[Bug] Fix Loop-Invariant-Cache for dynamic indexed pointers

.github/workflows/scripts/common-utils.sh

@@ -37,7 +37,7 @@ install_taichi_wheel() {
     python3 -m pip install dist/*.whl
     if [ -z "$GPU_TEST" ]; then
         python3 -m pip install -r requirements_test.txt
-        python3 -m pip install "torch>1.12.0; python_version < '3.10'"
+        python3 -m pip install "torch>1.12.0,<2.0; python_version < '3.10'"
         # Paddle's develop package doesn't support CI's MACOS machine at present
         if [[ $OSTYPE == "linux-"* ]]; then
             python3 -m pip install "paddlepaddle==2.3.0; python_version < '3.10'"

requirements_test.txt

@@ -7,7 +7,7 @@ numpy
 psutil
 autograd
 requests
-matplotlib
+matplotlib<=3.7.3
 cffi
 scipy
 setproctitle
@@ -16,3 +16,4 @@ marko
 PyYAML
 pyevtk
 wget
+urllib3<2.0

taichi/analysis/gather_dynamically_indexed_pointers.cpp

@@ -0,0 +1,124 @@
+#include "taichi/analysis/gather_uniquely_accessed_pointers.h"
+#include "taichi/ir/ir.h"
+#include "taichi/ir/analysis.h"
+#include "taichi/ir/statements.h"
+#include "taichi/ir/visitors.h"
+#include <algorithm>
+
+namespace taichi::lang {
+
+bool is_leaf_nodes_on_same_branch(SNode *snode0, SNode *snode1) {
+  // Verify: place snode
+  if (!snode0->is_place() || !snode1->is_place()) {
+    return false;
+  }
+
+  // Check parent snode
+  if (snode0->parent != snode1->parent) {
+    return false;
+  }
+
+  return true;
+}
+
+class DynamicIndexingAnalyzer : public BasicStmtVisitor {
+  void record_dynamic_indexed_ptr(ExternalPtrStmt *extern_ptr) {
+    dynamically_indexed_ptrs_.insert(extern_ptr);
+    // Find aliased ExternPtrStmt
+    for (auto *other_extern_ptr : extern_ptrs_) {
+      if (other_extern_ptr != extern_ptr &&
+          other_extern_ptr->base_ptr == extern_ptr->base_ptr) {
+        // Aliased ExternalPtrStmt, with same base_ptr and outter index
+        dynamically_indexed_ptrs_.insert(other_extern_ptr);
+      }
+    }
+  }
+
+  void record_dynamic_indexed_ptr(GlobalPtrStmt *global_ptr) {
+    dynamically_indexed_ptrs_.insert(global_ptr);
+    // Find aliased GlobalPtrStmt
+    for (auto *other_global_ptr : global_ptrs_) {
+      if (other_global_ptr != global_ptr &&
+          is_leaf_nodes_on_same_branch(other_global_ptr->snode,
+                                       global_ptr->snode)) {
+        dynamically_indexed_ptrs_.insert(other_global_ptr);
+      }
+    }
+  }
+
+ public:
+  explicit DynamicIndexingAnalyzer(IRNode *node) {
+  }
+
+  void visit(GlobalPtrStmt *stmt) override {
+    for (auto *index_stmt : stmt->indices) {
+      if (!index_stmt->is<ConstStmt>() && !index_stmt->is<LoopIndexStmt>()) {
+        record_dynamic_indexed_ptr(stmt);
+      }
+    }
+
+    global_ptrs_.insert(stmt);
+  }
+
+  void visit(ExternalPtrStmt *stmt) override {
+    for (auto *index_stmt : stmt->indices) {
+      if (!index_stmt->is<ConstStmt>() && !index_stmt->is<LoopIndexStmt>()) {
+        record_dynamic_indexed_ptr(stmt);
+      }
+    }
+
+    extern_ptrs_.insert(stmt);
+  }
+
+  void visit(MatrixPtrStmt *stmt) override {
+    GlobalPtrStmt *global_ptr = nullptr;
+    ExternalPtrStmt *extern_ptr = nullptr;
+
+    if (stmt->origin->is<GlobalPtrStmt>()) {
+      global_ptr = stmt->origin->as<GlobalPtrStmt>();
+    } else if (stmt->origin->is<ExternalPtrStmt>()) {
+      extern_ptr = stmt->origin->as<ExternalPtrStmt>();
+    } else {
+      return;
+    }
+
+    // Is dynamic index
+    if (stmt->offset->is<ConstStmt>()) {
+      return;
+    }
+
+    if (global_ptr) {
+      record_dynamic_indexed_ptr(global_ptr);
+    }
+
+    if (extern_ptr) {
+      record_dynamic_indexed_ptr(extern_ptr);
+    }
+  }
+
+  std::unordered_set<Stmt *> get_dynamically_indexed_ptrs() {
+    return dynamically_indexed_ptrs_;
+  }
+
+ private:
+  using BasicStmtVisitor::visit;
+  std::unordered_set<Stmt *> dynamically_indexed_ptrs_;
+  std::unordered_set<GlobalPtrStmt *> global_ptrs_;
+  std::unordered_set<ExternalPtrStmt *> extern_ptrs_;
+};
+
+namespace irpass::analysis {
+
+std::unordered_set<Stmt *> gather_dynamically_indexed_pointers(IRNode *root) {
+  DynamicIndexingAnalyzer pass(root);
+
+  // This pass is intended to run twice
+  root->accept(&pass);
+  root->accept(&pass);
+
+  auto dynamically_indexed_ptrs = pass.get_dynamically_indexed_ptrs();
+  return dynamically_indexed_ptrs;
+}
+
+}  // namespace irpass::analysis
+}  // namespace taichi::lang

taichi/ir/analysis.h

@@ -111,6 +111,7 @@ std::tuple<std::unordered_map<const SNode *, GlobalPtrStmt *>,
                               hashing::Hasher<std::vector<int>>>,
            std::unordered_set<MatrixPtrStmt *>>
 gather_uniquely_accessed_pointers(IRNode *root);
+std::unordered_set<Stmt *> gather_dynamically_indexed_pointers(IRNode *root);
 
 std::unique_ptr<std::unordered_set<AtomicOpStmt *>> gather_used_atomics(
     IRNode *root);

taichi/transforms/cache_loop_invariant_global_vars.cpp

@@ -26,6 +26,8 @@ class CacheLoopInvariantGlobalVars : public LoopInvariantDetector {
       loop_unique_arr_ptr_;
   std::unordered_set<MatrixPtrStmt *> loop_unique_matrix_ptr_;
 
+  std::unordered_set<Stmt *> dynamic_indexed_ptrs_;
+
   OffloadedStmt *current_offloaded;
 
   explicit CacheLoopInvariantGlobalVars(const CompileConfig &config)
@@ -44,6 +46,9 @@ class CacheLoopInvariantGlobalVars : public LoopInvariantDetector {
           std::move(std::get<2>(uniquely_accessed_pointers));
     }
     current_offloaded = stmt;
+    dynamic_indexed_ptrs_ =
+        irpass::analysis::gather_dynamically_indexed_pointers(stmt);
+
     // We don't need to visit TLS/BLS prologues/epilogues.
     if (stmt->body) {
       if (stmt->task_type == OffloadedStmt::TaskType::range_for ||
@@ -56,6 +61,28 @@ class CacheLoopInvariantGlobalVars : public LoopInvariantDetector {
     current_offloaded = nullptr;
   }
 
+  bool is_dynamically_indexed(Stmt *stmt) {
+    // Handle GlobalPtrStmt
+    Stmt *ptr_stmt = nullptr;
+    if (stmt->is<GlobalPtrStmt>()) {
+      ptr_stmt = stmt->as<GlobalPtrStmt>();
+    } else if (stmt->is<MatrixPtrStmt>() &&
+               stmt->as<MatrixPtrStmt>()->origin->is<GlobalPtrStmt>()) {
+      ptr_stmt = stmt->as<MatrixPtrStmt>()->origin->as<GlobalPtrStmt>();
+    } else if (stmt->is<ExternalPtrStmt>()) {
+      ptr_stmt = stmt->as<ExternalPtrStmt>();
+    } else if (stmt->is<MatrixPtrStmt>() &&
+               stmt->as<MatrixPtrStmt>()->origin->is<ExternalPtrStmt>()) {
+      ptr_stmt = stmt->as<MatrixPtrStmt>()->origin->as<ExternalPtrStmt>();
+    }
+
+    if (ptr_stmt && dynamic_indexed_ptrs_.count(ptr_stmt)) {
+      return true;
+    }
+
+    return false;
+  }
+
   bool is_offload_unique(Stmt *stmt) {
     if (current_offloaded->task_type == OffloadedTaskType::serial) {
       return true;
@@ -174,6 +201,9 @@ class CacheLoopInvariantGlobalVars : public LoopInvariantDetector {
 
   std::optional<int> find_cache_depth_if_cacheable(Stmt *operand,
                                                    Block *current_scope) {
+    if (is_dynamically_indexed(operand)) {
+      return std::nullopt;
+    }
     if (!is_offload_unique(operand)) {
       return std::nullopt;
     }

tests/python/test_cache_loop_invariant.py

@@ -0,0 +1,27 @@
+import pytest
+from taichi.lang import impl
+
+import taichi as ti
+from tests import test_utils
+
+
+@test_utils.test(arch=[ti.cuda, ti.cpu])
+def test_local_matrix_non_constant_index_real_matrix():
+    N = 1
+    x = ti.Vector.field(3, float, shape=1)
+
+    @ti.kernel
+    def test_invariant_cache():
+        for i in range(1):
+            x[i][1] = x[i][1] + 1.0
+            for j in range(1):
+                x[i][1] = x[i][1] - 5.0
+                for z in range(1):
+                    idx = 0
+                    if z == 0:
+                        idx = 1
+                    x_print = x[i][idx]
+
+                    assert x_print == x[i][1]
+
+    test_invariant_cache()