diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-02-23 23:41:40 +0000 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-02-24 10:08:28 +0000 |
| commit | 6f45cf77e784d1bfc2a57119671937a3c064eac2 (patch) | |
| tree | c97e0ea611fe6ef3d114ba837ce3844dc3247038 | |
| parent | 85e16c2431f1adce1dd166d52caece9066546895 (diff) | |
| download | ComputeLibrary-6f45cf77e784d1bfc2a57119671937a3c064eac2.tar.gz | |
Account Ksections correctly for interleaved and hybrid_indirect gemm kernels
Resolves: COMPMID-4131
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I6ab15d5dd2e89ccda6acf06e1bef7fd85e31ce54
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5169
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp | 4 | ||||
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp | 4 |
2 files changed, 4 insertions, 4 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp index d06c2183d1..5d5f21507f 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp @@ -577,7 +577,7 @@ public: // Note: Current hybrid kernels don't actually round up height (they // have paths for each possible height). Might need to make this // configurable in future. - uint64_t total_macs = static_cast<uint64_t>(args._nbatches) * args._nmulti * args._Msize * roundup(args._Nsize, strategy::out_width()) * roundup(args._Ksize, strategy::k_unroll()); + uint64_t total_macs = static_cast<uint64_t>(args._nbatches) * args._nmulti * args._Msize * roundup(args._Nsize, strategy::out_width()) * get_ktotal(args); float mac_cycles = static_cast<float>(total_macs) / params.kernel_macs_cycle; @@ -596,7 +596,7 @@ public: const Requantize32 *qp = reinterpret_cast<const Requantize32 *>(&os); // Row sums: need to consider each value in A (batch * multi * M * K)... - uint64_t rowsum_bytes = static_cast<uint64_t>(args._nbatches) * args._nmulti * args._Msize * roundup(args._Ksize, strategy::k_unroll()); + uint64_t rowsum_bytes = static_cast<uint64_t>(args._nbatches) * args._nmulti * args._Msize * get_ktotal(args); // ... but row sums are skipped if B offset==0. if (qp->b_offset == 0) { diff --git a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp index 92c1086a5f..7f870b83d7 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp @@ -1022,8 +1022,8 @@ public: static uint64_t estimate_cycles(const GemmArgs &args, const PerformanceParameters ¶ms) { unsigned int k_blocks = iceildiv(args._Ksize, get_k_block_size(args)); - uint64_t total_macs = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * roundup(args._Nsize, strategy::out_width()) * roundup(args._Ksize, strategy::k_unroll()); - uint64_t prepare_bytes = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * roundup(args._Ksize, strategy::k_unroll()) * sizeof(Toi); + uint64_t total_macs = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * roundup(args._Nsize, strategy::out_width()) * get_ktotal(args); + uint64_t prepare_bytes = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * get_ktotal(args) * sizeof(Toi); uint64_t merge_bytes = static_cast<uint16_t>(args._nbatches) * args._nmulti * k_blocks * roundup(args._Msize, strategy::out_height()) * roundup(args._Nsize, strategy::out_width()) * sizeof(Tr); float mac_cycles = static_cast<float>(total_macs) / params.kernel_macs_cycle; |