Faiss GPU: bfloat16 brute-force kNN support (facebookresearch#4018)

Summary:


This diff adds support for bfloat16 vector/query data types with the GPU brute-force k-nearest neighbor function (`bfKnn`).

The change is largely just plumbing the new data type through the template hierarchy (so distances can be computed in bfloat16).

Of note, by design, all final distance results are produced in float32 regardless of input data type (float32, float16, bfloat16). This is because the true nearest neighbors in many data sets can often differ by only ~1000 float32 ULPs in terms of distance which will result in possible false equivalency. This seems to be one area where lossy compression/quantization thoughout does not work as well (and is also why `CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION` is set in `StandardGpuResources.cpp`. However, given that there is native bf16 x bf16 = fp32 tensor core support on Ampere+ architectures, the matrix multiplication itself should 

WARNING: The one thing this diff does not yet handle properly is header inclusion / compilation for GPUs older than Ampere. This will need to be fixed before landing (so that compiling with an older CUDA SDK or compiling for the Volta architecture will simply error out at runtime properly with lack of support, instead of failing to compile (?)

Differential Revision: D65459723

faiss/gpu/GpuDistance.cu

@@ -30,6 +30,7 @@
 #include <faiss/gpu/utils/ConversionOperators.cuh>
 #include <faiss/gpu/utils/CopyUtils.cuh>
 #include <faiss/gpu/utils/DeviceTensor.cuh>
+#include <faiss/gpu/utils/Float16.cuh>
 
 #if defined USE_NVIDIA_RAFT
 #include <faiss/gpu/utils/RaftUtils.h>
@@ -242,7 +243,7 @@ void bfKnn(GpuResourcesProvider* prov, const GpuDistanceParams& args) {
     FAISS_THROW_IF_NOT_MSG(
             args.vectorType == args.queryType,
             "limitation: both vectorType and queryType must currently "
-            "be the same (F32 or F16");
+            "be the same (F32 / F16 / BF16");
 
 #if defined USE_NVIDIA_RAFT
     // Note: For now, RAFT bfknn requires queries and vectors to be same layout
@@ -374,6 +375,17 @@ void bfKnn(GpuResourcesProvider* prov, const GpuDistanceParams& args) {
         bfKnnConvert<float>(prov, args);
     } else if (args.vectorType == DistanceDataType::F16) {
         bfKnnConvert<half>(prov, args);
+    } else if (args.vectorType == DistanceDataType::BF16) {
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+        if (prov->getResources()->supportsBFloat16CurrentDevice()) {
+          bfKnnConvert<__nv_bfloat16>(prov, args);
+        } else {
+            FAISS_THROW_MSG("not compiled with bfloat16 support");
+        }
+#else
+        FAISS_THROW_MSG("no AMD bfloat16 support");
+#endif
     } else {
         FAISS_THROW_MSG("unknown vectorType");
     }
@@ -440,8 +452,10 @@ void bfKnn_single_query_shard(
             args.k > 0,
             "bfKnn_tiling: tiling vectors is only supported for k > 0");
     size_t distance_size = args.vectorType == DistanceDataType::F32 ? 4
-            : args.vectorType == DistanceDataType::F16              ? 2
-                                                                    : 0;
+            : (args.vectorType == DistanceDataType::F16 ||
+               args.vectorType == DistanceDataType::BF16)
+            ? 2
+            : 0;
     FAISS_THROW_IF_NOT_MSG(
             distance_size > 0, "bfKnn_tiling: unknown vectorType");
     size_t shard_size = vectorsMemoryLimit / (args.dims * distance_size);
@@ -498,8 +512,10 @@ void bfKnn_tiling(
             args.k > 0,
             "bfKnn_tiling: tiling queries is only supported for k > 0");
     size_t distance_size = args.queryType == DistanceDataType::F32 ? 4
-            : args.queryType == DistanceDataType::F16              ? 2
-                                                                   : 0;
+            : (args.queryType == DistanceDataType::F16 ||
+               args.queryType == DistanceDataType::BF16)
+            ? 2
+            : 0;
     FAISS_THROW_IF_NOT_MSG(
             distance_size > 0, "bfKnn_tiling: unknown queryType");
     size_t label_size = args.outIndicesType == IndicesDataType::I64 ? 8

faiss/gpu/GpuDistance.h

@@ -19,6 +19,7 @@ class GpuResourcesProvider;
 enum class DistanceDataType {
     F32 = 1,
     F16,
+    BF16,
 };
 
 // Scalar type of the indices data

faiss/gpu/GpuResources.cpp

@@ -161,6 +161,10 @@ GpuMemoryReservation::~GpuMemoryReservation() {
 
 GpuResources::~GpuResources() = default;
 
+bool GpuResources::supportsBFloat16CurrentDevice() {
+    return supportsBFloat16(getCurrentDevice());
+}
+
 cublasHandle_t GpuResources::getBlasHandleCurrentDevice() {
     return getBlasHandle(getCurrentDevice());
 }

faiss/gpu/GpuResources.h

@@ -205,6 +205,9 @@ class GpuResources {
     /// of demand
     virtual void initializeForDevice(int device) = 0;
 
+    /// Does the given GPU support bfloat16?
+    virtual bool supportsBFloat16(int device) = 0;
+
     /// Returns the cuBLAS handle that we use for the given device
     virtual cublasHandle_t getBlasHandle(int device) = 0;
 
@@ -252,6 +255,9 @@ class GpuResources {
     /// Functions provided by default
     ///
 
+    /// Does the current GPU support bfloat16?
+    bool supportsBFloat16CurrentDevice();
+
     /// Calls getBlasHandle with the current device
     cublasHandle_t getBlasHandleCurrentDevice();
 

faiss/gpu/StandardGpuResources.cpp

@@ -202,6 +202,13 @@ size_t StandardGpuResourcesImpl::getDefaultTempMemForGPU(
     return requested;
 }
 
+/// Does the given GPU support bfloat16?
+bool StandardGpuResourcesImpl::supportsBFloat16(int device) {
+    initializeForDevice(device);
+    auto& prop = getDeviceProperties(device);
+    return prop.major >= 8;
+}
+
 void StandardGpuResourcesImpl::noTempMemory() {
     setTempMemory(0);
 }
@@ -687,6 +694,14 @@ std::shared_ptr<GpuResources> StandardGpuResources::getResources() {
     return res_;
 }
 
+bool StandardGpuResources::supportsBFloat16(int device) {
+    return res_->supportsBFloat16(device);
+}
+
+bool StandardGpuResources::supportsBFloat16CurrentDevice() {
+    return res_->supportsBFloat16CurrentDevice();
+}
+
 void StandardGpuResources::noTempMemory() {
     res_->noTempMemory();
 }

faiss/gpu/StandardGpuResources.h

@@ -48,6 +48,9 @@ class StandardGpuResourcesImpl : public GpuResources {
 
     ~StandardGpuResourcesImpl() override;
 
+    /// Does the given GPU support bfloat16?
+    bool supportsBFloat16(int device) override;
+
     /// Disable allocation of temporary memory; all temporary memory
     /// requests will call cudaMalloc / cudaFree at the point of use
     void noTempMemory();
@@ -199,6 +202,12 @@ class StandardGpuResources : public GpuResourcesProvider {
 
     std::shared_ptr<GpuResources> getResources() override;
 
+    /// Whether or not the given device supports native bfloat16 arithmetic
+    bool supportsBFloat16(int device);
+
+    /// Whether or not the current device supports native bfloat16 arithmetic
+    bool supportsBFloat16CurrentDevice();
+
     /// Disable allocation of temporary memory; all temporary memory
     /// requests will call cudaMalloc / cudaFree at the point of use
     void noTempMemory();

faiss/gpu/impl/Distance.cu

@@ -504,6 +504,30 @@ void runAllPairwiseL2Distance(
             outDistances);
 }
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runAllPairwiseL2Distance(
+        GpuResources* res,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<float, 1, true>* vectorNorms,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        Tensor<float, 2, true>& outDistances) {
+    runAllPairwiseDistance<__nv_bfloat16>(
+            true,
+            res,
+            stream,
+            vectors,
+            vectorsRowMajor,
+            vectorNorms,
+            queries,
+            queriesRowMajor,
+            outDistances);
+}
+#endif // USE_AMD_ROCM
+
 void runAllPairwiseIPDistance(
         GpuResources* res,
         cudaStream_t stream,
@@ -544,6 +568,29 @@ void runAllPairwiseIPDistance(
             outDistances);
 }
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runAllPairwiseIPDistance(
+        GpuResources* res,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        Tensor<float, 2, true>& outDistances) {
+    runAllPairwiseDistance<__nv_bfloat16>(
+            false,
+            res,
+            stream,
+            vectors,
+            vectorsRowMajor,
+            nullptr,
+            queries,
+            queriesRowMajor,
+            outDistances);
+}
+#endif // USE_AMD_ROCM
+
 void runL2Distance(
         GpuResources* res,
         cudaStream_t stream,
@@ -596,6 +643,35 @@ void runL2Distance(
             ignoreOutDistances);
 }
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runL2Distance(
+        GpuResources* res,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<float, 1, true>* vectorNorms,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        int k,
+        Tensor<float, 2, true>& outDistances,
+        Tensor<idx_t, 2, true>& outIndices,
+        bool ignoreOutDistances) {
+    runL2Distance<__nv_bfloat16>(
+            res,
+            stream,
+            vectors,
+            vectorsRowMajor,
+            vectorNorms,
+            queries,
+            queriesRowMajor,
+            k,
+            outDistances,
+            outIndices,
+            ignoreOutDistances);
+}
+#endif // USE_AMD_ROCM
+
 void runIPDistance(
         GpuResources* res,
         cudaStream_t stream,
@@ -640,5 +716,30 @@ void runIPDistance(
             outIndices);
 }
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runIPDistance(
+        GpuResources* res,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        int k,
+        Tensor<float, 2, true>& outDistances,
+        Tensor<idx_t, 2, true>& outIndices) {
+    runIPDistance<__nv_bfloat16>(
+            res,
+            stream,
+            vectors,
+            vectorsRowMajor,
+            queries,
+            queriesRowMajor,
+            k,
+            outDistances,
+            outIndices);
+}
+#endif // USE_AMD_ROCM
+
 } // namespace gpu
 } // namespace faiss

faiss/gpu/impl/Distance.cuh

@@ -41,6 +41,19 @@ void runAllPairwiseL2Distance(
         bool queriesRowMajor,
         Tensor<float, 2, true>& outDistances);
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runAllPairwiseL2Distance(
+        GpuResources* res,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<float, 1, true>* vectorNorms,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        Tensor<float, 2, true>& outDistances);
+#endif // USE_AMD_ROCM
+
 void runAllPairwiseIPDistance(
         GpuResources* res,
         cudaStream_t stream,
@@ -59,6 +72,18 @@ void runAllPairwiseIPDistance(
         bool queriesRowMajor,
         Tensor<float, 2, true>& outDistances);
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runAllPairwiseIPDistance(
+        GpuResources* res,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        Tensor<float, 2, true>& outDistances);
+#endif // USE_AMD_ROCM
+
 /// Calculates brute-force L2 distance between `vectors` and
 /// `queries`, returning the k closest results seen
 void runL2Distance(
@@ -91,6 +116,22 @@ void runL2Distance(
         Tensor<idx_t, 2, true>& outIndices,
         bool ignoreOutDistances = false);
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runL2Distance(
+        GpuResources* resources,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<float, 1, true>* vectorNorms,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        int k,
+        Tensor<float, 2, true>& outDistances,
+        Tensor<idx_t, 2, true>& outIndices,
+        bool ignoreOutDistances = false);
+#endif // USE_AMD_ROCM
+
 /// Calculates brute-force inner product distance between `vectors`
 /// and `queries`, returning the k closest results seen
 void runIPDistance(
@@ -115,6 +156,20 @@ void runIPDistance(
         Tensor<float, 2, true>& outDistances,
         Tensor<idx_t, 2, true>& outIndices);
 
+// no bf16 support for AMD
+#ifndef USE_AMD_ROCM
+void runIPDistance(
+        GpuResources* resources,
+        cudaStream_t stream,
+        Tensor<__nv_bfloat16, 2, true>& vectors,
+        bool vectorsRowMajor,
+        Tensor<__nv_bfloat16, 2, true>& queries,
+        bool queriesRowMajor,
+        int k,
+        Tensor<float, 2, true>& outDistances,
+        Tensor<idx_t, 2, true>& outIndices);
+#endif // USE_AMD_ROCM
+
 //
 // General distance implementation, assumes that all arguments are on the
 // device. This is the top-level internal distance function to call to dispatch