facebookresearch/faiss
1
0
Fork 0
mirror of https://github.com/facebookresearch/faiss.git synced 2026-10-12 02:13:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix FAISS build with ROCm7 (#4567)

Browse Source 
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
This commit is contained in:
Haoyu Zhang
2025-08-29 08:18:35 -07:00
committed by Facebook GitHub Bot
parent 50b3eb48d2
commit d98ff432b9
1 changed files with 26 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
faiss/gpu/utils/MatrixMult-inl.cuh
+26 -5
View File
@@ -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(