From: Debadri Samaddar <s.debadri@samsung.com>
Date: Thu, 31 Aug 2023 09:09:20 +0000 (+0530)
Subject: [blas/neon] Optimized SGEMM when both inputs are transposed
X-Git-Tag: accepted/tizen/8.0/unified/20231005.093407~32
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=b2520ac23338e63dc50f9defac805692d8c68284;p=platform%2Fcore%2Fml%2Fnntrainer.git

[blas/neon] Optimized SGEMM when both inputs are transposed

Optimized sgemm stub when both A and B are transposed

**Self evaluation:**
1. Build test:	 [X]Passed [ ]Failed [ ]Skipped
2. Run test:	 [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Debadri Samaddar <s.debadri@samsung.com>
---

diff --git a/nntrainer/tensor/blas_neon.cpp b/nntrainer/tensor/blas_neon.cpp
index 3249ab2..4110546 100644
--- a/nntrainer/tensor/blas_neon.cpp
+++ b/nntrainer/tensor/blas_neon.cpp
@@ -810,17 +810,20 @@ void sgemm_neon_fp16(const __fp16 *A, const __fp16 *B, __fp16 *C, uint32_t M,
       }
     }
   } else { // TransA && TransB
-    for (unsigned int m = 0; m < M; m++) {
-      for (unsigned int n = 0; n < N; n++) {
-        __fp16 sum = 0;
-        for (int k = 0; k < K; k++) {
-          __fp16 a = A[k * M + m];
-          __fp16 b = B[n * K + k];
-          sum += a * b;
+    __fp16 vals[8];
+    for (unsigned int n = 0; n < N; n++) {
+      for (unsigned int k = 0; k < K; k++) {
+
+        __fp16 b = alpha * B[n * K + k];
+        for (unsigned int m = 0; m < M; m += 8) {
+          float16x8_t a = vld1q_f16(&A[k * M + m]);
+          a = vmulq_n_f16(a, b);
+          vst1q_f16(vals, a);
+
+          // calculations for all M values
+          for (unsigned int idx = m; idx < m + 8; idx++)
+            C[idx * N + n] += vals[idx - m];
         }
-
-        sum = alpha * sum;
-        C[m * N + n] += sum;
       }
     }
   }