fix(windows/dnn/ci): fix cpu x86_windows_xp flags build issue

*as xp 32 mkl crash issue after "11.1.4 (build date 20140806)"
we can not upgrade intel_mkl_windows_xp_archive to
MKL 2019.0.1 (build date 20180928), so we make a difference between
xp with linux/macos/windows_no_xp

GitOrigin-RevId: 97b049d7893fec217aaa837f0a6909991b788cf0

dnn/src/x86/matrix_mul/algos.cpp

@@ -14,21 +14,9 @@
 #include "src/common/utils.h"
 #include "src/fallback/matrix_mul/gemm_impl.h"
 #include "src/x86/matrix_mul/int8/strategy.h"
-#include "src/x86/utils.h"
 
 #include "src/x86/matrix_mul/f32/strategy.h"
 
-#if MEGDNN_X86_WITH_MKL
-#include <mkl.h>
-#include <mkl_cblas.h>
-#elif MEGDNN_X86_WITH_OPENBLAS
-#include <cblas.h>
-#else
-#endif
-
-#if MEGDNN_X86_WITH_MKL_DNN
-#include <mkldnn.h>
-#endif
 
 MIDOUT_DECL(megdnn_x86_matmul_kern)
 MIDOUT_DECL(megdnn_x86_matmul_kern_mk8_8x8)
@@ -55,7 +43,7 @@ void f32_blas_kern(const MatrixMulImpl::KernParam& kern_param) {
 #endif
 }
 
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
 void f32_blas_kern_only_packA(const MatrixMulImpl::KernParam& kern_param,
                                  const void* a_panel, const void* b_panel) {
   MEGDNN_MARK_USED_VAR(b_panel);
@@ -93,7 +81,7 @@ MatrixMulImpl::kern_t MatrixMulImpl::AlgoF32Blas::get_kern(
 }
 
 /* ===================== AlgoF32BlasPackA====================== */
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
 bool MatrixMulImpl::AlgoF32MKLPackA::usable(
         const KernSizeParam& kern_size_param) const {
     return kern_size_param.compute_mode == Param::ComputeMode::DEFAULT &&

dnn/src/x86/matrix_mul/algos.h

@@ -28,7 +28,7 @@ public:
     PackMode packmode() const override { return PackMode::NO_PACK; }
 };
 
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
 class MatrixMulImpl::AlgoF32MKLPackA : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }

dnn/src/x86/matrix_mul/opr_impl.cpp

@@ -12,7 +12,6 @@
 #include "src/x86/matrix_mul/opr_impl.h"
 #include "src/common/metahelper.h"
 #include "src/x86/matrix_mul/algos.h"
-#include "src/x86/utils.h"
 using namespace megdnn;
 using namespace x86;
 
@@ -25,7 +24,7 @@ void* const MatrixMulImpl::sm_x86_algo_type = &x86_algo_type_storage;
 class MatrixMulImpl::AlgoPack : NonCopyableObj {
     AlgoF32Blas f32blas;
 
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
     AlgoF32MKLPackA f32mkl_packa;
 #endif
 #if MEGDNN_X86_WITH_VNNI
@@ -57,7 +56,7 @@ public:
         all_algos.emplace_back(&algoint8x8x32mkldnn);
 #endif
         all_algos.emplace_back(&f32blas);
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
         all_algos.emplace_back(&f32mkl_packa);
 #endif
     }

dnn/src/x86/matrix_mul/opr_impl.h

@@ -9,8 +9,22 @@
  * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  */
 #pragma once
+
 #include "src/common/utils.h"
 #include "src/fallback/matrix_mul/opr_impl.h"
+#include "src/x86/utils.h"
+
+#if MEGDNN_X86_WITH_MKL
+#include <mkl.h>
+#include <mkl_cblas.h>
+#elif MEGDNN_X86_WITH_OPENBLAS
+#include <cblas.h>
+#else
+#endif
+
+#if MEGDNN_X86_WITH_MKL_DNN
+#include <mkldnn.h>
+#endif
 
 namespace megdnn {
 namespace x86 {
@@ -26,7 +40,7 @@ public:
 protected:
     static void* const sm_x86_algo_type;
     class AlgoF32Blas;
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
     class AlgoF32MKLPackA;
 #endif
 #if MEGDNN_X86_WITH_VNNI

dnn/src/x86/utils.h

@@ -14,6 +14,17 @@
 #include <vector>
 #include "src/common/utils.h"
 
+#if MEGDNN_X86_WITH_MKL
+#include <mkl.h>
+//! As INTEL_MKL_VERSION >= 20190001 support SUPPORT_MKL_PACKED_GEMM
+#if INTEL_MKL_VERSION >= 20190001
+#define SUPPORT_MKL_PACKED_GEMM 1
+#else
+#define SUPPORT_MKL_PACKED_GEMM 0
+#endif
+
+#endif
+
 namespace megdnn {
 namespace x86 {
 

dnn/test/x86/conv_bias.cpp

@@ -853,7 +853,7 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32) {
 #undef cb
 }
 
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
 TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32_PACKA) {
     using namespace conv_bias;
     std::vector<TestArg> args;

dnn/test/x86/matrix_mul.cpp

@@ -52,7 +52,7 @@ TEST_F(X86, MATRIX_MUL_SSE_8X8X32) {
                                  handle(), "X86_INT8X8X32_SSE_4X8X2");
 }
 
-#if MEGDNN_X86_WITH_MKL
+#if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
 TEST_F(X86, MATRIX_MUL_MKL_PACKA) {
     matrix_mul::check_matrix_mul(dtype::Float32{}, dtype::Float32{},
                                  dtype::Float32{}, handle(),