vinx13 commented on a change in pull request #9727: URL: https://github.com/apache/tvm/pull/9727#discussion_r768136729
########## File path: src/tir/ir/index_map.cc ########## @@ -0,0 +1,153 @@ +/* + * 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 index_map.cc + */ + +#include "tvm/tir/index_map.h" + +#include <tvm/arith/analyzer.h> +#include <tvm/arith/iter_affine_map.h> +#include <tvm/tir/op.h> + +#include <sstream> + +namespace tvm { +namespace tir { + +IndexMap::IndexMap(Array<Var> initial_indices, Array<PrimExpr> final_indices) { + auto n = make_object<IndexMapNode>(); + n->initial_indices = std::move(initial_indices); + n->final_indices = std::move(final_indices); + data_ = std::move(n); +} + +IndexMap IndexMap::Inverse(Array<Range> initial_ranges) const { + // Dummy variables to represent the inverse's inputs. + Array<Var> output_vars; + for (size_t i = 0; i < (*this)->final_indices.size(); i++) { + PrimExpr index = (*this)->final_indices[i]; + // TODO(Lunderberg): Better names for these variables. A variable + // that is passed through unmodified (`index` is an element of + // `initial_indices`) should use that input index's name. A pair + // of output indices variables split from a single input index + // should be named (X.outer,X.inner). + std::stringstream ss; + ss << "axis" << i; + Var var_index(ss.str(), index.dtype()); + output_vars.push_back(var_index); + } + + // Dummy ranges for the extent of each input. + Map<Var, Range> input_iters; + ICHECK_EQ((*this)->initial_indices.size(), initial_ranges.size()); + for (size_t i = 0; i < initial_ranges.size(); i++) { + input_iters.Set((*this)->initial_indices[i], initial_ranges[i]); + } + + // Unpack the output indices into linear combinations of the initial + // indices. + arith::Analyzer analyzer; + auto iter_map = DetectIterMap((*this)->final_indices, input_iters, 1, true, &analyzer); + CHECK(iter_map.size()) << "Index transformation was not bijective."; + + // Determine expressions for the input variables, in terms of the + // output variables. + Map<Var, PrimExpr> inverse_exprs_map = + InverseAffineIterMap(iter_map, Array<PrimExpr>(output_vars.begin(), output_vars.end())); + + // Unpack the map to an array, maintaining the same parameter order. + Array<PrimExpr> inverse_exprs; + for (const auto& index : (*this)->initial_indices) { + inverse_exprs.push_back(inverse_exprs_map.at(index)); + } + + return IndexMap(output_vars, inverse_exprs); +} + +Array<PrimExpr> IndexMapNode::MapIndices(const Array<PrimExpr>& indices) const { + ICHECK_EQ(indices.size(), initial_indices.size()); + + arith::Analyzer analyzer; + + for (size_t i = 0; i < initial_indices.size(); i++) { + analyzer.Bind(initial_indices[i], indices[i]); + } + + Array<PrimExpr> output; + for (const auto& output_dim : final_indices) { + output.push_back(analyzer.Simplify(output_dim)); + } + + return output; +} + +Array<Range> IndexMapNode::MapRanges(const Array<Range>& ranges) const { + ICHECK_EQ(ranges.size(), initial_indices.size()); + + Map<Var, Range> input_iters; + for (size_t i = 0; i < initial_indices.size(); i++) { + input_iters.Set(initial_indices[i], ranges[i]); + } + + arith::Analyzer analyzer; + auto iter_sums = DetectIterMap(final_indices, input_iters, 1, true, &analyzer); + + Array<Range> output; + for (const auto& iter_sum : iter_sums) { + PrimExpr min = iter_sum->base; + PrimExpr extent = 0; + for (const auto& term : iter_sum->args) { + extent += term->extent * term->scale; + } + output.push_back(Range::FromMinExtent(min, extent)); + } + + return output; +} + +Array<PrimExpr> IndexMapNode::MapShape(const Array<PrimExpr>& shape) const { + ICHECK_EQ(shape.size(), initial_indices.size()); + + Array<Range> ranges; + for (auto& dim : shape) { + ranges.push_back(Range(0, dim)); + } + Array<Range> mapped = MapRanges(std::move(ranges)); + + Array<PrimExpr> output; + for (auto& range : mapped) { + ICHECK(is_zero(range->min)); + output.push_back(range->extent); + } + Review comment: shall we also check the size (product of each dimension) is equal before and after the mapping (this is like a simplified check of surjection, or do we allow non-surjective mapping?) ########## File path: src/tir/transforms/ir_utils.cc ########## @@ -111,26 +111,56 @@ class IRConvertSSA final : public StmtExprMutator { return StmtExprMutator::VisitExpr_(op); } } + PrimExpr VisitExpr_(const LoadNode* op) final { - PrimExpr expr = StmtExprMutator::VisitExpr_(op); - op = expr.as<LoadNode>(); - const VarNode* v = op->buffer_var.get(); - if (scope_.count(v) && !scope_[v].empty()) { - return Load(op->dtype, scope_[v].back(), op->index, op->predicate); - } else { - return expr; - } + LOG(FATAL) << "Unexpected use of deprecated LoadNode. Please use BufferLoadNode instead."; + return PrimExpr(); } + Stmt VisitStmt_(const StoreNode* op) final { - Stmt stmt = StmtExprMutator::VisitStmt_(op); - op = stmt.as<StoreNode>(); - const VarNode* v = op->buffer_var.get(); - if (scope_.count(v) && !scope_[v].empty()) { - return Store(scope_[v].back(), op->value, op->index, op->predicate); - } else { - return stmt; + LOG(FATAL) << "Unexpected use of deprecated StoreNode. Please use BufferStoreNode instead."; + return Stmt(); + } + + PrimExpr VisitExpr_(const BufferLoadNode* op) final { + auto node = Downcast<BufferLoad>(StmtExprMutator::VisitExpr_(op)); + return VisitBufferAccess(node); + } + + Stmt VisitStmt_(const BufferStoreNode* op) final { + auto node = Downcast<BufferStore>(StmtExprMutator::VisitStmt_(op)); + return VisitBufferAccess(node); Review comment: ```suggestion return VisitBufferAccess(std::move(node)); ``` ########## File path: src/tir/transforms/ir_utils.cc ########## @@ -111,26 +111,56 @@ class IRConvertSSA final : public StmtExprMutator { return StmtExprMutator::VisitExpr_(op); } } + PrimExpr VisitExpr_(const LoadNode* op) final { - PrimExpr expr = StmtExprMutator::VisitExpr_(op); - op = expr.as<LoadNode>(); - const VarNode* v = op->buffer_var.get(); - if (scope_.count(v) && !scope_[v].empty()) { - return Load(op->dtype, scope_[v].back(), op->index, op->predicate); - } else { - return expr; - } + LOG(FATAL) << "Unexpected use of deprecated LoadNode. Please use BufferLoadNode instead."; + return PrimExpr(); } + Stmt VisitStmt_(const StoreNode* op) final { - Stmt stmt = StmtExprMutator::VisitStmt_(op); - op = stmt.as<StoreNode>(); - const VarNode* v = op->buffer_var.get(); - if (scope_.count(v) && !scope_[v].empty()) { - return Store(scope_[v].back(), op->value, op->index, op->predicate); - } else { - return stmt; + LOG(FATAL) << "Unexpected use of deprecated StoreNode. Please use BufferStoreNode instead."; + return Stmt(); + } + + PrimExpr VisitExpr_(const BufferLoadNode* op) final { + auto node = Downcast<BufferLoad>(StmtExprMutator::VisitExpr_(op)); + return VisitBufferAccess(node); Review comment: ```suggestion return VisitBufferAccess(std::move(node)); ``` need `std::move` to enable `CopyOnWrite` ########## File path: src/tir/ir/index_map.cc ########## @@ -0,0 +1,153 @@ +/* + * 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 index_map.cc + */ + +#include "tvm/tir/index_map.h" + +#include <tvm/arith/analyzer.h> +#include <tvm/arith/iter_affine_map.h> +#include <tvm/tir/op.h> + +#include <sstream> + +namespace tvm { +namespace tir { + +IndexMap::IndexMap(Array<Var> initial_indices, Array<PrimExpr> final_indices) { + auto n = make_object<IndexMapNode>(); + n->initial_indices = std::move(initial_indices); + n->final_indices = std::move(final_indices); + data_ = std::move(n); +} + +IndexMap IndexMap::Inverse(Array<Range> initial_ranges) const { + // Dummy variables to represent the inverse's inputs. + Array<Var> output_vars; + for (size_t i = 0; i < (*this)->final_indices.size(); i++) { + PrimExpr index = (*this)->final_indices[i]; + // TODO(Lunderberg): Better names for these variables. A variable + // that is passed through unmodified (`index` is an element of + // `initial_indices`) should use that input index's name. A pair + // of output indices variables split from a single input index + // should be named (X.outer,X.inner). + std::stringstream ss; + ss << "axis" << i; + Var var_index(ss.str(), index.dtype()); + output_vars.push_back(var_index); + } + + // Dummy ranges for the extent of each input. + Map<Var, Range> input_iters; + ICHECK_EQ((*this)->initial_indices.size(), initial_ranges.size()); + for (size_t i = 0; i < initial_ranges.size(); i++) { + input_iters.Set((*this)->initial_indices[i], initial_ranges[i]); + } + + // Unpack the output indices into linear combinations of the initial + // indices. + arith::Analyzer analyzer; + auto iter_map = DetectIterMap((*this)->final_indices, input_iters, 1, true, &analyzer); + CHECK(iter_map.size()) << "Index transformation was not bijective."; + + // Determine expressions for the input variables, in terms of the + // output variables. + Map<Var, PrimExpr> inverse_exprs_map = + InverseAffineIterMap(iter_map, Array<PrimExpr>(output_vars.begin(), output_vars.end())); + + // Unpack the map to an array, maintaining the same parameter order. + Array<PrimExpr> inverse_exprs; + for (const auto& index : (*this)->initial_indices) { + inverse_exprs.push_back(inverse_exprs_map.at(index)); + } + + return IndexMap(output_vars, inverse_exprs); +} + +Array<PrimExpr> IndexMapNode::MapIndices(const Array<PrimExpr>& indices) const { + ICHECK_EQ(indices.size(), initial_indices.size()); + + arith::Analyzer analyzer; + + for (size_t i = 0; i < initial_indices.size(); i++) { + analyzer.Bind(initial_indices[i], indices[i]); + } + + Array<PrimExpr> output; + for (const auto& output_dim : final_indices) { + output.push_back(analyzer.Simplify(output_dim)); + } + + return output; +} + +Array<Range> IndexMapNode::MapRanges(const Array<Range>& ranges) const { + ICHECK_EQ(ranges.size(), initial_indices.size()); + + Map<Var, Range> input_iters; + for (size_t i = 0; i < initial_indices.size(); i++) { + input_iters.Set(initial_indices[i], ranges[i]); + } + + arith::Analyzer analyzer; + auto iter_sums = DetectIterMap(final_indices, input_iters, 1, true, &analyzer); Review comment: Is `DetectIterMap` necessary? This looks like a strong constraint that the mapping should be affine (and is supported by our affine analysis). On the other hand, if we would like to get the range after the mapping, int set analysis will suffice. I'm think about the following alternative: ``` std::unordered_map<const VarNode*, arith::IntSet> dom_map; for (size_t i = 0; i < initial_indices.size(); i++) { dom_map.emplace(initial_indices[i], IntSet::FromRange(ranges[I])); } NDIntSet nd_int_set = NDIntSetFromPoint(final_indices); nd_int_set = NDIntSetEval(nd_int_set, dom_map); Array<Range> output; for (const auto& int_set : nd_int_set) { output.push_back(Range::FromMinExtent(int_set.min(), int_set.max() - int_set.min() + 1)); } ``` ########## File path: src/te/schedule/schedule_ops.cc ########## @@ -343,12 +367,16 @@ Stmt ScheduleOps(Schedule sch, Map<IterVar, Range> dom_map_, bool debug_keep_tri Stage s = sch->stages[i - 1]; ICHECK_NE(s->attach_type, kInline) << "call schedule.normalize before scheduleops"; ICHECK(s->op.defined()); - // no need to specify place holder op. - if (s->op.as<PlaceholderOpNode>()) continue; // Remove grouping sugar, get the real attach spec. Stage attach_spec = s.GetAttachSpec(); - if (scan_init.count(s->op)) { + if (s->op.as<PlaceholderOpNode>()) { + // Placeholders don't need any realize/provide statements, but + // may be annotated with set_physical_layout to indicate the + // physical layout of an input, and must still have the + // attribute given. + body = WrapPhysicalLayout(s, std::move(body)); Review comment: looks like this is a leftover of `physical_layout` renaming ########## File path: src/target/llvm/codegen_llvm.cc ########## @@ -931,15 +931,15 @@ llvm::Value* CodeGenLLVM::CreateIntrinsic(const CallNode* op) { } else if (op->op.same_as(builtin::tvm_storage_sync())) { return CreateStorageSync(op); } else if (op->op.same_as(builtin::address_of())) { - const LoadNode* l = op->args[0].as<LoadNode>(); - ICHECK(op->args.size() == 1 && l); - TypedPointer buffer_ptr; - if (const RampNode* r = l->index.as<RampNode>()) { - PrimExpr index = r->base / make_const(DataType::Int(32), r->lanes); - buffer_ptr = CreateBufferPtr(l->dtype, MakeValue(l->buffer_var), MakeValue(index)); - } else { - buffer_ptr = CreateBufferPtr(l->dtype, MakeValue(l->buffer_var), MakeValue(l->index)); + const BufferLoadNode* load = op->args[0].as<BufferLoadNode>(); + ICHECK(op->args.size() == 1 && load); + ICHECK_EQ(load->indices.size(), 0) << "LLVM only supports flat memory allocations."; + PrimExpr index = load->indices[0]; Review comment: nit: use `const PrimExpr&` to avoid copy ########## File path: src/target/llvm/codegen_llvm.cc ########## @@ -1189,14 +1189,24 @@ llvm::Value* CodeGenLLVM::VisitExpr_(const LetNode* op) { } llvm::Value* CodeGenLLVM::VisitExpr_(const LoadNode* op) { + LOG(FATAL) << "Unexpected deprecated LoadNode. Use BufferLoadNode instead."; + return NULL; +} + +llvm::Value* CodeGenLLVM::VisitExpr_(const BufferLoadNode* op) { + ICHECK_EQ(op->indices.size(), 1) << "CodeGenLLVM expects flattened 1-d buffers."; + DataType t = op->dtype; - bool is_volatile = volatile_buf_.count(op->buffer_var.get()); - llvm::Value* buffer = MakeValue(op->buffer_var); - llvm::Value* index = MakeValue(op->index); + Var buffer_var = op->buffer->data; + PrimExpr buffer_index = op->indices[0]; Review comment: const reference -- 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]
