std::string ref_scope = storage_scope_[expr][0]; + for (std::string& scope : storage_scope_[expr]) { + if (scope != ref_scope) { + return true; + } + } + } + return false; + } + + bool SupportsTextureStorage(const CallNode* call) const { + bool supports_texture_storage = false; + if (auto attrs = call->attrs.as<Conv2DAttrs>()) { + if (attrs->data_layout == "NCHW4c" && attrs->kernel_layout == "OIHW4o") { + supports_texture_storage = true; + } else if (attrs->data_layout == "NHWC4c" && + (attrs->kernel_layout == "HWOI4o" || attrs->kernel_layout == "HWIO4o" || + attrs->kernel_layout == "OIHW4o")) { + supports_texture_storage = true; + } + } else if (auto attrs = call->attrs.as<Conv2DWinogradAttrs>()) { + if ((attrs->data_layout == "NCHW4c" || attrs->data_layout == "NHWC4c") && + (attrs->kernel_layout == "OIHW4o" || attrs->kernel_layout == "HWIO4o")) { + supports_texture_storage = true; + } + } else if (auto attrs = call->attrs.as<GlobalPool2DAttrs>()) { + if (attrs->layout == "NCHW4c") { + supports_texture_storage = true; + } + } else if (auto attrs = call->attrs.as<MaxPool2DAttrs>()) { + if (attrs->layout == "NCHW4c") { + supports_texture_storage = true; + } + } else if (auto attrs = call->attrs.as<AvgPool2DAttrs>()) { + if (attrs->layout == "NCHW4c") { + supports_texture_storage = true; + } + } + + return supports_texture_storage; + } + + /*! \brief Temporary state for marking whether a visited function + * primitive supports texture storage scope */ + bool primitive_supports_texture_ = false; + /*! \brief expr storage scope mapping for each output */ + std::unordered_map<const ExprNode*, std::vector<std::string>> storage_scope_; + /*! \brief output storage scopes used by consumers of expr key */ + std::unordered_map<const ExprNode*, std::vector<std::string>> consumer_storage_scopes_; + /*! \brief mapping of arguments to call to function variables*/ + std::unordered_map<Expr, std::vector<Var>, ObjectPtrHash, ObjectPtrEqual> args_to_vars_; +}; + +} // namespace + +/** + * @brief rewrite of virtual devices, memory_scope part for expressions defined + * by the StorageInfo analysis pass + * + * Currently this workflow supports analysis and rewriting of VirtualDevice for + * Constants and function Variables + */ +class VDRewriter : public transform::DeviceAwareExprMutator { Review Comment: VDRewriter is a bit too generic class name could be `RewriteVDStorageScopes` or similar ########## src/relay/transforms/annotate_texture_storage.cc: ########## @@ -0,0 +1,509 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * \file annotate_texture_storage.cc + * \brief Collection of target specific relay passes which + * storage scope related information. + * + * - CollectStorageInfo returns a mapping from relay expr + * to a list of output storage scopes for each output. + * These scopes are used during memory planning as well + * as downstream when doing codegen and in the graph runtime when doing runtime dataspace + * allocations. + * + * - AnnotateMemoryScope calls *target.CollectStorageInfo for all target been represented + * in the graph and rewrites graph modifying or inserting of VirtualDevice with required + * memory_scop collected from the CollectStorageInfo + */ + +#include <tvm/relay/attrs/nn.h> +#include <tvm/relay/expr.h> +#include <tvm/relay/expr_functor.h> +#include <tvm/relay/transform.h> +#include <tvm/tir/expr.h> + +#include <memory> +#include <unordered_map> + +#include "../transforms/device_aware_visitors.h" + +namespace tvm { +namespace relay { +namespace { + +/** + * @brief Analyzes the graph and returns mapping of expressions vs desired memory scope + */ +class StorageInfo : private transform::DeviceAwareExprVisitor { + public: + StorageInfo() : transform::DeviceAwareExprVisitor(Optional<IRModule>()) {} + + static Map<Expr, Array<String>> GetStorageMap(const Expr& expr) { + StorageInfo storage_info; + storage_info.VisitExpr(expr); + storage_info.LegalizeProducerStorage(); + Map<Expr, Array<String>> storage_map; + for (auto& kv : storage_info.storage_scope_) { + std::vector<String> storage_scopes; + std::copy(kv.second.begin(), kv.second.end(), std::back_inserter(storage_scopes)); + storage_map.Set(GetRef<Expr>(kv.first), Array<String>{storage_scopes}); + } + + // Filling the input arguments by "global" scope to handle PlanDevice algo which propagates + // virtual devices from outputs to inputs. At the same time outputs must be unconstrained + // to avoid useless device_copy + for (const auto& cs : storage_info.consumer_storage_scopes_) { + // we have record in consumers that mean that potentially consumer + // dealt with textures anyhow, it's safe to mark this expr as global scope + // even without verification of the consumer's outputs scope + if (storage_info.CanConsumeTextures(cs.second) && + storage_map.find(GetRef<Expr>(cs.first)) == storage_map.end()) { + storage_map.Set(GetRef<Expr>(cs.first), Array<String>{"global"}); + } + } + + // initial algo assumes mapping of outputs of the expr that is not enough, need to update + // VirtualDevice for function variables to get proper codegen. Adding vars to storage_map + for (const auto& a : storage_info.args_to_vars_) { + if (storage_map.count(a.first)) { + for (const auto& v : a.second) { + storage_map.Set(v, storage_map[a.first]); + } + } + } + return storage_map; + } + + private: + void Visit(const Expr& expr) { + // Pre-order traversal to enable upward propagation + // of consumer storage scopes to producers when desirable. + if (const auto* fn = expr.as<FunctionNode>()) { + this->VisitExpr(fn->body); + for (const auto& param : fn->params) { + this->VisitExpr(param); + } + } else { + this->VisitExpr(expr); + } + } + + void VisitExpr_(const VarNode* vn) final { ApplyConsumerScopeToInputs(vn); } + + void VisitExpr_(const ConstantNode* cn) final { ApplyConsumerScopeToInputs(cn); } + + void DeviceAwareVisitExpr_(const CallNode* call) final { + // Check the contents of this primitive function + if (const auto* fn = call->op.as<FunctionNode>()) { + if (fn->HasNonzeroAttr(attr::kPrimitive)) { + primitive_supports_texture_ = false; + Visit(call->op); + if (primitive_supports_texture_) { + if (call->checked_type().as<TensorTypeNode>()) { + std::string scope = "global.texture"; + if (const auto* ttype = call->checked_type().as<TensorTypeNode>()) { + if (ttype->shape.size() == 5) { + scope = Scope(ttype->shape, GetVirtualDevice(GetRef<Expr>(call))); + } + } + storage_scope_[call].push_back(scope); + } else { + const auto* tuple_type = call->type_as<TupleTypeNode>(); + ICHECK(tuple_type); + // TODO(csullivan): Add support for mixed output storage scope. + // In current adreno storage planner all outputs of a + // primitive function are assumed to be of the same storage + // type. This should be easy to extend in the future. + for (size_t i = 0; i < tuple_type->fields.size(); i++) { + storage_scope_[call].push_back("global.texture"); + } + } + for (size_t i = 0; i < fn->params.size(); i++) { + args_to_vars_[call->args[i]].push_back(fn->params[i]); + } + } + // Add consumer storage scope information for call arguments + for (auto& arg : call->args) { + if (storage_scope_.count(call)) { + ICHECK(!HasMixedStorageOutputs(call)) + << "Mixed output storage scopes are not currently supported"; + consumer_storage_scopes_[arg.operator->()].push_back("global.texture"); + } else { + consumer_storage_scopes_[arg.operator->()].push_back("global"); + } + } + } + } + + primitive_supports_texture_ = SupportsTextureStorage(call); + + for (auto& arg : call->args) { + Visit(arg); + } + // We have all callees filled into storage_scope_ if they support textures + // We need to verify if this call expects texture and if it does not, remove from + // storage_scope_ since initially storage_scope_ is filled only based on knowledge + // that function able to work with textures, but not necessary that this texture is + // expected by function callee + for (auto& arg : call->args) { + if (consumer_storage_scopes_.count(arg.operator->()) && + GetConsumerScope(consumer_storage_scopes_[arg.operator->()]) != "global.texture") { + storage_scope_.erase(arg.operator->()); + if (const auto* cn = arg.as<CallNode>()) { + if (const auto* fn = cn->op.as<FunctionNode>()) { + storage_scope_.erase(fn->body.operator->()); + } + } + } + } + } + + std::string Scope(Array<PrimExpr> shape, const VirtualDevice& vd) { Review Comment: This function needs some documentation explaining its intent to check various limits on texture and assign the correct texture layout to the storage scope based on the limit comparisons. ########## src/relay/transforms/annotate_texture_storage.cc: ########## @@ -0,0 +1,509 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * \file annotate_texture_storage.cc + * \brief Collection of target specific relay passes which + * storage scope related information. + * + * - CollectStorageInfo returns a mapping from relay expr + * to a list of output storage scopes for each output. + * These scopes are used during memory planning as well + * as downstream when doing codegen and in the graph runtime when doing runtime dataspace + * allocations. + * + * - AnnotateMemoryScope calls *target.CollectStorageInfo for all target been represented + * in the graph and rewrites graph modifying or inserting of VirtualDevice with required + * memory_scop collected from the CollectStorageInfo + */ + +#include <tvm/relay/attrs/nn.h> +#include <tvm/relay/expr.h> +#include <tvm/relay/expr_functor.h> +#include <tvm/relay/transform.h> +#include <tvm/tir/expr.h> + +#include <memory> +#include <unordered_map> + +#include "../transforms/device_aware_visitors.h" + +namespace tvm { +namespace relay { +namespace { + +/** + * @brief Analyzes the graph and returns mapping of expressions vs desired memory scope + */ +class StorageInfo : private transform::DeviceAwareExprVisitor { + public: + StorageInfo() : transform::DeviceAwareExprVisitor(Optional<IRModule>()) {} + + static Map<Expr, Array<String>> GetStorageMap(const Expr& expr) { + StorageInfo storage_info; + storage_info.VisitExpr(expr); + storage_info.LegalizeProducerStorage(); + Map<Expr, Array<String>> storage_map; + for (auto& kv : storage_info.storage_scope_) { + std::vector<String> storage_scopes; + std::copy(kv.second.begin(), kv.second.end(), std::back_inserter(storage_scopes)); + storage_map.Set(GetRef<Expr>(kv.first), Array<String>{storage_scopes}); + } + + // Filling the input arguments by "global" scope to handle PlanDevice algo which propagates + // virtual devices from outputs to inputs. At the same time outputs must be unconstrained + // to avoid useless device_copy + for (const auto& cs : storage_info.consumer_storage_scopes_) { + // we have record in consumers that mean that potentially consumer + // dealt with textures anyhow, it's safe to mark this expr as global scope + // even without verification of the consumer's outputs scope + if (storage_info.CanConsumeTextures(cs.second) && + storage_map.find(GetRef<Expr>(cs.first)) == storage_map.end()) { + storage_map.Set(GetRef<Expr>(cs.first), Array<String>{"global"}); + } + } + + // initial algo assumes mapping of outputs of the expr that is not enough, need to update + // VirtualDevice for function variables to get proper codegen. Adding vars to storage_map + for (const auto& a : storage_info.args_to_vars_) { + if (storage_map.count(a.first)) { + for (const auto& v : a.second) { + storage_map.Set(v, storage_map[a.first]); + } + } + } + return storage_map; + } + + private: + void Visit(const Expr& expr) { + // Pre-order traversal to enable upward propagation + // of consumer storage scopes to producers when desirable. + if (const auto* fn = expr.as<FunctionNode>()) { + this->VisitExpr(fn->body); + for (const auto& param : fn->params) { + this->VisitExpr(param); + } + } else { + this->VisitExpr(expr); + } + } + + void VisitExpr_(const VarNode* vn) final { ApplyConsumerScopeToInputs(vn); } + + void VisitExpr_(const ConstantNode* cn) final { ApplyConsumerScopeToInputs(cn); } + + void DeviceAwareVisitExpr_(const CallNode* call) final { + // Check the contents of this primitive function + if (const auto* fn = call->op.as<FunctionNode>()) { + if (fn->HasNonzeroAttr(attr::kPrimitive)) { + primitive_supports_texture_ = false; + Visit(call->op); + if (primitive_supports_texture_) { + if (call->checked_type().as<TensorTypeNode>()) { + std::string scope = "global.texture"; + if (const auto* ttype = call->checked_type().as<TensorTypeNode>()) { + if (ttype->shape.size() == 5) { + scope = Scope(ttype->shape, GetVirtualDevice(GetRef<Expr>(call))); + } + } + storage_scope_[call].push_back(scope); + } else { + const auto* tuple_type = call->type_as<TupleTypeNode>(); + ICHECK(tuple_type); + // TODO(csullivan): Add support for mixed output storage scope. + // In current adreno storage planner all outputs of a + // primitive function are assumed to be of the same storage + // type. This should be easy to extend in the future. Review Comment: Can we relax this requirement now? I don't recall the outcome of our discussion previously on this. I thought I recall you having mentioned some change on the topic of heterogeneous scopes in a tuple. ########## src/relay/transforms/annotate_texture_storage.cc: ########## @@ -0,0 +1,509 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * \file annotate_texture_storage.cc + * \brief Collection of target specific relay passes which + * storage scope related information. + * + * - CollectStorageInfo returns a mapping from relay expr + * to a list of output storage scopes for each output. + * These scopes are used during memory planning as well + * as downstream when doing codegen and in the graph runtime when doing runtime dataspace + * allocations. + * + * - AnnotateMemoryScope calls *target.CollectStorageInfo for all target been represented + * in the graph and rewrites graph modifying or inserting of VirtualDevice with required + * memory_scop collected from the CollectStorageInfo Review Comment: ```suggestion * memory_scope collected from the CollectStorageInfo ``` ########## tests/python/relay/test_conv2d_nchw_texture.py: ########## @@ -435,3 +435,558 @@ def test_conv2d_vgg16_winograd_4d(): graph = build_run_compare(mod, params1, {"data": input_shape}, dtype, target) matches = re.findall("winograd", graph) assert len(matches) > 0 + + [email protected]_opencl +def test_residual_block(): + target = "opencl --device=adreno" + dtype = "float16" + + input_shape = (1, 32, 40, 40) + filter_shape1 = (32, 32, 2, 2) + filter_shape2 = (32, 32, 1, 1) + filter_shape3 = (32, 32, 2, 2) + bias_shape1 = (1, 32, 1, 1) + A = relay.var("data", shape=input_shape, dtype=dtype) + W1 = relay.var("weight1", shape=filter_shape1, dtype=dtype) + B1 = relay.var("bias1", shape=bias_shape1, dtype=dtype) + W2 = relay.var("weight2", shape=filter_shape2, dtype=dtype) + W3 = relay.var("weight3", shape=filter_shape3, dtype=dtype) + + conv1 = relay.nn.conv2d( + A, + W1, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + D = relay.op.add(conv1, B1) + D = relay.op.nn.relu(D) + + conv2 = relay.nn.conv2d( + D, + W2, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(1, 1), + ) + D = relay.op.add(conv2, D) + D = D * relay.const(0.15, "float16") + D = relay.op.nn.relu(D) + + conv3 = relay.nn.conv2d( + D, + W3, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + D = relay.op.nn.relu(conv3) + + mod = relay.Function([A, W1, B1, W2, W3], D) + np.random.seed(0) + initializer = relay.testing.init.Xavier() + filter_data1 = np.zeros(filter_shape1).astype(dtype) + bias_data1 = np.zeros(bias_shape1).astype(dtype) + initializer("weight", filter_data1) + initializer("bias", bias_data1) + filter_data2 = np.zeros(filter_shape2).astype(dtype) + initializer("weight", filter_data2) + filter_data3 = np.zeros(filter_shape3).astype(dtype) + initializer("weight", filter_data3) + params1 = { + "weight1": tvm.nd.array(filter_data1), + "bias1": tvm.nd.array(bias_data1), + "weight2": tvm.nd.array(filter_data2), + "weight3": tvm.nd.array(filter_data3), + } + + static_memory_scope = [ + "", + "global", + "global.texture-weight", + "global.texture-weight", + "global.texture", + "global.texture-weight", + "global", + "global.texture", + "global.texture-weight", + "", + "", + ] + + build_run_compare(mod, params1, {"data": input_shape}, dtype, target, static_memory_scope) + + [email protected]_opencl +def test_plan_device_issue1(): + target = "opencl --device=adreno" + dtype = "float16" + + input_shape = (1, 32, 40, 40) + filter_shape1 = (32, 32, 2, 2) + filter_shape2 = (32, 32, 1, 1) + A = relay.var("data", shape=input_shape, dtype=dtype) + W1 = relay.var("weight1", shape=filter_shape1, dtype=dtype) + W2 = relay.var("weight2", shape=filter_shape2, dtype=dtype) + + conv1 = relay.nn.conv2d( + A, + W1, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + conv2 = relay.nn.conv2d( + conv1, + W2, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(1, 1), + ) + + mod = relay.Function([A, W1, W2], conv2) + np.random.seed(0) + initializer = relay.testing.init.Xavier() + filter_data1 = np.zeros(filter_shape1).astype(dtype) + initializer("weight", filter_data1) + filter_data2 = np.zeros(filter_shape2).astype(dtype) + initializer("weight", filter_data2) + params1 = { + "weight1": tvm.nd.array(filter_data1), + "weight2": tvm.nd.array(filter_data2), + } + + static_memory_scope = [ + "", + "global", + "global.texture-weight", + "global.texture", + "global.texture-weight", + "", + "", + ] + + static_memory_scope = [] + + build_run_compare(mod, params1, {"data": input_shape}, dtype, target, static_memory_scope) + + [email protected]_opencl +def test_branch_textures(): + target = "opencl --device=adreno" + dtype = "float16" + + input_shape = (1, 32, 40, 40) + filter_shape1 = (96, 32, 2, 2) + filter_shape2 = (32, 96, 2, 2) + filter_shape3 = (5, 96, 2, 2) + bias_shape1 = (1, 96, 1, 1) + bias_shape2 = (1, 32, 1, 1) + A = relay.var("data", shape=input_shape, dtype=dtype) + W1 = relay.var("weight1", shape=filter_shape1, dtype=dtype) + B1 = relay.var("bias1", shape=bias_shape1, dtype=dtype) + W2 = relay.var("weight2", shape=filter_shape2, dtype=dtype) + W3 = relay.var("weight3", shape=filter_shape3, dtype=dtype) + B2 = relay.var("bias2", shape=bias_shape2, dtype=dtype) + + # C = relay.nn.relu(A) + conv1 = relay.nn.conv2d( + A, + W1, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=96, + kernel_size=(2, 2), + ) + D = relay.op.add(conv1, B1) + D = relay.op.nn.relu(D) + + conv2 = relay.nn.conv2d( + D, + W2, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + conv2 = relay.op.add(conv2, B2) + conv2 = relay.op.nn.relu(conv2) + + conv3 = relay.nn.conv2d( + D, + W3, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=5, + kernel_size=(2, 2), + ) + + t = relay.Tuple([conv2, conv3]) + c = relay.op.concatenate(t, axis=1) + + mod = relay.Function([A, W1, B1, W2, B2, W3], c) + np.random.seed(0) + initializer = relay.testing.init.Xavier() + filter_data1 = np.zeros(filter_shape1).astype(dtype) + bias_data1 = np.zeros(bias_shape1).astype(dtype) + initializer("weight", filter_data1) + initializer("bias", bias_data1) + filter_data2 = np.zeros(filter_shape2).astype(dtype) + bias_data2 = np.zeros(bias_shape2).astype(dtype) + initializer("weight", filter_data2) + initializer("bias", bias_data2) + filter_data3 = np.zeros(filter_shape3).astype(dtype) + initializer("weight", filter_data3) + params1 = { + "weight1": tvm.nd.array(filter_data1), + "bias1": tvm.nd.array(bias_data1), + "weight2": tvm.nd.array(filter_data2), + "bias2": tvm.nd.array(bias_data2), + "weight3": tvm.nd.array(filter_data3), + } + + static_memory_scope = [ + "", + "global", + "global.texture-weight", + "global.texture-weight", + "global", + "global.texture-weight", + "global.texture-weight", + "", + "", + "", + "", + "", + ] + + static_memory_scope = [] + + build_run_compare(mod, params1, {"data": input_shape}, dtype, target, static_memory_scope) + + [email protected]_opencl +def test_branch1_texture_params(): + target = "opencl --device=adreno" + dtype = "float16" + + input_shape = (1, 32, 40, 40) + filter_shape0 = (32, 32, 1, 1) + filter_shape1 = (32, 32, 2, 2) + filter_shape2 = (32, 32, 1, 1) + filter_shape3 = (32, 32, 2, 2) + bias_shape1 = (1, 32, 1, 1) + # bias_shape2 = (1, 32, 1, 1) + A = relay.var("data", shape=input_shape, dtype=dtype) + W0 = relay.var("weight0", shape=filter_shape0, dtype=dtype) + W1 = relay.var("weight1", shape=filter_shape1, dtype=dtype) + B1 = relay.var("bias1", shape=bias_shape1, dtype=dtype) + W2 = relay.var("weight2", shape=filter_shape2, dtype=dtype) + W3 = relay.var("weight3", shape=filter_shape3, dtype=dtype) + + conv0 = relay.nn.conv2d( + A, + W0, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(1, 1), + ) + + pool = relay.nn.avg_pool2d(conv0, pool_size=(2, 2), strides=(2, 2)) + conv1 = relay.nn.conv2d( + pool, + W1, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 1, 1], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + conv1 = relay.op.add(conv1, B1) + conv1 = relay.op.nn.relu(conv1) + + conv2 = relay.nn.conv2d( + pool, + W2, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(1, 1), + ) + + conv3 = relay.nn.conv2d( + pool, + W3, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 1, 1, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + conv3 = relay.op.nn.relu(conv3) + res = relay.op.add(conv1, conv2) + res = relay.op.add(res, conv3) + + mod = relay.Function([A, W0, W1, B1, W2, W3], res) + np.random.seed(0) + initializer = relay.testing.init.Xavier() + filter_data0 = np.zeros(filter_shape0).astype(dtype) + filter_data1 = np.zeros(filter_shape1).astype(dtype) + bias_data1 = np.zeros(bias_shape1).astype(dtype) + initializer("weight", filter_data1) + initializer("bias", bias_data1) + filter_data2 = np.zeros(filter_shape2).astype(dtype) + initializer("weight", filter_data2) + filter_data3 = np.zeros(filter_shape3).astype(dtype) + initializer("weight", filter_data3) + params1 = { + "weight0": tvm.nd.array(filter_data0), + "weight1": tvm.nd.array(filter_data1), + "bias1": tvm.nd.array(bias_data1), + "weight2": tvm.nd.array(filter_data2), + "weight3": tvm.nd.array(filter_data3), + } + + static_memory_scope = [ + "", + "global", + "global.texture-weight", + "global.texture", + "global.texture", + "global.texture-weight", + "global.texture-weight", + "global.texture-weight", + "global.texture", + "global.texture-weight", + "global.texture", + "", + "", + ] + + build_run_compare(mod, params1, {"data": input_shape}, dtype, target, static_memory_scope) + + +# conv2d <- to get textures Review Comment: This diagram is helpful. Can you format it in a docstring. The above comment about making the static_memory_scopes would greatly help in tying back that list to this diagram. ########## tests/python/relay/test_conv2d_nchw_texture.py: ########## @@ -435,3 +435,558 @@ def test_conv2d_vgg16_winograd_4d(): graph = build_run_compare(mod, params1, {"data": input_shape}, dtype, target) matches = re.findall("winograd", graph) assert len(matches) > 0 + + [email protected]_opencl +def test_residual_block(): Review Comment: Can you give a docstring description here? Specifically below there are some ops in global.texture and others in global. It would be nice to know why that's expected. ########## tests/python/relay/test_conv2d_nchw_texture.py: ########## @@ -435,3 +435,558 @@ def test_conv2d_vgg16_winograd_4d(): graph = build_run_compare(mod, params1, {"data": input_shape}, dtype, target) matches = re.findall("winograd", graph) assert len(matches) > 0 + + [email protected]_opencl +def test_residual_block(): + target = "opencl --device=adreno" + dtype = "float16" + + input_shape = (1, 32, 40, 40) + filter_shape1 = (32, 32, 2, 2) + filter_shape2 = (32, 32, 1, 1) + filter_shape3 = (32, 32, 2, 2) + bias_shape1 = (1, 32, 1, 1) + A = relay.var("data", shape=input_shape, dtype=dtype) + W1 = relay.var("weight1", shape=filter_shape1, dtype=dtype) + B1 = relay.var("bias1", shape=bias_shape1, dtype=dtype) + W2 = relay.var("weight2", shape=filter_shape2, dtype=dtype) + W3 = relay.var("weight3", shape=filter_shape3, dtype=dtype) + + conv1 = relay.nn.conv2d( + A, + W1, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + D = relay.op.add(conv1, B1) + D = relay.op.nn.relu(D) + + conv2 = relay.nn.conv2d( + D, + W2, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(1, 1), + ) + D = relay.op.add(conv2, D) + D = D * relay.const(0.15, "float16") + D = relay.op.nn.relu(D) + + conv3 = relay.nn.conv2d( + D, + W3, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + D = relay.op.nn.relu(conv3) + + mod = relay.Function([A, W1, B1, W2, W3], D) + np.random.seed(0) + initializer = relay.testing.init.Xavier() + filter_data1 = np.zeros(filter_shape1).astype(dtype) + bias_data1 = np.zeros(bias_shape1).astype(dtype) + initializer("weight", filter_data1) + initializer("bias", bias_data1) + filter_data2 = np.zeros(filter_shape2).astype(dtype) + initializer("weight", filter_data2) + filter_data3 = np.zeros(filter_shape3).astype(dtype) + initializer("weight", filter_data3) + params1 = { + "weight1": tvm.nd.array(filter_data1), + "bias1": tvm.nd.array(bias_data1), + "weight2": tvm.nd.array(filter_data2), + "weight3": tvm.nd.array(filter_data3), + } + + static_memory_scope = [ + "", + "global", + "global.texture-weight", + "global.texture-weight", + "global.texture", + "global.texture-weight", + "global", + "global.texture", + "global.texture-weight", + "", Review Comment: It's great to have unit tests explicitly checking the scopes! That said, this list is a bit confusing. For example, this list has `global.texture-weight` scope as the final entry, which is a scope used for a weight, but that certainly is not the final expr in topo order. Even nicer would be a map from expr to scope, or since its checking against the json maybe function/kernel name for the key? This would make the test more intuitive. ########## tests/python/relay/test_conv2d_nchw_texture.py: ########## @@ -435,3 +435,558 @@ def test_conv2d_vgg16_winograd_4d(): graph = build_run_compare(mod, params1, {"data": input_shape}, dtype, target) matches = re.findall("winograd", graph) assert len(matches) > 0 + + [email protected]_opencl +def test_residual_block(): + target = "opencl --device=adreno" + dtype = "float16" + + input_shape = (1, 32, 40, 40) + filter_shape1 = (32, 32, 2, 2) + filter_shape2 = (32, 32, 1, 1) + filter_shape3 = (32, 32, 2, 2) + bias_shape1 = (1, 32, 1, 1) + A = relay.var("data", shape=input_shape, dtype=dtype) + W1 = relay.var("weight1", shape=filter_shape1, dtype=dtype) + B1 = relay.var("bias1", shape=bias_shape1, dtype=dtype) + W2 = relay.var("weight2", shape=filter_shape2, dtype=dtype) + W3 = relay.var("weight3", shape=filter_shape3, dtype=dtype) + + conv1 = relay.nn.conv2d( + A, + W1, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + D = relay.op.add(conv1, B1) + D = relay.op.nn.relu(D) + + conv2 = relay.nn.conv2d( + D, + W2, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[1, 1], + out_dtype=dtype, + channels=32, + kernel_size=(1, 1), + ) + D = relay.op.add(conv2, D) + D = D * relay.const(0.15, "float16") + D = relay.op.nn.relu(D) + + conv3 = relay.nn.conv2d( + D, + W3, + data_layout="NCHW", + kernel_layout="OIHW", + padding=[0, 0, 0, 0], + strides=[2, 2], + out_dtype=dtype, + channels=32, + kernel_size=(2, 2), + ) + D = relay.op.nn.relu(conv3) + + mod = relay.Function([A, W1, B1, W2, W3], D) + np.random.seed(0) + initializer = relay.testing.init.Xavier() + filter_data1 = np.zeros(filter_shape1).astype(dtype) + bias_data1 = np.zeros(bias_shape1).astype(dtype) + initializer("weight", filter_data1) + initializer("bias", bias_data1) + filter_data2 = np.zeros(filter_shape2).astype(dtype) + initializer("weight", filter_data2) + filter_data3 = np.zeros(filter_shape3).astype(dtype) + initializer("weight", filter_data3) + params1 = { + "weight1": tvm.nd.array(filter_data1), + "bias1": tvm.nd.array(bias_data1), + "weight2": tvm.nd.array(filter_data2), + "weight3": tvm.nd.array(filter_data3), + } + + static_memory_scope = [ + "", + "global", + "global.texture-weight", + "global.texture-weight", + "global.texture", + "global.texture-weight", + "global", + "global.texture", + "global.texture-weight", + "", + "", + ] + + build_run_compare(mod, params1, {"data": input_shape}, dtype, target, static_memory_scope) + + [email protected]_opencl +def test_plan_device_issue1(): Review Comment: Need a better name and a docstring. Same for the below tests, just a simple description can help. ########## src/relay/transforms/annotate_texture_storage.cc: ########## @@ -0,0 +1,509 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +/*! + * \file annotate_texture_storage.cc + * \brief Collection of target specific relay passes which + * storage scope related information. + * + * - CollectStorageInfo returns a mapping from relay expr + * to a list of output storage scopes for each output. + * These scopes are used during memory planning as well + * as downstream when doing codegen and in the graph runtime when doing runtime dataspace + * allocations. + * + * - AnnotateMemoryScope calls *target.CollectStorageInfo for all target been represented + * in the graph and rewrites graph modifying or inserting of VirtualDevice with required + * memory_scop collected from the CollectStorageInfo + */ + +#include <tvm/relay/attrs/nn.h> +#include <tvm/relay/expr.h> +#include <tvm/relay/expr_functor.h> +#include <tvm/relay/transform.h> +#include <tvm/tir/expr.h> + +#include <memory> +#include <unordered_map> + +#include "../transforms/device_aware_visitors.h" + +namespace tvm { +namespace relay { +namespace { + +/** + * @brief Analyzes the graph and returns mapping of expressions vs desired memory scope + */ +class StorageInfo : private transform::DeviceAwareExprVisitor { + public: + StorageInfo() : transform::DeviceAwareExprVisitor(Optional<IRModule>()) {} + + static Map<Expr, Array<String>> GetStorageMap(const Expr& expr) { + StorageInfo storage_info; + storage_info.VisitExpr(expr); + storage_info.LegalizeProducerStorage(); + Map<Expr, Array<String>> storage_map; + for (auto& kv : storage_info.storage_scope_) { + std::vector<String> storage_scopes; + std::copy(kv.second.begin(), kv.second.end(), std::back_inserter(storage_scopes)); + storage_map.Set(GetRef<Expr>(kv.first), Array<String>{storage_scopes}); + } + + // Filling the input arguments by "global" scope to handle PlanDevice algo which propagates + // virtual devices from outputs to inputs. At the same time outputs must be unconstrained + // to avoid useless device_copy + for (const auto& cs : storage_info.consumer_storage_scopes_) { + // we have record in consumers that mean that potentially consumer + // dealt with textures anyhow, it's safe to mark this expr as global scope + // even without verification of the consumer's outputs scope + if (storage_info.CanConsumeTextures(cs.second) && + storage_map.find(GetRef<Expr>(cs.first)) == storage_map.end()) { + storage_map.Set(GetRef<Expr>(cs.first), Array<String>{"global"}); + } + } + + // initial algo assumes mapping of outputs of the expr that is not enough, need to update + // VirtualDevice for function variables to get proper codegen. Adding vars to storage_map + for (const auto& a : storage_info.args_to_vars_) { + if (storage_map.count(a.first)) { + for (const auto& v : a.second) { + storage_map.Set(v, storage_map[a.first]); + } + } + } + return storage_map; + } + + private: + void Visit(const Expr& expr) { + // Pre-order traversal to enable upward propagation + // of consumer storage scopes to producers when desirable. + if (const auto* fn = expr.as<FunctionNode>()) { + this->VisitExpr(fn->body); + for (const auto& param : fn->params) { + this->VisitExpr(param); + } + } else { + this->VisitExpr(expr); + } + } + + void VisitExpr_(const VarNode* vn) final { ApplyConsumerScopeToInputs(vn); } + + void VisitExpr_(const ConstantNode* cn) final { ApplyConsumerScopeToInputs(cn); } + + void DeviceAwareVisitExpr_(const CallNode* call) final { + // Check the contents of this primitive function + if (const auto* fn = call->op.as<FunctionNode>()) { + if (fn->HasNonzeroAttr(attr::kPrimitive)) { + primitive_supports_texture_ = false; + Visit(call->op); + if (primitive_supports_texture_) { + if (call->checked_type().as<TensorTypeNode>()) { + std::string scope = "global.texture"; + if (const auto* ttype = call->checked_type().as<TensorTypeNode>()) { + if (ttype->shape.size() == 5) { + scope = Scope(ttype->shape, GetVirtualDevice(GetRef<Expr>(call))); + } + } + storage_scope_[call].push_back(scope); + } else { + const auto* tuple_type = call->type_as<TupleTypeNode>(); + ICHECK(tuple_type); + // TODO(csullivan): Add support for mixed output storage scope. + // In current adreno storage planner all outputs of a + // primitive function are assumed to be of the same storage + // type. This should be easy to extend in the future. + for (size_t i = 0; i < tuple_type->fields.size(); i++) { + storage_scope_[call].push_back("global.texture"); + } + } + for (size_t i = 0; i < fn->params.size(); i++) { + args_to_vars_[call->args[i]].push_back(fn->params[i]); + } + } + // Add consumer storage scope information for call arguments + for (auto& arg : call->args) { + if (storage_scope_.count(call)) { + ICHECK(!HasMixedStorageOutputs(call)) + << "Mixed output storage scopes are not currently supported"; + consumer_storage_scopes_[arg.operator->()].push_back("global.texture"); + } else { + consumer_storage_scopes_[arg.operator->()].push_back("global"); + } + } + } + } + + primitive_supports_texture_ = SupportsTextureStorage(call); + + for (auto& arg : call->args) { + Visit(arg); + } + // We have all callees filled into storage_scope_ if they support textures + // We need to verify if this call expects texture and if it does not, remove from + // storage_scope_ since initially storage_scope_ is filled only based on knowledge + // that function able to work with textures, but not necessary that this texture is + // expected by function callee + for (auto& arg : call->args) { + if (consumer_storage_scopes_.count(arg.operator->()) && + GetConsumerScope(consumer_storage_scopes_[arg.operator->()]) != "global.texture") { + storage_scope_.erase(arg.operator->()); + if (const auto* cn = arg.as<CallNode>()) { + if (const auto* fn = cn->op.as<FunctionNode>()) { + storage_scope_.erase(fn->body.operator->()); + } + } + } + } + } + + std::string Scope(Array<PrimExpr> shape, const VirtualDevice& vd) { + if (vd != VirtualDevice::FullyUnconstrained()) { + std::map<int, std::string> diffs; + int limit = + vd->target->GetAttr<Integer>("texture_spatial_limit").value_or(Integer(16384))->value; + int a0 = shape[0].as<IntImmNode>()->value; + int a1 = shape[1].as<IntImmNode>()->value; + int a2 = shape[2].as<IntImmNode>()->value; + int a3 = shape[3].as<IntImmNode>()->value; + + int d3l = a0 * a1 * a2; + int d3r = a3; + int diff3 = d3l > d3r ? d3l - d3r : d3r - d3l; + if (d3l < limit && d3r < limit) diffs[diff3] = ""; + + int d2l = a0 * a1; + int d2r = a2 * a3; + int diff2 = d2l > d2r ? d2l - d2r : d2r - d2l; + if (d2l < limit && d2r < limit) diffs[diff2] = "nhwc"; + + int d1l = a0; + int d1r = a1 * a2 * a3; + int diff1 = d1l > d1r ? d1l - d1r : d1r - d1l; + if (d1l < limit && d1r < limit) diffs[diff1] = "weight"; + if (!diffs.empty()) { + std::string scope = "global.texture"; + if (!diffs.begin()->second.empty()) { + scope += ("-" + diffs.begin()->second); + } + return scope; + } + } + return "global"; + } + + void ApplyConsumerScopeToInputs(const ExprNode* expr) { + std::string scope; + auto consumer_scopes_it = consumer_storage_scopes_.find(expr); + if (consumer_scopes_it != consumer_storage_scopes_.end()) { + std::string consumer_scope = GetConsumerScope(consumer_scopes_it->second); + ICHECK(!storage_scope_.count(expr)) + << "Already propagated consumer scopes to input: " << GetRef<Expr>(expr); + + bool expr_is_rgba_vectorizable = false; + if (const auto* ttype = expr->checked_type().as<TensorTypeNode>()) { + if (ttype->shape.size() == 5) { Review Comment: This is essentially checking for packing of C with layouts like `NHWCxc and NCHWxc` but `==5` is pretty mystical to the reader without that context. Can we do this in a way that is more clear? If not perhaps pull that check into Scope function and add a comment. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
