Integrate new pretranspose_b_array with extra fused transpose of B

This patch fuses the transposition taking place in Acl with the transformations done in arm_gemm (called pretranspose_b_array) if the underlying kernel and transform supports it. This should improve start-up time (as it's for constant Rhs matrices) and memory footprint. The transformations in arm_gemm are kernel specific. The Rhs matrix is transformed into certain layouts to improve the performance.

Resolves: COMPMID-6595

Change-Id: Id2932dd966e59f903c279417bebcea83d9a42464
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11144
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 14 insertions, 6 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
index 362a3e30ea..4f732f7d94 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2023 Arm Limited.
+ * Copyright (c) 2017-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -1067,11 +1067,18 @@ public:
         }
     }
 
-    void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override {
-        pretranspose_B_array_part(in_buffer, B, ldb, B_multi_stride, 0, get_B_pretranspose_window_size());
+    // Support for transposed B is a property of the strategy::transpose type
+    bool B_pretranspose_supports_transpose() const override {
+        typename transform_type<strategy, MergeStep && std::is_same<OutputStage, Requantize32>::value>::type transforms;
+
+        return transforms.PrepareB_supports_transpose();
+    }
+
+    void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, const bool transposed) override {
+        pretranspose_B_array_part(in_buffer, B, ldb, B_multi_stride, transposed, 0, get_B_pretranspose_window_size());
     }
 
-    void pretranspose_B_array_part(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, size_t start, size_t end) override {
+    void pretranspose_B_array_part(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, const bool transposed, size_t start, size_t end) override {
         // Perform column sums etc as part of the last block.
         if (end >= get_B_pretranspose_window_size()) {
             requantize_bias(in_buffer, B, ldb, B_multi_stride);
@@ -1134,7 +1141,8 @@ public:
                         strat.transforms.PrepareB(buffer, B + (current.multi() * B_multi_stride), ldb,
                                                   x0, xmax,
                                                   (k_section_base * _Ksize) + k_offset,               // K starting point - compute row to read based on our section and the true section length.
-                                                  (k_section_base * _Ksize) + k_offset + k_length);   // K end point - starting point plus length computed above.
+                                                  (k_section_base * _Ksize) + k_offset + k_length,    // K end point - starting point plus length computed above.
+                                                  transposed);
 
                         // We need to modify our position based on the ROUNDED version of what we just did.
                         unsigned int padded_length = roundup(k_length, strategy::k_unroll());
@@ -1149,7 +1157,7 @@ public:
                 // In the single K section case, can process the whole lot in one go.
                 // Caution: 'blockwalker::kmax()' rounds up, so clamp to valid _Ksize.
                 strat.transforms.PrepareB(buffer, B + (current.multi() * B_multi_stride), ldb,
-                                          current.x0(), current.xmax(), current.k0(), std::min(current.kmax(), _Ksize));
+                                          current.x0(), current.xmax(), current.k0(), std::min(current.kmax(), _Ksize), transposed);
                 buffer += roundup(current.xmax() - current.x0(), strategy::out_width()) * roundup(current.kmax() - current.k0(), strategy::k_unroll());
             }