Enable premultiplication for depthwise convolution

with fp16 and quantized types

Resolves: COMPMID-6337
Change-Id: I81542e51c9c0329f202ac8452f173b138e51a0f6
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9883
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

14 files changed, 113 insertions, 63 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp
index 3b76e52206..8fef6f8ae0 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp
@@ -66,9 +66,47 @@ namespace depthwise {
 
 namespace
 {
+#if defined(__aarch64__)
+#if defined(ENABLE_FP16_KERNELS) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+  bool prefer_premultiply(const DepthwiseArgs &args) {
+    if ((args.stride_rows != args.stride_cols) || (args.kernel_rows != args.kernel_cols))
+    {
+      return false;
+    }
+
+    unsigned int threshold;
+
+    if (args.stride_rows == 1 && args.kernel_rows == 3)
+    {
+      threshold = 30;
+    }
+    else if (args.stride_rows == 1 && args.kernel_rows == 5)
+    {
+      threshold = 31;
+    }
+    else if (args.stride_rows == 2 && args.kernel_rows == 3)
+    {
+      threshold = 11;
+    }
+    else if (args.stride_rows == 2 && args.kernel_rows == 5)
+    {
+      threshold = 19;
+    } else
+    {
+      return false;
+    }
+
+    return args.channel_multiplier <= threshold;
+  }
+
   template <class Strategy>
   unsigned int cycle_estimate(const DepthwiseArgs &args, const Nothing &)
   {
+    if (args.channel_multiplier > 1 && !prefer_premultiply(args))
+    {
+      return std::numeric_limits<unsigned int>::max();
+    }
+
     // First-pass: compute the number of output pixels which will be computed.
     return arm_gemm::roundup(args.output_rows, Strategy::output_rows) *
            arm_gemm::roundup(args.output_cols, Strategy::output_cols) *
@@ -90,13 +128,18 @@ namespace
           );
   }
 
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+  unsigned int multiplier_cycle_estimate(const DepthwiseArgs &args, const Nothing &)
+  {
+    return prefer_premultiply(args)? std::numeric_limits<unsigned int>::max() : 0;
+  }
+
   unsigned int not_preferred(const DepthwiseArgs &, const Nothing &) __attribute__ ((unused));
   unsigned int not_preferred(const DepthwiseArgs &, const Nothing &)
   {
     return std::numeric_limits<unsigned int>::max();
   }
-#endif  // defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#endif  // defined(ENABLE_FP16_KERNELS) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#endif  // defined(__aarch64__)
 }
 
 static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] = {
@@ -108,7 +151,6 @@ static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] =
     DepthwiseMethod::DEPTHFIRST,
     "sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst",
     constraint(is_supported<sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst>,
-               has_no_channel_multiplier,
                cpu_has_sme2),
     cycle_estimate<sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst>,
     [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * {
@@ -120,7 +162,6 @@ static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] =
     DepthwiseMethod::DEPTHFIRST,
     "sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst",
     constraint(is_supported<sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst>,
-               has_no_channel_multiplier,
                cpu_has_sme2),
     cycle_estimate<sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst>,
     [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * {
@@ -132,7 +173,6 @@ static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] =
     DepthwiseMethod::DEPTHFIRST,
     "sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint(is_supported<sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-              has_no_channel_multiplier,
               cpu_has_sme2),
     cycle_estimate<sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst>,
     [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * {
@@ -144,7 +184,6 @@ static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] =
     DepthwiseMethod::DEPTHFIRST,
     "sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint(is_supported<sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-               has_no_channel_multiplier,
                cpu_has_sme2),
     cycle_estimate<sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst>,
     [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * {
@@ -156,7 +195,6 @@ static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] =
     DepthwiseMethod::DEPTHFIRST,
     "sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint(is_supported<sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-               has_no_channel_multiplier,
                cpu_has_sme2),
     cycle_estimate<sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst>,
     [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * {
@@ -291,7 +329,7 @@ static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] =
     DepthwiseMethod::DEPTHFIRST,
     "a64_fp16_nhwc_generic_with_multiplier_output2x8_mla_depthfirst",
     constraint(cpu_has_fp16, has_channel_multiplier),
-    nullptr,
+    multiplier_cycle_estimate,
     [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * {
       auto kern = new a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst(args.cpu_info);
       auto strat = new GenericDepthfirstMultiplierStrategy<__fp16>(kern, args);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp
index b0f606332b..760328f3ba 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp
@@ -116,7 +116,7 @@ namespace
   {
     if (args.channel_multiplier > 1 && !prefer_premultiply(args))
     {
-      return UINT32_MAX;
+      return std::numeric_limits<unsigned int>::max();
     }
 
     // First-pass: compute the number of output pixels which will be computed.
@@ -154,7 +154,7 @@ namespace
 
   unsigned int multiplier_cycle_estimate(const DepthwiseArgs &args, const Nothing &)
   {
-    return prefer_premultiply(args)? UINT32_MAX : 0;
+    return prefer_premultiply(args)? std::numeric_limits<unsigned int>::max() : 0;
   }
 
   unsigned int not_preferred(const DepthwiseArgs &, const Nothing &)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp
index 2d03183c59..6ecdc36bf0 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -75,6 +75,11 @@ bool qp_weights_are_symmetric(const DepthwiseArgs &, const void *_qp)
   const auto qp = static_cast<const arm_gemm::Requantize32 *>(_qp);
   return qp->b_offset == 0;
 }
+
+uint64_t not_preferred(const DepthwiseArgs &, const Requantize32 &)
+{
+  return std::numeric_limits<uint64_t>::max();
+}
 #endif // defined(__aarch64__)
 }
 
@@ -139,7 +144,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst",
     constraint<Requantize32>(is_supported<sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              qp_weights_are_symmetric,
                              cpu_has_sve2),
@@ -153,7 +157,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst",
     constraint<Requantize32>(is_supported<sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -166,7 +169,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "sve_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -179,7 +181,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "sve_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -192,7 +193,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "sve_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -208,7 +208,7 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
                              qp_has_no_left_shift,
                              has_channel_multiplier,
                              cpu_has_sve2),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
       auto strat = new sve_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<int8_t, int8_t, int8_t, int32_t, false>(strat, args, qp);
@@ -221,7 +221,7 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
                              qp_has_no_left_shift,
                              has_channel_multiplier,
                              cpu_has_sve2),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
       auto strat = new sve_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<int8_t, int8_t, int8_t, int32_t, false>(strat, args, qp);
@@ -232,7 +232,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst",
     constraint<Requantize32>(is_supported<a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_weights_are_symmetric,
                              qp_has_no_left_shift,
                              cpu_has_dot_product),
@@ -246,7 +245,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst",
     constraint<Requantize32>(is_supported<a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_dot_product),
     nullptr,
@@ -259,7 +257,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
@@ -271,7 +268,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
@@ -283,7 +279,6 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
@@ -294,7 +289,7 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
   {
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8q_nhwc_generic_output3x3_mla_depthfirst",
-    constraint<Requantize32>(has_no_channel_multiplier),
+    nullptr,
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
       auto kernel = new a64_s8q_nhwc_generic_output9_mla_depthfirst(args.cpu_info);
@@ -309,7 +304,7 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
                              qp_has_no_left_shift,
                              has_channel_multiplier,
                              cpu_has_dot_product),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
       auto strat = new a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<int8_t, int8_t, int8_t, int32_t, false>(strat, args, qp);
@@ -322,7 +317,7 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
                              qp_has_no_left_shift,
                              has_channel_multiplier,
                              cpu_has_dot_product),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
       auto strat = new a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<int8_t, int8_t, int8_t, int32_t, false>(strat, args, qp);
@@ -332,7 +327,7 @@ static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depth
     DepthwiseMethod::DEPTHFIRST,
     "a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst",
     constraint<Requantize32>(has_channel_multiplier),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<int8_t, int8_t, int8_t> * {
       auto kern = new a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst(args.cpu_info);
       auto strat = new GenericDepthfirstMultiplierStrategy<int8_t>(kern, args);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp
index 9dbd89fb52..236930ee26 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -71,6 +71,16 @@ using arm_gemm::Requantize32;
 namespace arm_conv {
 namespace depthwise {
 
+namespace
+{
+#if defined(__aarch64__)
+uint64_t not_preferred(const DepthwiseArgs &, const Requantize32 &)
+{
+  return std::numeric_limits<uint64_t>::max();
+}
+#endif // defined(__aarch64__)
+}
+
 static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> depthwise_u8q_methods[] = {
 #if defined(__aarch64__)
 #if defined(ARM_COMPUTE_ENABLE_SVE)
@@ -132,7 +142,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -145,7 +154,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -158,7 +166,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -171,7 +178,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -187,7 +193,7 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
                              qp_has_no_left_shift,
                              has_channel_multiplier,
                              cpu_has_sve2),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
       auto strat = new sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<uint8_t, uint8_t, uint8_t, int32_t, false>(strat, args, qp);
@@ -200,7 +206,7 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
                              qp_has_no_left_shift,
                              has_channel_multiplier,
                              cpu_has_sve2),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
       auto strat = new sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<uint8_t, uint8_t, uint8_t, int32_t, false>(strat, args, qp);
@@ -212,7 +218,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     "a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst>,
                              cpu_has_dot_product,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
@@ -225,7 +230,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_zero_a_offset,
                              qp_has_no_left_shift),
     nullptr,
@@ -238,7 +242,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_zero_a_offset,
                              qp_has_no_left_shift),
     nullptr,
@@ -251,7 +254,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_zero_a_offset,
                              qp_has_no_left_shift),
     nullptr,
@@ -265,7 +267,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
@@ -277,7 +278,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
@@ -289,7 +289,6 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
@@ -300,7 +299,7 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
   {
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8q_nhwc_generic_output3x3_mla_depthfirst",
-    constraint<Requantize32>(has_no_channel_multiplier),
+    nullptr,
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
       auto kernel = new a64_u8q_nhwc_generic_output9_mla_depthfirst(args.cpu_info);
@@ -315,7 +314,7 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
                              cpu_has_dot_product,
                              has_channel_multiplier,
                              qp_has_no_left_shift),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
       auto strat = new a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<uint8_t, uint8_t, uint8_t, int32_t, false>(strat, args, qp);
@@ -328,7 +327,7 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
                              cpu_has_dot_product,
                              has_channel_multiplier,
                              qp_has_no_left_shift),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
       auto strat = new a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst(args.cpu_info);
       return new DepthwiseDepthfirstMultiplier<uint8_t, uint8_t, uint8_t, int32_t, false>(strat, args, qp);
@@ -338,7 +337,7 @@ static const DepthwiseImplementation<uint8_t, uint8_t, uint8_t, Requantize32> de
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst",
     constraint<Requantize32>(has_channel_multiplier),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, uint8_t, uint8_t> * {
       auto kern = new a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst(args.cpu_info);
       auto strat = new GenericDepthfirstMultiplierStrategy<uint8_t>(kern, args);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp
index 0665c67fbb..a888958b76 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -58,6 +58,16 @@ using arm_gemm::Requantize32;
 namespace arm_conv {
 namespace depthwise {
 
+namespace
+{
+#if defined(__aarch64__)
+uint64_t not_preferred(const DepthwiseArgs &, const Requantize32 &)
+{
+  return std::numeric_limits<uint64_t>::max();
+}
+#endif // defined(__aarch64__)
+}
+
 static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> depthwise_u8q_methods[] = {
 #if defined(__aarch64__)
 #if defined(ARM_COMPUTE_ENABLE_SVE)
@@ -119,7 +129,6 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -132,7 +141,6 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -145,7 +153,6 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift,
                              cpu_has_sve2),
     nullptr,
@@ -159,7 +166,6 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, int8_t, uint8_t> * {
@@ -171,7 +177,6 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, int8_t, uint8_t> * {
@@ -183,7 +188,6 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst",
     constraint<Requantize32>(is_supported<a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst>,
-                             has_no_channel_multiplier,
                              qp_has_no_left_shift),
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, int8_t, uint8_t> * {
@@ -194,7 +198,7 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
   {
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8s8u8q_nhwc_generic_output3x3_mla_depthfirst",
-    constraint<Requantize32>(has_no_channel_multiplier),
+    nullptr,
     nullptr,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, int8_t, uint8_t> * {
       auto kernel = new a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst(args.cpu_info);
@@ -206,7 +210,7 @@ static const DepthwiseImplementation<uint8_t, int8_t, uint8_t, Requantize32> dep
     DepthwiseMethod::DEPTHFIRST,
     "a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst",
     constraint<Requantize32>(has_channel_multiplier),
-    nullptr,
+    not_preferred,
     [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon<uint8_t, int8_t, uint8_t> * {
       auto kern = new a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst(args.cpu_info);
       auto strat = new GenericDepthfirstMultiplierStrategy<uint8_t, int8_t>(kern, args);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp
index 314f09a0c5..19264c9fce 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp
@@ -42,7 +42,7 @@ size_t interleave_a64_u8q_3x3_dot::get_packed_size(const DepthwiseArgs &args)
 {
   // We store 7 vectors for every <vector_of_ints> of channels.
   const unsigned int n = arm_gemm::roundup(
-    arm_gemm::iceildiv((long unsigned int) args.input_channels,
+    arm_gemm::iceildiv((long unsigned int) args.input_channels * args.channel_multiplier,
                        get_vector_length<int32_t>(arm_gemm::VLType::None)), 4lu
   );
   return n * 7 * get_vector_length<uint8_t>(arm_gemm::VLType::None);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp
index 3a4999296a..5d7b54f235 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp
@@ -42,7 +42,7 @@ size_t interleave_sve_s8q_3x3_dot::get_packed_size(const DepthwiseArgs &args)
 {
   // We store 7 vectors for every <vector_of_ints> of channels.
   const unsigned int n = arm_gemm::roundup(
-    arm_gemm::iceildiv((long unsigned int) args.input_channels,
+    arm_gemm::iceildiv((long unsigned int) args.input_channels * args.channel_multiplier,
                        get_vector_length<int32_t>(arm_gemm::VLType::SVE)), 4lu
   );
   return n * 7 * get_vector_length<int8_t>(arm_gemm::VLType::SVE);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp
index 7c5d3c4904..c3da81448b 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp
@@ -42,7 +42,7 @@ size_t interleave_sve_u8q_3x3_dot::get_packed_size(const DepthwiseArgs &args)
 {
   // We store 7 vectors for every <vector_of_ints> of channels.
   const unsigned int n = arm_gemm::roundup(
-    arm_gemm::iceildiv((long unsigned int) args.input_channels,
+    arm_gemm::iceildiv((long unsigned int) args.input_channels * args.channel_multiplier,
                        get_vector_length<int32_t>(arm_gemm::VLType::SVE)), 4lu
   );
   return n * 7 * get_vector_length<uint8_t>(arm_gemm::VLType::SVE);
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
index fc83aaf5d2..bd2941b3d6 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
@@ -64,7 +64,7 @@ class a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst : public DepthwiseDepthfirst
   ) const override
   {
     interleave_a64_s8q_3x3_dot::pack_parameters(
-      args.input_channels, buffer, reinterpret_cast<const int32_t *>(biases),
+      args.input_channels * args.channel_multiplier, buffer, reinterpret_cast<const int32_t *>(biases),
       reinterpret_cast<const int8_t *>(weights), qp, ld_weight_col, ld_weight_row
     );
   }
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
index bea97a54b6..ed24f8fa3c 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
@@ -64,7 +64,7 @@ class a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst : public DepthwiseDepthfirstS
   ) const override
   {
     interleave_a64_u8q_3x3_dot::pack_parameters(
-      args.input_channels, buffer, reinterpret_cast<const int32_t *>(biases),
+      args.input_channels * args.channel_multiplier, buffer, reinterpret_cast<const int32_t *>(biases),
       reinterpret_cast<const uint8_t *>(weights), qp, ld_weight_col, ld_weight_row
     );
   }
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
index 32ea009e8a..04cf0d4036 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
@@ -64,7 +64,7 @@ class sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst : public DepthwiseDepthfirstS
   ) const override
   {
     interleave_sve_s8q_3x3_dot::pack_parameters(
-      args.input_channels, buffer, reinterpret_cast<const int32_t *>(biases),
+      args.input_channels * args.channel_multiplier, buffer, reinterpret_cast<const int32_t *>(biases),
       reinterpret_cast<const int8_t *>(weights), qp, ld_weight_col, ld_weight_row
     );
   }
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
index 1730574933..6799b10ed9 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
@@ -64,7 +64,7 @@ class sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst : public DepthwiseDepthfirst
   ) const override
   {
     interleave_sve_s8q_3x3_dot::pack_parameters(
-      args.input_channels, buffer, reinterpret_cast<const int32_t *>(biases),
+      args.input_channels * args.channel_multiplier, buffer, reinterpret_cast<const int32_t *>(biases),
       reinterpret_cast<const int8_t *>(weights), qp, ld_weight_col, ld_weight_row
     );
   }
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
index 9432cd7550..6b006e8d51 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
@@ -64,7 +64,7 @@ class sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst : public DepthwiseDepthfirstS
   ) const override
   {
     interleave_sve_u8q_3x3_dot::pack_parameters(
-      args.input_channels, buffer, reinterpret_cast<const int32_t *>(biases),
+      args.input_channels * args.channel_multiplier, buffer, reinterpret_cast<const int32_t *>(biases),
       reinterpret_cast<const uint8_t *>(weights), qp, ld_weight_col, ld_weight_row
     );
   }
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp b/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp
index ad4c821cfb..8a49c775d3 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp
@@ -45,7 +45,9 @@ void do_premultiply_float_6(const float       *in_ptr,
         {
             const float *ip = ip2;
             float       *op = op2;
-            for(unsigned int c = 0; c < input_channels; c += BLOCK_SIZE)
+
+            unsigned int num_blocks = input_channels / BLOCK_SIZE;
+            for(unsigned int c = 0; c < num_blocks; c++)
             {
                 float vals[BLOCK_SIZE];
                 for(unsigned int v = 0; v < BLOCK_SIZE; v++)
@@ -63,6 +65,18 @@ void do_premultiply_float_6(const float       *in_ptr,
                     op += CHANNEL_MULTIPLIER;
                 }
             }
+
+            unsigned int rem = input_channels - num_blocks * BLOCK_SIZE;
+            for(unsigned int c = 0; c < rem; c++)
+            {
+                float val = ip[c];
+                for(unsigned int r = 0; r < CHANNEL_MULTIPLIER; r++)
+                {
+                    op[r] = val;
+                }
+                op += CHANNEL_MULTIPLIER;
+            }
+
             ip2 += ld_col;
             op2 += out_ld_col;
         }