WIP

NVIDIA · Nov 12, 2024 · 21c6e2e · 21c6e2e
1 parent 2dec4a7
commit 21c6e2e
Show file tree

Hide file tree

Showing 4 changed files with 87 additions and 23 deletions.
diff --git a/csrc/ir/internal_nodes.h b/csrc/ir/internal_nodes.h
@@ -2579,40 +2579,40 @@ class SdpaBwdOp : public Expr {
   std::string toString(int indent_size = 0) const override;
   std::string toInlineString(int indent_size = 0) const override;
 
-  Val* grad_query() const {
-    return output(0);
+  TensorView* grad_query() const {
+    return output(0)->as<TensorView>();
   }
 
-  Val* grad_key() const {
-    return output(1);
+  TensorView* grad_key() const {
+    return output(1)->as<TensorView>();
   }
 
-  Val* grad_value() const {
-    return output(2);
+  TensorView* grad_value() const {
+    return output(2)->as<TensorView>();
   }
 
-  Val* grad_attn() const {
-    return input(0);
+  TensorView* grad_attn() const {
+    return input(0)->as<TensorView>();
   }
 
-  Val* query() const {
-    return input(1);
+  TensorView* query() const {
+    return input(1)->as<TensorView>();
   }
 
-  Val* key() const {
-    return input(2);
+  TensorView* key() const {
+    return input(2)->as<TensorView>();
   }
 
-  Val* value() const {
-    return input(3);
+  TensorView* value() const {
+    return input(3)->as<TensorView>();
   }
 
-  Val* attn_out() const {
-    return input(4);
+  TensorView* attn_out() const {
+    return input(4)->as<TensorView>();
   }
 
-  Val* logsumexp() const {
-    return input(5);
+  TensorView* logsumexp() const {
+    return input(5)->as<TensorView>();
   }
 
   Val* dropout_p() const {

diff --git a/csrc/ir/utils.h b/csrc/ir/utils.h
@@ -658,6 +658,25 @@ std::optional<std::vector<int64_t>> computePermutation(
   return permutation;
 }
 
+template <typename T>
+std::vector<T> applyPermutation(
+    const std::vector<T>& in,
+    const std::vector<int64_t>& permutation) {
+  NVF_CHECK(in.size() == permutation.size());
+
+  std::vector<int64_t> identity(permutation.size());
+  std::iota(identity.begin(), identity.end(), 0);
+  NVF_CHECK(std::is_permutation(
+      permutation.begin(), permutation.end(), identity.begin()));
+
+  std::vector<T> out;
+  out.reserve(permutation.size());
+  for (auto i : permutation) {
+    out.push_back(in[i]);
+  }
+  return out;
+}
+
 bool hasTrivialAllocationDomain(const TensorView* tv);
 
 // Returns true if all expr outputs should be mapped unconditionally

diff --git a/csrc/preseg_passes/allocation_order_inference.cpp b/csrc/preseg_passes/allocation_order_inference.cpp
@@ -336,6 +336,18 @@ void inferenceAllocationOrder(
   }
 }
 
+namespace {
+void propagateAllocation(TensorView* in, TensorView* out) {
+  auto in_order = ir_utils::computePermutation(
+      in->getLogicalDomain(), in->getMaybeAllocationDomain());
+  if (!in_order.has_value()) {
+    return;
+  }
+  out->setAllocationDomain(
+      ir_utils::applyPermutation(out->getLogicalDomain(), *in_order), true);
+}
+} // namespace
+
 void AllocationDomainPass::runPass(Fusion* fusion) {
   // mark input TensorViews as propagation sources
   auto input_tvs = ir_utils::filterByType<TensorView>(fusion->inputs());
@@ -351,15 +363,48 @@ void AllocationDomainPass::runPass(Fusion* fusion) {
   // hint, but they should respect semantic requirement.
   // see issue: https://github.com/NVIDIA/Fuser/pull/2425
   for (TensorView* output : output_tvs) {
-    if (output->isDefinitionType<LinearOp>() ||
-        output->isDefinitionType<MatmulOp>() ||
-        output->isDefinitionType<MmaOp>()) {
-      continue;
+    if (Expr* def = output->definition()) {
+      if (def->isOneOf<LinearOp, SdpaFwdOp, SdpaBwdOp, MatmulOp, MmaOp>()) {
+        continue;
+      }
     }
     dsts.push_back(output);
   }
   // propagate allocation domain from sources to destinations
   inferenceAllocationOrder(fusion, srcs, dsts);
+
+  for (Expr* e : fusion->exprs()) {
+    if (auto* sdpa_fwd = dynamic_cast<SdpaFwdOp*>(e)) {
+      std::optional<std::vector<int64_t>> out_order = {};
+      for (TensorView* in :
+           {sdpa_fwd->query(), sdpa_fwd->key(), sdpa_fwd->value()}) {
+        auto in_order = ir_utils::computePermutation(
+            in->getLogicalDomain(), in->getMaybeAllocationDomain());
+        if (!in_order.has_value()) {
+          out_order = std::nullopt;
+          break;
+        }
+        if ((*out_order).empty()) {
+          out_order = *in_order;
+          continue;
+        }
+        if (*out_order != *in_order) {
+          out_order = std::nullopt;
+          break;
+        }
+      }
+      if (out_order.has_value()) {
+        TensorView* out = sdpa_fwd->attn_out();
+        out->setAllocationDomain(
+            ir_utils::applyPermutation(out->getLogicalDomain(), *out_order),
+            true);
+      }
+    } else if (auto* sdpa_bwd = dynamic_cast<SdpaBwdOp*>(e)) {
+      propagateAllocation(sdpa_bwd->query(), sdpa_bwd->grad_query());
+      propagateAllocation(sdpa_bwd->key(), sdpa_bwd->grad_key());
+      propagateAllocation(sdpa_bwd->value(), sdpa_bwd->grad_value());
+    }
+  }
 }
 
 } // namespace nvfuser::preseg_passes
diff --git a/csrc/runtime/fusion_kernel_runtime.cpp b/csrc/runtime/fusion_kernel_runtime.cpp
@@ -84,7 +84,7 @@ FusionKernelRuntime::FusionKernelRuntime(
     if (!communicator.is_available() || communicator.local_rank() == 0) {
       debug() << "Fusion IR after pre-segmenter optimization passes:"
               << std::endl;
-      fusion->printMath();
+      fusion->print();
     }
   }