cpu: aarch64: Add key_generic_acc to booking scratchpad for dst_acc

Author: annop-w

src/common/memory_tracking.hpp

@@ -227,6 +227,7 @@ enum {
     key_gemm_blocked_a,
     key_gemm_blocked_b,
     key_gemm_accumulator,
+    key_generic_acc,
     key_gnorm_cvt,
     key_gnorm_reduction,
     key_gnorm_tmp_mean,

src/cpu/aarch64/acl_convolution_utils.hpp

@@ -104,7 +104,7 @@ status_t execute_forward_conv_acl(
     // If we have an unfused sum post op, put the result in a scratchpad tensor.
     // Result will be summed to the dst during acl_post_ops.execute
     auto dst_base = use_dst_acc_for_sum
-            ? scratchpad.get<void>(memory_tracking::names::key_none)
+            ? scratchpad.get<void>(memory_tracking::names::key_generic_acc)
             : CTX_OUT_MEM(dst_data_t *, DNNL_ARG_DST);
     acl_conv_obj.dst_tensor.allocator()->import_memory(dst_base);
 

src/cpu/aarch64/acl_deconvolution.cpp

@@ -37,7 +37,7 @@ status_t acl_deconvolution_fwd_t::execute_forward(const exec_ctx_t &ctx) const {
     // If we have an unfused sum post op, put the result in a scratchpad tensor.
     // Result will be summed to the dst during acl_post_ops.execute
     auto dst_base = use_dst_acc_for_sum
-            ? scratchpad.get<void>(memory_tracking::names::key_none)
+            ? scratchpad.get<void>(memory_tracking::names::key_generic_acc)
             : CTX_OUT_MEM(void *, DNNL_ARG_DST);
 
     // Retrieve primitive resource and configured Compute Library objects

src/cpu/aarch64/acl_deconvolution.hpp

@@ -89,7 +89,7 @@ struct acl_deconvolution_fwd_t : public primitive_t {
             , post_ops() {}
 
         DECLARE_COMMON_PD_T(
-                "acl:deconv", acl_deconvolution_fwd_t, USE_GLOBAL_SCRATCHPAD);
+                "acl", acl_deconvolution_fwd_t, USE_GLOBAL_SCRATCHPAD);
 
         status_t init(engine_t *engine) {
             using namespace data_type;
@@ -296,7 +296,7 @@ struct acl_deconvolution_fwd_t : public primitive_t {
 
             if (acl_pd_conf.use_dst_acc_for_sum) {
                 auto scratchpad = scratchpad_registry().registrar();
-                scratchpad.book(memory_tracking::names::key_none,
+                scratchpad.book(memory_tracking::names::key_generic_acc,
                         dst_d.nelems(), dst_d.data_type_size());
             }
 

src/cpu/aarch64/acl_depthwise_convolution.hpp

@@ -99,7 +99,7 @@ struct acl_depthwise_convolution_fwd_t : public primitive_t {
             if (acp_.use_dst_acc_for_sum) {
                 const memory_desc_wrapper dst_d(&dst_md_);
                 auto scratchpad = scratchpad_registry().registrar();
-                scratchpad.book(memory_tracking::names::key_none,
+                scratchpad.book(memory_tracking::names::key_generic_acc,
                         dst_d.nelems(), dst_d.data_type_size());
             }
 

src/cpu/aarch64/acl_gemm_convolution.hpp

@@ -94,7 +94,7 @@ struct acl_gemm_convolution_fwd_t : public primitive_t {
             if (acp_.use_dst_acc_for_sum) {
                 const memory_desc_wrapper dst_d(&dst_md_);
                 auto scratchpad = scratchpad_registry().registrar();
-                scratchpad.book(memory_tracking::names::key_none,
+                scratchpad.book(memory_tracking::names::key_generic_acc,
                         dst_d.nelems(), dst_d.data_type_size());
             }
 

src/cpu/aarch64/acl_indirect_gemm_convolution.hpp

@@ -103,7 +103,7 @@ struct acl_indirect_gemm_convolution_fwd_t : public primitive_t {
             if (acp_.use_dst_acc_for_sum) {
                 const memory_desc_wrapper dst_d(&dst_md_);
                 auto scratchpad = scratchpad_registry().registrar();
-                scratchpad.book(memory_tracking::names::key_none,
+                scratchpad.book(memory_tracking::names::key_generic_acc,
                         dst_d.nelems(), dst_d.data_type_size());
             }
 

src/cpu/aarch64/acl_inner_product.cpp

@@ -46,7 +46,7 @@ status_t acl_inner_product_fwd_t::execute_forward(const exec_ctx_t &ctx) const {
     // If we have an unfused sum post op, put the result in a scratchpad tensor.
     // Result will be summed to the dst during acl_post_ops.execute
     auto dst_base = use_dst_acc_for_sum
-            ? scratchpad.get<void>(memory_tracking::names::key_none)
+            ? scratchpad.get<void>(memory_tracking::names::key_generic_acc)
             : CTX_OUT_MEM(void *, DNNL_ARG_DST);
     acl_obj.dst_tensor.allocator()->import_memory(dst_base);
 

src/cpu/aarch64/acl_inner_product.hpp

@@ -121,7 +121,7 @@ struct acl_inner_product_fwd_t : public primitive_t {
             if (aip.use_dst_acc_for_sum) {
                 const memory_desc_wrapper dst_d(&dst_md_);
                 auto scratchpad = scratchpad_registry().registrar();
-                scratchpad.book(memory_tracking::names::key_none,
+                scratchpad.book(memory_tracking::names::key_generic_acc,
                         dst_d.nelems(), dst_d.data_type_size());
             }
 

src/cpu/aarch64/acl_winograd_convolution.hpp

@@ -106,7 +106,7 @@ struct acl_wino_convolution_fwd_t : public primitive_t {
             if (acp_.use_dst_acc_for_sum) {
                 const memory_desc_wrapper dst_d(&dst_md_);
                 auto scratchpad = scratchpad_registry().registrar();
-                scratchpad.book(memory_tracking::names::key_none,
+                scratchpad.book(memory_tracking::names::key_generic_acc,
                         dst_d.nelems(), dst_d.data_type_size());
             }