f

oneapi-src · Dec 13, 2024 · 9b4efc6 · 9b4efc6
1 parent 21b89c3
commit 9b4efc6
Show file tree

Hide file tree

Showing 4 changed files with 14 additions and 15 deletions.
diff --git a/src/common/primitive_attr_quant.hpp b/src/common/primitive_attr_quant.hpp
@@ -138,15 +138,10 @@ struct scales_t : public c_compatible {
         CHECK(scales_[arg].set(mask, data_type, group_ndims, group_dims));
         return status::success;
     }
-
-    // Used directly in the following scenarios:
-    // * CPU GeMM-based matmul to transfer scale as `alpha` arg, referred as
-    //   `gemm_applies_output_scales_`.
-    status_t erase(int arg) {
-        if (!check_arg(arg)) return status::invalid_arguments;
-        const auto it = scales_.find(arg);
-        if (it != scales_.end()) scales_.erase(it);
-        return status::success;
+    // Use this interface with `default_quant_entry` when need to remove a
+    // specific scale.
+    status_t set(int arg, const quant_entry_t &other) {
+        return scales_[arg].set(other);
     }
 
     bool has_default_values(const std::vector<int> &supported_args = {}) const {

diff --git a/src/cpu/matmul/gemm_bf16_matmul.cpp b/src/cpu/matmul/gemm_bf16_matmul.cpp
@@ -148,9 +148,11 @@ status_t gemm_bf16_matmul_t<dst_type>::pd_t::check_and_configure_attributes(
             = attr()->scales_.get_mask(DNNL_ARG_WEIGHTS) == 0 && !with_bias();
 
     if (params_.gemm_applies_output_scales_) {
-        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.erase(DNNL_ARG_SRC),
+        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.set(
+                                    DNNL_ARG_SRC, default_quant_entry()),
                 VERBOSE_UNSUPPORTED_SCALES_CFG);
-        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.erase(DNNL_ARG_WEIGHTS),
+        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.set(
+                                    DNNL_ARG_WEIGHTS, default_quant_entry()),
                 VERBOSE_UNSUPPORTED_SCALES_CFG);
     }
 

diff --git a/src/cpu/matmul/gemm_f32_matmul.cpp b/src/cpu/matmul/gemm_f32_matmul.cpp
@@ -133,9 +133,11 @@ status_t gemm_f32_matmul_t::pd_t::configure_attributes() {
     params_.gemm_applies_output_scales_
             = attr()->scales_.get_mask(DNNL_ARG_WEIGHTS) == 0 && !with_bias();
     if (params_.gemm_applies_output_scales_) {
-        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.erase(DNNL_ARG_SRC),
+        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.set(
+                                    DNNL_ARG_SRC, default_quant_entry()),
                 VERBOSE_UNSUPPORTED_SCALES_CFG);
-        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.erase(DNNL_ARG_WEIGHTS),
+        VDISPATCH_MATMUL_SC(params_.pp_attr_.scales_.set(
+                                    DNNL_ARG_WEIGHTS, default_quant_entry()),
                 VERBOSE_UNSUPPORTED_SCALES_CFG);
     }
 

diff --git a/src/cpu/ref_fused_convolution.hpp b/src/cpu/ref_fused_convolution.hpp
@@ -227,8 +227,8 @@ struct ref_fused_convolution_fwd_t : public primitive_t {
                 auto &scale
                         = attr_1x1.scales_.get(DNNL_ARG_ATTR_POST_OP_DW | arg);
                 if (!scale.has_default_values())
-                    CHECK(attr_1x1.scales_.erase(
-                            DNNL_ARG_ATTR_POST_OP_DW | arg));
+                    CHECK(attr_1x1.scales_.set(DNNL_ARG_ATTR_POST_OP_DW | arg,
+                            default_quant_entry()));
             }
             // erase post-ops after fusion as they will be handled separately
             auto &e = attr_1x1.post_ops_.entry_;