commit 499b5bca1a897461d4105ba52e4c766ddb5f564a
author Gunes Bayir <gunes.bayir@arm.com> Fri Apr 26 13:15:05 2024 +0100
committer Gunes Bayir <gunes.bayir@arm.com> Fri Apr 26 13:06:06 2024 +0000
tree 58b462f7a743f9a2769a52aa4c37d7fbfbe2cab5
parent ada3200f5cec0b6a37f898d5d6f8e69395d7bcb1

Disable SME2 Gemm kernel selection in case results needs to be accumulated

SME2 kernels use a different accumulation buffer and destination tensor is not copied to this buffer as initial value, thus causing mismatches. This patch modifies the kernel selection algorithm such that it does not select SME2 kernels if accumulation is required.

Resolves: COMPMID-6995

Change-Id: I82da3cba41729f938a046f26b41b63ff5716c02d
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11500
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
index 290fe87..af0d38e 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
@@ -124,14 +124,14 @@
 {
     GemmMethod::GEMM_HYBRID,
     "sme2_gemv_fp32bf16fp32_dot_16VL",
-    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input && !args._accumulate; },
     nullptr,
     [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_fp32bf16fp32_dot_16VL, float, float>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "sme2_gemv_fp32_mla_16VL",
-    [](const GemmArgs &args) { return args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input; },
+    [](const GemmArgs &args) { return args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input && !args._accumulate; },
     nullptr,
     [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_fp32_mla_16VL, float, float>(args); }
 },
@@ -139,7 +139,7 @@
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL",
-    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2(); },
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && !args._accumulate; },
     [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
                                return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL, float, float>(args); }
@@ -148,7 +148,7 @@
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_fp32_mopa_1VLx4VL",
-    [](const GemmArgs &args) { return args._ci->has_sme2(); },
+    [](const GemmArgs &args) { return args._ci->has_sme2() && !args._accumulate; },
     [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
                                return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_fp32_mopa_1VLx4VL, float, float>(args); }
@@ -157,7 +157,7 @@
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL",
-    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2(); },
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && !args._accumulate; },
     [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
                                return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL, float, float>(args); }
@@ -166,7 +166,7 @@
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_fp32_mopa_4VLx1VL",
-    [](const GemmArgs &args) { return args._ci->has_sme2(); },
+    [](const GemmArgs &args) { return args._ci->has_sme2() && !args._accumulate; },
     [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
                                return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_fp32_mopa_4VLx1VL, float, float>(args); }
@@ -175,7 +175,7 @@
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL",
-    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2(); },
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && !args._accumulate; },
     nullptr,
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL, float, float>(args); }
 },
@@ -183,7 +183,7 @@
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_fp32_mopa_2VLx2VL",
-    [](const GemmArgs &args) { return args._ci->has_sme2(); },
+    [](const GemmArgs &args) { return args._ci->has_sme2() && !args._accumulate; },
     nullptr,
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_fp32_mopa_2VLx2VL, float, float>(args); }
 },