[SE] KernelSpec return best PTX

Summary:
Before, the kernel spec would only return PTX for exactly the requested
compute capability. With this patch it will now return the PTX with the
largest compute capability that does not exceed that requested compute
capability.

Reviewers: jlebar

Subscribers: jprice, jlebar, parallel_libs-commits

Differential Revision: https://reviews.llvm.org/D24531

3 files changed, 15 insertions, 12 deletions

diff --git a/parallel-libs/streamexecutor/include/streamexecutor/KernelSpec.h b/parallel-libs/streamexecutor/include/streamexecutor/KernelSpec.h
index caf6f1bdc4f..a6a293001ec 100644
--- a/parallel-libs/streamexecutor/include/streamexecutor/KernelSpec.h
+++ b/parallel-libs/streamexecutor/include/streamexecutor/KernelSpec.h
@@ -121,12 +121,11 @@ public:
       llvm::StringRef KernelName,
       const llvm::ArrayRef<CUDAPTXInMemorySpec::PTXSpec> SpecList);
 
-  /// Returns a pointer to the PTX code for the requested compute capability.
+  /// Returns a pointer to the PTX code for the greatest compute capability not
+  /// exceeding the requested compute capability.
   ///
-  /// Returns nullptr on failed lookup (if the requested compute capability is
-  /// not available). Matches exactly the specified compute capability. Doesn't
-  /// try to do anything smart like finding the next best compute capability if
-  /// the specified capability cannot be found.
+  /// Returns nullptr on failed lookup (if the requested version is not
+  /// available and no lower versions are available).
   const char *getCode(int ComputeCapabilityMajor,
                       int ComputeCapabilityMinor) const;
 
diff --git a/parallel-libs/streamexecutor/lib/KernelSpec.cpp b/parallel-libs/streamexecutor/lib/KernelSpec.cpp
index b5753a489d1..951ea8fc41c 100644
--- a/parallel-libs/streamexecutor/lib/KernelSpec.cpp
+++ b/parallel-libs/streamexecutor/lib/KernelSpec.cpp
@@ -31,12 +31,13 @@ CUDAPTXInMemorySpec::CUDAPTXInMemorySpec(
 
 const char *CUDAPTXInMemorySpec::getCode(int ComputeCapabilityMajor,
                                          int ComputeCapabilityMinor) const {
-  auto PTXIter =
-      PTXByComputeCapability.find(CUDAPTXInMemorySpec::ComputeCapability{
+  auto Iterator =
+      PTXByComputeCapability.upper_bound(CUDAPTXInMemorySpec::ComputeCapability{
           ComputeCapabilityMajor, ComputeCapabilityMinor});
-  if (PTXIter == PTXByComputeCapability.end())
+  if (Iterator == PTXByComputeCapability.begin())
     return nullptr;
-  return PTXIter->second;
+  --Iterator;
+  return Iterator->second;
 }
 
 CUDAFatbinInMemorySpec::CUDAFatbinInMemorySpec(llvm::StringRef KernelName,
diff --git a/parallel-libs/streamexecutor/unittests/CoreTests/KernelSpecTest.cpp b/parallel-libs/streamexecutor/unittests/CoreTests/KernelSpecTest.cpp
index fc9eb549968..486a3504091 100644
--- a/parallel-libs/streamexecutor/unittests/CoreTests/KernelSpecTest.cpp
+++ b/parallel-libs/streamexecutor/unittests/CoreTests/KernelSpecTest.cpp
@@ -30,8 +30,9 @@ TEST(CUDAPTXInMemorySpec, SingleComputeCapability) {
   const char *PTXCodeString = "Dummy PTX code";
   se::CUDAPTXInMemorySpec Spec("KernelName", {{{1, 0}, PTXCodeString}});
   EXPECT_EQ("KernelName", Spec.getKernelName());
+  EXPECT_EQ(nullptr, Spec.getCode(0, 5));
   EXPECT_EQ(PTXCodeString, Spec.getCode(1, 0));
-  EXPECT_EQ(nullptr, Spec.getCode(2, 0));
+  EXPECT_EQ(PTXCodeString, Spec.getCode(2, 0));
 }
 
 TEST(CUDAPTXInMemorySpec, TwoComputeCapabilities) {
@@ -40,9 +41,10 @@ TEST(CUDAPTXInMemorySpec, TwoComputeCapabilities) {
   se::CUDAPTXInMemorySpec Spec(
       "KernelName", {{{1, 0}, PTXCodeString10}, {{3, 0}, PTXCodeString30}});
   EXPECT_EQ("KernelName", Spec.getKernelName());
+  EXPECT_EQ(nullptr, Spec.getCode(0, 5));
   EXPECT_EQ(PTXCodeString10, Spec.getCode(1, 0));
   EXPECT_EQ(PTXCodeString30, Spec.getCode(3, 0));
-  EXPECT_EQ(nullptr, Spec.getCode(2, 0));
+  EXPECT_EQ(PTXCodeString10, Spec.getCode(2, 0));
 }
 
 TEST(CUDAFatbinInMemorySpec, BasicUsage) {
@@ -89,8 +91,9 @@ TEST(MultiKernelLoaderSpec, Registration) {
   EXPECT_TRUE(MultiSpec.hasOpenCLTextInMemory());
 
   EXPECT_EQ(KernelName, MultiSpec.getCUDAPTXInMemory().getKernelName());
+  EXPECT_EQ(nullptr, MultiSpec.getCUDAPTXInMemory().getCode(0, 5));
   EXPECT_EQ(PTXCodeString, MultiSpec.getCUDAPTXInMemory().getCode(1, 0));
-  EXPECT_EQ(nullptr, MultiSpec.getCUDAPTXInMemory().getCode(2, 0));
+  EXPECT_EQ(PTXCodeString, MultiSpec.getCUDAPTXInMemory().getCode(2, 0));
 
   EXPECT_EQ(KernelName, MultiSpec.getCUDAFatbinInMemory().getKernelName());
   EXPECT_EQ(FatbinBytes, MultiSpec.getCUDAFatbinInMemory().getBytes());