Unify GemmKaiKernelExecutor::execute

Author: aobolensk

src/plugins/intel_cpu/src/emitters/snippets/aarch64/jit_gemm_emitter.cpp

@@ -133,10 +133,7 @@ void jit_gemm_emitter::emit_call(const std::vector<size_t>& mem_ptrs_idxs) const
     Xbyak_aarch64::XReg x0(0);
     Xbyak_aarch64::XReg x1(1);
 
-    auto execute_func_ptr =
-        static_cast<void (*)(const GemmKaiKernelExecutor*, const GemmKaiKernelExecutor::call_args*)>(
-            GemmKaiKernelExecutor::execute);
-    h->mov(call_address_reg, reinterpret_cast<uintptr_t>(execute_func_ptr));
+    h->mov(call_address_reg, reinterpret_cast<uintptr_t>(GemmKaiKernelExecutor::execute));
 
     h->mov(x0, reinterpret_cast<uintptr_t>(m_kernel_executor_kai.get()));
     h->mov(x1, h->sp);
@@ -153,10 +150,7 @@ uintptr_t jit_gemm_emitter::get_compiled_kernel_ptr() const {
 }
 
 uintptr_t jit_gemm_emitter::get_execute_function_ptr() {
-    auto execute_func_ptr =
-        static_cast<void (*)(const GemmKaiKernelExecutor*, const GemmKaiKernelExecutor::call_args*)>(
-            GemmKaiKernelExecutor::execute);
-    return reinterpret_cast<const uintptr_t>(execute_func_ptr);
+    return reinterpret_cast<const uintptr_t>(GemmKaiKernelExecutor::execute);
 }
 
 }  // namespace ov::intel_cpu::aarch64

src/plugins/intel_cpu/src/emitters/snippets/aarch64/kernel_executors/gemm.cpp

@@ -54,9 +54,10 @@ void GemmKaiKernelExecutor::update_config(const ov::snippets::lowered::Expressio
     config.update(M, N, K, LDA, LDB, LDC, beta);
 }
 
-void GemmKaiKernelExecutor::execute(const GemmKaiKernelExecutor* executor, void* in0, void* in1, void* out0) {
+void GemmKaiKernelExecutor::execute(const GemmKaiKernelExecutor* executor, const call_args* args) {
     OV_CPU_JIT_EMITTER_ASSERT(executor, "has nullptr executor");
-    // matmul for input1 and slices of repacked input2
+    OV_CPU_JIT_EMITTER_ASSERT(args, "has nullptr args");
+
     const auto& config = static_cast<const GemmKernelKaiConfig&>(executor->get_config());
     const auto& kernel = executor->get_kernel();
     const auto& ukernel = *kernel->gemm_ukernel;
@@ -80,12 +81,12 @@ void GemmKaiKernelExecutor::execute(const GemmKaiKernelExecutor* executor, void*
         const size_t dst_offset = ukernel.get_dst_offset(0, n_start, dst_stride_row);
         // in0, in1, out is point to current block memory, based on block loop info, and shift done in loop begin and
         // end emitters(adjusted copyb loop info as repack outside block loops).
-        float* rhs_ptr = static_cast<float*>(in1) + rhs_packed_offset / sizeof(float);
-        float* dst_ptr = (static_cast<float*>(out0) + dst_offset / (sizeof(float)));
+        float* rhs_ptr = const_cast<float*>(static_cast<const float*>(args->B)) + rhs_packed_offset / sizeof(float);
+        float* dst_ptr = const_cast<float*>(static_cast<const float*>(args->C)) + dst_offset / (sizeof(float));
         ukernel.run_matmul(M,
                            n_block_size,
                            K,
-                           in0,
+                           const_cast<void*>(args->A),
                            lhs_stride,
                            rhs_ptr,
                            dst_ptr,
@@ -96,11 +97,4 @@ void GemmKaiKernelExecutor::execute(const GemmKaiKernelExecutor* executor, void*
     }
 }
 
-void GemmKaiKernelExecutor::execute(const GemmKaiKernelExecutor* executor, const call_args* args) {
-    OV_CPU_JIT_EMITTER_ASSERT(executor, "has nullptr executor");
-    OV_CPU_JIT_EMITTER_ASSERT(args, "has nullptr args");
-
-    execute(executor, const_cast<void*>(args->A), const_cast<void*>(args->B), args->C);
-}
-
 }  // namespace ov::intel_cpu::aarch64

src/plugins/intel_cpu/src/emitters/snippets/aarch64/kernel_executors/gemm.hpp

@@ -65,10 +65,6 @@ class GemmKaiKernelExecutor : public snippets::KernelExecutor<GemmKernelKaiConfi
     // No need kernel update, just update config is enough for update. The universal ukernel is reused with any config.
     void update_kernel(const GemmKernelKaiConfig& config,
                        std::shared_ptr<GemmCompiledKernel>& kernel) const override final;
-
-    // Function that will be called in runtime to execute the kernel
-    static void execute(const GemmKaiKernelExecutor* executor, void* in0, void* in1, void* out0);
-
     // ABI-compliant execute function that takes call_args structure
     static void execute(const GemmKaiKernelExecutor* executor, const call_args* args);