!47 enhance vectorization analysis in auto tiling

Merge pull request !47 from yangsijia/enhance-vectorization-analysis

src/poly/space_analyzer.cc

@@ -311,24 +311,45 @@ void SpaceAnalyzer::IdentifyInsnType() {
 void SpaceAnalyzer::IdentifyVectorizedAxes() {
   if (provides_ana_.empty()) return;
   std::string attr_key = "VECTORIZED";
-  std::unordered_map<std::string, const For *> last_axes;
-
+  std::unordered_set<std::string> unsupported_insn = {"REDUCE", "TRANSFORM", "TRANSPOSE"};
+  std::unordered_map<std::string, const For *> mark_dst_axes;
   for (auto it : provides_ana_) {
     std::vector<ProvideEntry> pes = it.second;
     for (auto pe : pes) {
-      if (pe.src.size() != 1U) continue;
+      bool skip = false;
+      for (auto insn : unsupported_insn) {
+        if (pe.basic_op_type.find(insn) != std::string::npos) {
+          skip = true;
+          break;
+        }
+      }
+      if (skip) {
+        continue;
+      }
       Tensor dst_tensor = pe.dst;
-      Tensor src_tensor = pe.src[0];
-      if (dst_tensor.var_names.size() != 1U || src_tensor.var_names.size() != 1U) continue;
-      const For *dst_last = GetBufferLastAxis(dst_tensor);
-      const For *src_last = GetBufferLastAxis(src_tensor);
-      if (dst_last != nullptr && src_last != nullptr && dst_last == src_last) {
-        last_axes[dst_tensor.name] = dst_last;
+      const For *dst_last = GetBufferInnerAxis(dst_tensor);
+      // skip if dst is scalar
+      if (dst_last == nullptr) {
+        continue;
+      }
+
+      const For *src_last = nullptr;
+      for (auto src : pe.src) {
+        const auto *last = GetBufferInnerAxis(src);
+        if (last != nullptr && last == dst_last) {
+          src_last = last;
+          break;
+        }
+      }
+      // skip if src tensor does not share same inner-most axis with dst tensor
+      if (src_last == nullptr && !pe.src.empty()) {
+        continue;
       }
+      mark_dst_axes[dst_tensor.name] = dst_last;
     }
   }
 
-  for (auto la : last_axes) {
+  for (auto la : mark_dst_axes) {
     TileAxis *last_axis = analyzer_->Axis(la.second);
     if (last_axis != nullptr) last_axis->MarkWithAttr(AttrInfo{attr_key, la.first});
   }
@@ -354,7 +375,7 @@ void SpaceAnalyzer::IdentifyDmaUnderCondition() {
       };
       ktvm::ir::PostOrderVisit(pe.cond->condition, DetectToT);
       if (!contain_tot) continue;
-      TileAxis *tot_axis = analyzer_->Axis(GetBufferLastAxis(pe.dst));
+      TileAxis *tot_axis = analyzer_->Axis(GetBufferInnerAxis(pe.dst));
       if (tot_axis != nullptr) tot_axis->MarkWithAttr(AttrInfo{attr_key, ""});
     }
   }
@@ -371,14 +392,14 @@ void SpaceAnalyzer::IdentifyAlignAxes() {
       std::vector<Tensor> src_tensors = pe.src;
       Tensor dst_tensor = pe.dst;
       if (pe.basic_op_type.find("TRANSPOSE") != std::string::npos) {
-        const For *dst_last = GetBufferLastAxis(dst_tensor);
+        const For *dst_last = GetBufferInnerAxis(dst_tensor);
         if (dst_last != nullptr) {
           align_axes_attrs[dst_last] = std::make_pair(dst_tensor.name, pe.basic_op_type);
         } else {
           analyzer_->RootAxis()->MarkWithAttr(AttrInfo{"TRANSFORM", dst_tensor.name});
         }
       } else if (pe.basic_op_type.find("DMA") != std::string::npos) {
-        const For *dst_last = GetBufferLastAxis(dst_tensor);
+        const For *dst_last = GetBufferInnerAxis(dst_tensor);
         if (dst_last != nullptr) {
           align_axes_attrs[dst_last] = std::make_pair(dst_tensor.name, pe.basic_op_type);
         } else {
@@ -394,7 +415,7 @@ void SpaceAnalyzer::IdentifyAlignAxes() {
         }
         for (auto t : src_tensors) {
           if (t.loops.size() <= dst_tensor.loops.size()) continue;
-          const For *src_last = GetBufferLastAxis(t);
+          const For *src_last = GetBufferInnerAxis(t);
           if (src_last != nullptr) {
             align_axes_attrs[src_last] = std::make_pair(t.name, pe.basic_op_type);
           }
@@ -407,7 +428,7 @@ void SpaceAnalyzer::IdentifyAlignAxes() {
                                                                         const Tensor dst_tensor) {
           for (auto src : src_tensors) {
             if (src.name != dst_tensor.name) {
-              src_last = GetBufferLastAxis(src);
+              src_last = GetBufferInnerAxis(src);
               src_name = src.name;
               break;
             }
@@ -416,7 +437,7 @@ void SpaceAnalyzer::IdentifyAlignAxes() {
           if (const auto i = src_last->extent.as<IntImm>()) gm_block = i->value;
         };
         if (pe.basic_op_type.find("REDUCE") != std::string::npos) {
-          const For *dst_last = GetBufferLastAxis(dst_tensor);
+          const For *dst_last = GetBufferInnerAxis(dst_tensor);
           int64_t ub_block = 1;
           if (dst_last != nullptr) {
             align_axes_attrs[dst_last] = std::make_pair(dst_tensor.name, pe.basic_op_type);
@@ -430,7 +451,7 @@ void SpaceAnalyzer::IdentifyAlignAxes() {
             }
           }
         } else if (pe.basic_op_type.find("BROADCAST") != std::string::npos) {
-          const For *dst_last = GetBufferLastAxis(dst_tensor);
+          const For *dst_last = GetBufferInnerAxis(dst_tensor);
           int64_t ub_block = 1;
           if (dst_last == nullptr) continue;
           if (const auto i = dst_last->extent.as<IntImm>()) ub_block = i->value;
@@ -451,7 +472,7 @@ void SpaceAnalyzer::IdentifyAlignAxes() {
   }
 }
 
-const For *SpaceAnalyzer::GetBufferLastAxis(Tensor t, int offset) {
+const For *SpaceAnalyzer::GetBufferInnerAxis(Tensor t, int offset) {
   int last_dim = static_cast<int>(t.var_names.size()) - offset;
   auto it = t.loops.find(last_dim);
   if (it != t.loops.end() && it->second.size() == 1U) return it->second[0];
@@ -470,11 +491,11 @@ void SpaceAnalyzer::IdentifyReduceAxes() {
     std::vector<ProvideEntry> pes = it.second;
     for (auto pe : pes) {
       if ((pe.basic_op_type.find("REDUCE") == std::string::npos)) continue;
-      const For *dst_last = GetBufferLastAxis(pe.dst);
+      const For *dst_last = GetBufferInnerAxis(pe.dst);
       if (dst_last == nullptr) {
         // Reduce op like A[i, 0] = A[i, 0] op B[i, j], we need to mark axis `i` as dst last for dma align.
         for (auto offset = 0; offset < static_cast<int>(pe.dst.var_names.size()); ++offset) {
-          dst_last = GetBufferLastAxis(pe.dst, offset + 1);
+          dst_last = GetBufferInnerAxis(pe.dst, offset + 1);
           if (dst_last == nullptr) continue;
           MarkAttr(dst_last, "REDUCE_DST_LAST", pe.dst.name);
           break;
@@ -484,7 +505,7 @@ void SpaceAnalyzer::IdentifyReduceAxes() {
       }
       for (Tensor src : pe.src) {
         if (src.name == pe.dst.name) continue;
-        const For *src_last = GetBufferLastAxis(src);
+        const For *src_last = GetBufferInnerAxis(src);
         MarkAttr(src_last, "REDUCE_SRC_LAST", src.name);
         std::string flow = src.name + "->" + pe.dst.name;
         root->MarkWithAttr(AttrInfo{"REDUCE_FLOW", flow});

src/poly/space_analyzer.h

@@ -67,7 +67,7 @@ class SpaceAnalyzer {
   // Provides stmt after analysis.
   std::unordered_map<const For *, std::vector<ProvideEntry>> provides_ana_;
 
-  const For *GetBufferLastAxis(Tensor t, int offset = 1);
+  const For *GetBufferInnerAxis(Tensor t, int offset = 1);
   // generalized cases
   void IdentifyInsnType();
   void IdentifyDmaUnderCondition();

src/poly/tiling_solver.cc

@@ -462,7 +462,9 @@ int64_t InequalitySolver::DetermineTileForStatic(TileAxis *axis, const Expr &mem
       ss << "--> Init factor " << final_factor;
 
       auto mod_value = cons.tile_mod_.as<IntImm>() ? cons.tile_mod_.as<IntImm>()->value : 1;
-      if (static_shape >= mod_value && final_factor % mod_value != 0) {
+      bool is_unaligned = (static_shape >= mod_value && final_factor % mod_value != 0);
+      bool need_to_align = (final_factor > mod_value || !axis->HasAttr("VECTORIZED"));
+      if (is_unaligned && need_to_align) {
         final_factor = std::max(static_cast<int>(final_factor / mod_value * mod_value), 1);
         ss << "--> Mod value " << mod_value << " --> Align to mod " << final_factor;
       }
@@ -810,12 +812,15 @@ bool TraverseSolver::IsTilable(TileInfo *info) {
   }
 
   if (level == LEVEL1) {
-    min_tile = cons.tile_mod_.as<IntImm>()->value;
-
-    if ((info->axis->forbid_iso && const_extent % min_tile != 0) || (cons.tile_min_.as<IntImm>()->value > min_tile) ||
-        (cons.tile_min_.as<IntImm>()->value == MIN_TILE)) {
+    bool use_tile_min = (info->axis->forbid_iso && const_extent % cons.tile_mod_.as<IntImm>()->value != 0) ||
+                        (cons.tile_min_.as<IntImm>()->value == MIN_TILE) || (axis->HasAttr("VECTORIZED")) ||
+                        (cons.tile_min_.as<IntImm>()->value > cons.tile_mod_.as<IntImm>()->value);
+    if (use_tile_min) {
       min_tile = cons.tile_min_.as<IntImm>()->value;
+    } else {
+      min_tile = cons.tile_mod_.as<IntImm>()->value;
     }
+
     if (axis->range_min > min_tile) {
       min_tile = axis->range_min;
     }

src/poly/tiling_strategy_manager.cc

@@ -222,20 +222,20 @@ void VectorizedStrategy::AddConstraint() {
     return;
   }
   for (auto axis : analyzer_->GetAxesOfAttr("VECTORIZED")) {
-    if (axis->HasAttr("DYNAMIC_BOUND")) {
+    if (axis->HasAttr("DYNAMIC_BOUND") || axis->range_extent.as<IntImm>() == nullptr) {
       continue;
     }
     int64_t min_byte = -1;
-    if (axis->data_size.empty()) {
-      min_byte = 1;
-    } else {
-      for (const auto &it : axis->data_size) {
-        if (min_byte == -1 || min_byte > it.second) {
-          min_byte = it.second;
-        }
+    for (const auto &it : axis->data_size) {
+      if (it.second == 0) {
+        continue;
+      }
+      if (min_byte == -1 || min_byte > it.second) {
+        min_byte = it.second;
       }
     }
-    CHECK_NE(min_byte, 0);
+    min_byte = min_byte == -1 ? 1 : min_byte;
+    CHECK_GT(min_byte, 0);
     axis->l1_constraints.tile_mod_ = CanonicalSimplify(CastIntToExpr(VECTORIZE_BYTE / min_byte));
   }
 }