[misc] Rc v1.1.1 patch1 (#5820)

* [bug] Fix incorrect autodiff_mode information in offline cache key (#5737) * [bug] Fix incorrect autodiff_mode infomation in offline cache key * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Error] Do not show warning when the offline cache path does not exist (#5747) * [Error] Do not show warning when the offline cache path does not exist * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> * [Bug] [type] Fix wrong type cast in codegen of storing quant floats (#5818) * [autodiff] Support shift ptr in dynamic index (#5770) * [autodiff] Support shift ptr in dynamic index * update the offset * update offset * add dynamic index test for ad * forcely enable dynamic index for polar decompose Co-authored-by: Mingming Zhang <pgzxb@qq.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Lin Jiang <linjiang@taichi.graphics> Co-authored-by: Yi Xu <xy_xuyi@foxmail.com> Co-authored-by: Mingrui Zhang <33411325+erizmr@users.noreply.github.com>
taichi-dev · Aug 18, 2022 · aae46db · aae46db
1 parent f5bb646
commit aae46db
Show file tree

Hide file tree

Showing 8 changed files with 151 additions and 24 deletions.
diff --git a/taichi/analysis/offline_cache_util.cpp b/taichi/analysis/offline_cache_util.cpp
@@ -161,14 +161,16 @@ std::string get_hashed_offline_cache_key(CompileConfig *config,
     compile_config_key = get_offline_cache_key_of_compile_config(config);
   }
 
+  std::string autodiff_mode =
+      std::to_string(static_cast<std::size_t>(kernel->autodiff_mode));
   picosha2::hash256_one_by_one hasher;
   hasher.process(compile_config_key.begin(), compile_config_key.end());
   hasher.process(kernel_ast_string.begin(), kernel_ast_string.end());
+  hasher.process(autodiff_mode.begin(), autodiff_mode.end());
   hasher.finish();
 
   auto res = picosha2::get_hash_hex_string(hasher);
-  res.insert(res.begin(),
-             kernel->autodiff_mode != AutodiffMode::kNone ? 'g' : 'n');
+  res.insert(res.begin(), 'T');  // The key must start with a letter
   return res;
 }
 

diff --git a/taichi/codegen/llvm/codegen_llvm_quant.cpp b/taichi/codegen/llvm/codegen_llvm_quant.cpp
@@ -221,7 +221,7 @@ void TaskCodeGenLLVM::visit(BitStructStoreStmt *stmt) {
           create_call("max_i32", {exponent_bits, tlctx->get_constant(0)});
 
       // Compute the bit pointer of the exponent bits.
-      val = builder->CreateBitCast(exponent_bits, physical_type);
+      val = builder->CreateIntCast(exponent_bits, physical_type, false);
       val = builder->CreateShl(val, bit_struct->get_member_bit_offset(exp));
 
       if (bit_struct_val == nullptr) {
@@ -238,7 +238,7 @@ void TaskCodeGenLLVM::visit(BitStructStoreStmt *stmt) {
                               tlctx->get_constant(0));
       val = builder->CreateSelect(exp_non_zero, digit_bits,
                                   tlctx->get_constant(0));
-      val = builder->CreateBitCast(val, physical_type);
+      val = builder->CreateIntCast(val, physical_type, false);
       val = builder->CreateShl(val, bit_struct->get_member_bit_offset(ch_id));
     } else {
       val = quant_int_or_quant_fixed_to_bits(val, dtype, physical_type);

diff --git a/taichi/runtime/llvm/llvm_offline_cache.cpp b/taichi/runtime/llvm/llvm_offline_cache.cpp
@@ -376,6 +376,9 @@ void LlvmOfflineCacheFileWriter::clean_cache(const std::string &path,
   {
     std::string lock_path = taichi::join_path(path, kMetadataFileLockName);
     if (!lock_with_file(lock_path)) {
+      if (!taichi::path_exists(path)) {
+        return;
+      }
       TI_WARN("Lock {} failed", lock_path);
       return;
     }

diff --git a/taichi/transforms/auto_diff.cpp b/taichi/transforms/auto_diff.cpp
@@ -14,12 +14,14 @@ class IndependentBlocksJudger : public BasicStmtVisitor {
 
   void visit(LocalLoadStmt *stmt) override {
     for (auto &lane : stmt->src.data) {
-      touched_allocas_.insert(lane.var->as<AllocaStmt>());
+      TI_ASSERT(lane.var->is<AllocaStmt>() || lane.var->is<PtrOffsetStmt>());
+      touched_allocas_.insert(lane.var);
     }
   }
 
   void visit(LocalStoreStmt *stmt) override {
-    touched_allocas_.insert(stmt->dest->as<AllocaStmt>());
+    TI_ASSERT(stmt->dest->is<AllocaStmt>() || stmt->dest->is<PtrOffsetStmt>());
+    touched_allocas_.insert(stmt->dest);
   }
 
   void visit(AtomicOpStmt *stmt) override {
@@ -75,7 +77,7 @@ class IndependentBlocksJudger : public BasicStmtVisitor {
   }
 
  private:
-  std::set<AllocaStmt *> touched_allocas_;
+  std::set<Stmt *> touched_allocas_;
   bool qualified_atomics_ = true;
   bool inner_most_loop_ = true;
   bool is_inside_loop_ = false;
@@ -578,6 +580,10 @@ class ADTransform : public IRVisitor {
     // do nothing.
   }
 
+  void visit(PtrOffsetStmt *stmt) override {
+    // do nothing.
+  }
+
   void visit(PrintStmt *print_stmt) override {
     // do nothing
   }
@@ -989,7 +995,21 @@ class MakeAdjoint : public ADTransform {
 
   void visit(GlobalLoadStmt *stmt) override {
     // issue global store to adjoint
-    GlobalPtrStmt *src = stmt->src->as<GlobalPtrStmt>();
+    if (stmt->src->is<ExternalPtrStmt>()) {
+      TI_ERROR(
+          "Importing data from external array (such as numpy array) not "
+          "supported in AutoDiff for now")
+    }
+
+    GlobalPtrStmt *src = nullptr;
+    bool is_ptr_offset = false;
+    if (stmt->src->is<PtrOffsetStmt>()) {
+      is_ptr_offset = true;
+      src = stmt->src->as<PtrOffsetStmt>()->origin->as<GlobalPtrStmt>();
+    } else {
+      src = stmt->src->as<GlobalPtrStmt>();
+    }
+
     TI_ASSERT(src->width() == 1);
     auto snodes = src->snodes;
     if (!snodes[0]->has_adjoint()) {
@@ -1003,12 +1023,30 @@ class MakeAdjoint : public ADTransform {
     TI_ASSERT(snodes[0]->get_adjoint() != nullptr);
     snodes[0] = snodes[0]->get_adjoint();
     auto adj_ptr = insert<GlobalPtrStmt>(snodes, src->indices);
+    if (is_ptr_offset) {
+      adj_ptr = insert<PtrOffsetStmt>(adj_ptr,
+                                      stmt->src->as<PtrOffsetStmt>()->offset);
+    }
     insert<AtomicOpStmt>(AtomicOpType::add, adj_ptr, load(adjoint(stmt)));
   }
 
   void visit(GlobalStoreStmt *stmt) override {
     // erase and replace with global load adjoint
-    GlobalPtrStmt *dest = stmt->dest->as<GlobalPtrStmt>();
+    if (stmt->dest->is<ExternalPtrStmt>()) {
+      TI_ERROR(
+          "Exporting data to external array (such as numpy array) not "
+          "supported in AutoDiff for now")
+    }
+
+    GlobalPtrStmt *dest = nullptr;
+    bool is_ptr_offset = false;
+    if (stmt->dest->is<PtrOffsetStmt>()) {
+      is_ptr_offset = true;
+      dest = stmt->dest->as<PtrOffsetStmt>()->origin->as<GlobalPtrStmt>();
+    } else {
+      dest = stmt->dest->as<GlobalPtrStmt>();
+    }
+
     TI_ASSERT(dest->width() == 1);
     auto snodes = dest->snodes;
     if (!snodes[0]->has_adjoint()) {
@@ -1018,24 +1056,40 @@ class MakeAdjoint : public ADTransform {
     TI_ASSERT(snodes[0]->get_adjoint() != nullptr);
     snodes[0] = snodes[0]->get_adjoint();
     auto adjoint_ptr = insert<GlobalPtrStmt>(snodes, dest->indices);
-    auto load = insert<GlobalLoadStmt>(adjoint_ptr);
-    accumulate(stmt->val, load);
+    if (is_ptr_offset) {
+      adjoint_ptr = insert<PtrOffsetStmt>(
+          adjoint_ptr, stmt->dest->as<PtrOffsetStmt>()->offset);
+    }
+    accumulate(stmt->val, insert<GlobalLoadStmt>(adjoint_ptr));
     stmt->parent->erase(stmt);
   }
 
   void visit(AtomicOpStmt *stmt) override {
     // erase and replace with global load adjoint
-    GlobalPtrStmt *dest = stmt->dest->as<GlobalPtrStmt>();
+    GlobalPtrStmt *dest = nullptr;
+    bool is_ptr_offset = false;
+    if (stmt->dest->is<PtrOffsetStmt>()) {
+      is_ptr_offset = true;
+      dest = stmt->dest->as<PtrOffsetStmt>()->origin->as<GlobalPtrStmt>();
+    } else {
+      dest = stmt->dest->as<GlobalPtrStmt>();
+    }
+
     TI_ASSERT(dest->width() == 1);
     auto snodes = dest->snodes;
-    if (snodes[0]->has_adjoint()) {
-      TI_ASSERT(snodes[0]->get_adjoint() != nullptr);
-      snodes[0] = snodes[0]->get_adjoint();
-      auto adjoint_ptr = insert<GlobalPtrStmt>(snodes, dest->indices);
-      accumulate(stmt->val, insert<GlobalLoadStmt>(adjoint_ptr));
-    } else {
+    if (!snodes[0]->has_adjoint()) {
       // no gradient (likely integer types)
+      return;
     }
+
+    TI_ASSERT(snodes[0]->get_adjoint() != nullptr);
+    snodes[0] = snodes[0]->get_adjoint();
+    auto adjoint_ptr = insert<GlobalPtrStmt>(snodes, dest->indices);
+    if (is_ptr_offset) {
+      adjoint_ptr = insert<PtrOffsetStmt>(
+          adjoint_ptr, stmt->dest->as<PtrOffsetStmt>()->offset);
+    }
+    accumulate(stmt->val, insert<GlobalLoadStmt>(adjoint_ptr));
     stmt->parent->erase(stmt);
   }
 };
@@ -1278,7 +1332,14 @@ class MakeDual : public ADTransform {
 
   void visit(GlobalLoadStmt *stmt) override {
     // issue global store to dual
-    GlobalPtrStmt *src = stmt->src->as<GlobalPtrStmt>();
+    GlobalPtrStmt *src = nullptr;
+    bool is_ptr_offset = false;
+    if (stmt->src->is<PtrOffsetStmt>()) {
+      is_ptr_offset = true;
+      src = stmt->src->as<PtrOffsetStmt>()->origin->as<GlobalPtrStmt>();
+    } else {
+      src = stmt->src->as<GlobalPtrStmt>();
+    }
     TI_ASSERT(src->width() == 1);
     auto snodes = src->snodes;
     if (!snodes[0]->has_dual()) {
@@ -1292,11 +1353,22 @@ class MakeDual : public ADTransform {
     TI_ASSERT(snodes[0]->get_dual() != nullptr);
     snodes[0] = snodes[0]->get_dual();
     auto dual_ptr = insert<GlobalPtrStmt>(snodes, src->indices);
+    if (is_ptr_offset) {
+      dual_ptr = insert<PtrOffsetStmt>(dual_ptr,
+                                       stmt->src->as<PtrOffsetStmt>()->offset);
+    }
     accumulate(stmt, insert<GlobalLoadStmt>(dual_ptr));
   }
 
   void visit(GlobalStoreStmt *stmt) override {
-    GlobalPtrStmt *dest = stmt->dest->as<GlobalPtrStmt>();
+    GlobalPtrStmt *dest = nullptr;
+    bool is_ptr_offset = false;
+    if (stmt->dest->is<PtrOffsetStmt>()) {
+      is_ptr_offset = true;
+      dest = stmt->dest->as<PtrOffsetStmt>()->origin->as<GlobalPtrStmt>();
+    } else {
+      dest = stmt->dest->as<GlobalPtrStmt>();
+    }
     TI_ASSERT(dest->width() == 1);
     auto snodes = dest->snodes;
     if (!snodes[0]->has_dual()) {
@@ -1306,11 +1378,22 @@ class MakeDual : public ADTransform {
     TI_ASSERT(snodes[0]->get_dual() != nullptr);
     snodes[0] = snodes[0]->get_dual();
     auto dual_ptr = insert<GlobalPtrStmt>(snodes, dest->indices);
+    if (is_ptr_offset) {
+      dual_ptr = insert<PtrOffsetStmt>(dual_ptr,
+                                       stmt->dest->as<PtrOffsetStmt>()->offset);
+    }
     insert<AtomicOpStmt>(AtomicOpType::add, dual_ptr, load(dual(stmt->val)));
   }
 
   void visit(AtomicOpStmt *stmt) override {
-    GlobalPtrStmt *dest = stmt->dest->as<GlobalPtrStmt>();
+    GlobalPtrStmt *dest = nullptr;
+    bool is_ptr_offset = false;
+    if (stmt->dest->is<PtrOffsetStmt>()) {
+      is_ptr_offset = true;
+      dest = stmt->dest->as<PtrOffsetStmt>()->origin->as<GlobalPtrStmt>();
+    } else {
+      dest = stmt->dest->as<GlobalPtrStmt>();
+    }
     TI_ASSERT(dest->width() == 1);
     auto snodes = dest->snodes;
     if (!snodes[0]->has_dual()) {
@@ -1320,6 +1403,10 @@ class MakeDual : public ADTransform {
     TI_ASSERT(snodes[0]->get_dual() != nullptr);
     snodes[0] = snodes[0]->get_dual();
     auto dual_ptr = insert<GlobalPtrStmt>(snodes, dest->indices);
+    if (is_ptr_offset) {
+      dual_ptr = insert<PtrOffsetStmt>(dual_ptr,
+                                       stmt->dest->as<PtrOffsetStmt>()->offset);
+    }
     insert<AtomicOpStmt>(AtomicOpType::add, dual_ptr, load(dual(stmt->val)));
   }
 };

diff --git a/taichi/util/io.h b/taichi/util/io.h
@@ -10,13 +10,19 @@
 #include <vector>
 #include <cstdio>
 #include <cstdlib>
+#include <sys/stat.h>
 
 #if defined(TI_PLATFORM_WINDOWS)
 #include <filesystem>
 #endif
 
 TI_NAMESPACE_BEGIN
 
+inline bool path_exists(const std::string &dir) {
+  struct stat buffer;
+  return stat(dir.c_str(), &buffer) == 0;
+}
+
 // TODO: move to std::filesystem after it's nonexperimental on all platforms
 inline void create_directories(const std::string &dir) {
 #if defined(TI_PLATFORM_WINDOWS)

diff --git a/tests/python/test_ad_dynamic_index.py b/tests/python/test_ad_dynamic_index.py
@@ -0,0 +1,28 @@
+import taichi as ti
+from tests import test_utils
+
+
+@test_utils.test(require=ti.extension.dynamic_index,
+                 dynamic_index=True,
+                 debug=True)
+def test_matrix_non_constant_index():
+    m = ti.Matrix.field(2, 2, ti.f32, 5, needs_grad=True)
+    n = ti.Matrix.field(2, 2, ti.f32, 5, needs_grad=True)
+    loss = ti.field(ti.f32, (), needs_grad=True)
+
+    n.fill(0)
+
+    @ti.kernel
+    def func1():
+        for i in range(5):
+            for j, k in ti.ndrange(2, 2):
+                m[i][j, k] = (j + 1) * (k + 1) * n[i][j, k]
+                loss[None] += m[i][j, k]
+
+    loss.grad[None] = 1.0
+    func1.grad()
+
+    for i in range(5):
+        for j in range(2):
+            for k in range(2):
+                assert n.grad[i][j, k] == (j + 1) * (k + 1)
diff --git a/tests/python/test_ad_math_func.py b/tests/python/test_ad_math_func.py
@@ -2,7 +2,7 @@
 from tests import test_utils
 
 
-@test_utils.test(require=ti.extension.adstack, dynamic_index=False)
+@test_utils.test(require=ti.extension.adstack, dynamic_index=True)
 def test_polar_decompose_2D():
     # `polar_decompose3d` in current Taichi version (v1.1) does not support autodiff,
     # becasue it mixed usage of for-loops and statements without looping.

diff --git a/tests/python/test_quant_float.py b/tests/python/test_quant_float.py
@@ -6,12 +6,13 @@
 from tests import test_utils
 
 
+@pytest.mark.parametrize('max_num_bits', [32, 64])
 @test_utils.test(require=ti.extension.quant)
-def test_quant_float_unsigned():
+def test_quant_float_unsigned(max_num_bits):
     qflt = ti.types.quant.float(exp=6, frac=13, signed=False)
     x = ti.field(dtype=qflt)
 
-    bitpack = ti.BitpackedFields(max_num_bits=32)
+    bitpack = ti.BitpackedFields(max_num_bits=max_num_bits)
     bitpack.place(x)
     ti.root.place(bitpack)