From d98ff432b98a4d7bea770c0cc221d8a0dab2e4a9 Mon Sep 17 00:00:00 2001
From: Haoyu Zhang <haoyuz@meta.com>
Date: Fri, 29 Aug 2025 08:18:35 -0700
Subject: [PATCH] Fix FAISS build with ROCm7 (#4567)

Summary:
Pull Request resolved: https://github.com/facebookresearch/faiss/pull/4567

D80101131 enabled FAISS library to build with AMD GPUs. However, certain libraries are written using previous version of ROCm / hipBLAS libraries and fails to build with ROCm7 (hipBLAS 3).

This library is a dependency of MVAI (minimal_viable_ai) so the build is currently broken for MVAI package with ROCm7.

New hipBLAS has API renamed that landed with ROCm 7 (hipBLAS 3.x). Specifically, `hipblasDatatype_t` was removed and replaced by two separate types:

  - `hipDataType` for A/B/C/scale datatypes
  - `hipblasComputeType_t` for the GEMM compute type

To make sure it build with both old and new versions, we need to add additional compile-time check to pick the right APIs.

Reviewed By: leitian

Differential Revision: D81293871

fbshipit-source-id: 755758ef485a74b35651412c5dbd3bc728145a07
---
 faiss/gpu/utils/MatrixMult-inl.cuh | 31 +++++++++++++++++++++++++-----
 1 file changed, 26 insertions(+), 5 deletions(-)

diff --git a/faiss/gpu/utils/MatrixMult-inl.cuh b/faiss/gpu/utils/MatrixMult-inl.cuh
index b28ea56da..a0ef20f6c 100644
--- a/faiss/gpu/utils/MatrixMult-inl.cuh
+++ b/faiss/gpu/utils/MatrixMult-inl.cuh
@@ -22,19 +22,26 @@ struct GetCudaType;
 
 #ifdef USE_AMD_ROCM
 
+#if ((hipblasVersionMajor == 2 && defined(HIPBLAS_V2)) || \
+     hipblasVersionMajor >= 3)
+using hipDataTypeCompat = hipDataType;
+#else
+using hipDataTypeCompat = hipblasDatatype_t;
+#endif
+
 template <>
 struct GetCudaType<float> {
-    static constexpr hipblasDatatype_t Type = HIPBLAS_R_32F;
+    static constexpr hipDataTypeCompat Type = HIPBLAS_R_32F;
 };
 
 template <>
 struct GetCudaType<half> {
-    static constexpr hipblasDatatype_t Type = HIPBLAS_R_16F;
+    static constexpr hipDataTypeCompat Type = HIPBLAS_R_16F;
 };
 
 template <>
 struct GetCudaType<__hip_bfloat16> {
-    static constexpr hipblasDatatype_t Type = HIPBLAS_R_16B;
+    static constexpr hipDataTypeCompat Type = HIPBLAS_R_16B;
 };
 
 #else
@@ -76,6 +83,13 @@ cublasStatus_t rawGemm(
     auto cBT = GetCudaType<BT>::Type;
 
 #ifdef USE_AMD_ROCM
+#if ((hipblasVersionMajor == 2 && defined(HIPBLAS_V2)) || \
+     hipblasVersionMajor >= 3)
+    auto computeType = HIPBLAS_COMPUTE_32F;
+#else
+    auto computeType = HIPBLAS_R_32F;
+#endif
+
     return hipblasGemmEx(
             handle,
             transa,
@@ -94,7 +108,7 @@ cublasStatus_t rawGemm(
             C,
             HIPBLAS_R_32F,
             ldc,
-            HIPBLAS_R_32F,
+            computeType,
             HIPBLAS_GEMM_DEFAULT);
 #else
 
@@ -177,6 +191,13 @@ cublasStatus_t rawBatchGemm(
 
     // Always accumulate in f32
 #ifdef USE_AMD_ROCM
+#if ((hipblasVersionMajor == 2 && defined(HIPBLAS_V2)) || \
+     hipblasVersionMajor >= 3)
+    auto computeType = HIPBLAS_COMPUTE_32F;
+#else
+    auto computeType = HIPBLAS_R_32F;
+#endif
+
     return hipblasGemmStridedBatchedEx(
             handle,
             transa,
@@ -199,7 +220,7 @@ cublasStatus_t rawBatchGemm(
             ldc,
             strideC,
             batchCount,
-            HIPBLAS_R_32F,
+            computeType,
             HIPBLAS_GEMM_DEFAULT);
 #else
     return cublasGemmStridedBatchedEx(