Re-enable AMX IGEMM kernels after fixing them.

PiperOrigin-RevId: 716385364
google · Jan 16, 2025 · 5ff4949 · 5ff4949
1 parent f909205
commit 5ff4949
Show file tree

Hide file tree

Showing 9 changed files with 178 additions and 35 deletions.
diff --git a/src/xnnpack/igemm.h b/src/xnnpack/igemm.h
@@ -870,7 +870,7 @@ DECLARE_QD8_F16_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f16_qc8w_igemm_minmax
 DECLARE_QD8_F16_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f16_qc8w_igemm_minmax_ukernel_1x64c4__avx512amx)
 DECLARE_QD8_F16_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f16_qc8w_igemm_minmax_ukernel_7x64c4__avx512amx)
 DECLARE_QD8_F16_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f16_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx)
-//DECLARE_QD8_F16_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f16_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm)
+DECLARE_QD8_F16_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f16_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm)
 
 #define DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(fn_name)               \
   XNN_INTERNAL void fn_name(                                                      \
@@ -1072,17 +1072,17 @@ DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c4__avx512amx)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_7x16c4__avx512amx)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx)
-//DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx_prfm)
+DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx_prfm)
 
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x32c4__avx512amx)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_7x32c4__avx512amx)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x32c4__avx512amx)
-//DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x32c4__avx512amx_prfm)
+DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x32c4__avx512amx_prfm)
 
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x64c4__avx512amx)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_7x64c4__avx512amx)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx)
-//DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm)
+DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm)
 
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c4__avx512vnni)
 DECLARE_QD8_F32_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x16c4__avx512vnni)
@@ -1588,17 +1588,17 @@ DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_uk
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16c4__avx512amx)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_7x16c4__avx512amx)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx)
-//DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx_prfm)
+DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx_prfm)
 
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x32c4__avx512amx)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_7x32c4__avx512amx)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x32c4__avx512amx)
-//DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x32c4__avx512amx_prfm)
+DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x32c4__avx512amx_prfm)
 
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x64c4__avx512amx)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_7x64c4__avx512amx)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x64c4__avx512amx)
-//DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x64c4__avx512amx_prfm)
+DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x64c4__avx512amx_prfm)
 
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16c4__avx512vnni)
 DECLARE_QS8_QC8W_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x16c4__avx512vnni)

diff --git a/test/qd8-f16-qc8w-igemm-minmax-3.cc b/test/qd8-f16-qc8w-igemm-minmax-3.cc
@@ -318,6 +318,26 @@ std::vector<GemmTestParams> CreateTests1(
       [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
         return info.param.test_name;
       });
+
+  INSTANTIATE_TEST_SUITE_P(
+      QD8_F16_QC8W_IGEMM_MINMAX_16X64C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/64, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/true,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qd8_f16_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm,
+                        xnn_init_f16_minmax_scalar_params,
+                        xnn_pack_qs8_conv_goki_w);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
 #endif  // XNN_ENABLE_AVX512AMX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 

diff --git a/test/qd8-f16-qc8w-igemm-minmax.yaml b/test/qd8-f16-qc8w-igemm-minmax.yaml
@@ -16,10 +16,10 @@
   init: xnn_init_f16_minmax_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qd8_f16_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm
-#  init: xnn_init_f16_minmax_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qd8_f16_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm
+  init: xnn_init_f16_minmax_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 # ARM NEONI8MM
 - name: xnn_qd8_f16_qc8w_igemm_minmax_ukernel_1x8c8__neoni8mm

diff --git a/test/qd8-f32-qc8w-igemm-minmax-2.cc b/test/qd8-f32-qc8w-igemm-minmax-2.cc
@@ -279,6 +279,26 @@ std::vector<GemmTestParams> CreateTests1(
 
 
 #if XNN_ENABLE_AVX512AMX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+  INSTANTIATE_TEST_SUITE_P(
+      QD8_F32_QC8W_IGEMM_MINMAX_16X16C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/16, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/true,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx_prfm,
+                        xnn_init_f32_minmax_scalar_params,
+                        xnn_pack_qs8_conv_goki_w);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
+
   INSTANTIATE_TEST_SUITE_P(
       QD8_F32_QC8W_IGEMM_MINMAX_7X32C4__AVX512AMX, GemmTest,
       testing::ValuesIn(CreateTests1(
@@ -318,6 +338,26 @@ std::vector<GemmTestParams> CreateTests1(
       [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
         return info.param.test_name;
       });
+
+  INSTANTIATE_TEST_SUITE_P(
+      QD8_F32_QC8W_IGEMM_MINMAX_16X64C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/64, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/true,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm,
+                        xnn_init_f32_minmax_scalar_params,
+                        xnn_pack_qs8_conv_goki_w);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
 #endif  // XNN_ENABLE_AVX512AMX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 

diff --git a/test/qd8-f32-qc8w-igemm-minmax-3.cc b/test/qd8-f32-qc8w-igemm-minmax-3.cc
@@ -319,6 +319,26 @@ std::vector<GemmTestParams> CreateTests1(
         return info.param.test_name;
       });
 
+  INSTANTIATE_TEST_SUITE_P(
+      QD8_F32_QC8W_IGEMM_MINMAX_16X32C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/32, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/true,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x32c4__avx512amx_prfm,
+                        xnn_init_f32_minmax_scalar_params,
+                        xnn_pack_qs8_conv_goki_w);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
+
   INSTANTIATE_TEST_SUITE_P(
       QD8_F32_QC8W_IGEMM_MINMAX_1X64C4__AVX512AMX, GemmTest,
       testing::ValuesIn(CreateTests1(

diff --git a/test/qd8-f32-qc8w-igemm-minmax.yaml b/test/qd8-f32-qc8w-igemm-minmax.yaml
@@ -16,10 +16,10 @@
   init: xnn_init_f32_minmax_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx_prfm
-#  init: xnn_init_f32_minmax_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x16c4__avx512amx_prfm
+  init: xnn_init_f32_minmax_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 - name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x32c4__avx512amx
   init: xnn_init_f32_minmax_scalar_params
@@ -33,10 +33,10 @@
   init: xnn_init_f32_minmax_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x32c4__avx512amx_prfm
-#  init: xnn_init_f32_minmax_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x32c4__avx512amx_prfm
+  init: xnn_init_f32_minmax_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 - name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x64c4__avx512amx
   init: xnn_init_f32_minmax_scalar_params
@@ -50,10 +50,10 @@
   init: xnn_init_f32_minmax_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm
-#  init: xnn_init_f32_minmax_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_16x64c4__avx512amx_prfm
+  init: xnn_init_f32_minmax_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 # Scalar
 - name: xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x2__scalar

diff --git a/test/qs8-qc8w-igemm-minmax-fp32-2.cc b/test/qs8-qc8w-igemm-minmax-fp32-2.cc
@@ -289,6 +289,48 @@ std::vector<GemmTestParams> CreateTests1(
       [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
         return info.param.test_name;
       });
+
+  INSTANTIATE_TEST_SUITE_P(
+      QS8_QC8W_IGEMM_MINMAX_FP32_16X16C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/16, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/false,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx_prfm,
+                        xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+                        xnn_pack_qs8_conv_goki_w,
+                        xnn_qs8_requantize_fp32);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
+
+  INSTANTIATE_TEST_SUITE_P(
+      QS8_QC8W_IGEMM_MINMAX_FP32_16X64C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/64, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/false,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x64c4__avx512amx_prfm,
+                        xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+                        xnn_pack_qs8_conv_goki_w,
+                        xnn_qs8_requantize_fp32);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
 #endif  // XNN_ENABLE_AVX512AMX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 

diff --git a/test/qs8-qc8w-igemm-minmax-fp32.cc b/test/qs8-qc8w-igemm-minmax-fp32.cc
@@ -332,6 +332,27 @@ std::vector<GemmTestParams> CreateTests1(
         return info.param.test_name;
       });
 
+  INSTANTIATE_TEST_SUITE_P(
+      QS8_QC8W_IGEMM_MINMAX_FP32_16X32C4__AVX512AMX_PRFM, GemmTest,
+      testing::ValuesIn(CreateTests1(
+          /*k_block=*/64,
+          /*adj_k_block=*/64,
+          /*mr=*/16, /*nr=*/32, /*kr=*/4, /*sr=*/1,
+          /*is_igemm=*/false,
+          /*unsigned_inputs=*/false,
+          [](GemmMicrokernelTester& tester) {
+            tester.Test(xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x32c4__avx512amx_prfm,
+                        xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+                        xnn_pack_qs8_conv_goki_w,
+                        xnn_qs8_requantize_fp32);
+          },
+          []() {
+            TEST_REQUIRES_X86_AVX512AMX;
+          })),
+      [](const testing::TestParamInfo<GemmTest::ParamType>& info) {
+        return info.param.test_name;
+      });
+
   INSTANTIATE_TEST_SUITE_P(
       QS8_QC8W_IGEMM_MINMAX_FP32_1X64C4__AVX512AMX, GemmTest,
       testing::ValuesIn(CreateTests1(

diff --git a/test/qs8-qc8w-igemm-minmax-fp32.yaml b/test/qs8-qc8w-igemm-minmax-fp32.yaml
@@ -16,10 +16,10 @@
   init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx_prfm
-#  init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x16c4__avx512amx_prfm
+  init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 - name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x32c4__avx512amx
   init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
@@ -33,10 +33,10 @@
   init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x32c4__avx512amx_prfm
-#  init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x32c4__avx512amx_prfm
+  init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 - name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x64c4__avx512amx
   init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
@@ -50,10 +50,10 @@
   init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
   pack: xnn_pack_qs8_conv_goki_w
   k-block: 64
-#- name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x64c4__avx512amx_prfm
-#  init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
-#  pack: xnn_pack_qs8_conv_goki_w
-#  k-block: 64
+- name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_16x64c4__avx512amx_prfm
+  init: xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params
+  pack: xnn_pack_qs8_conv_goki_w
+  k-block: 64
 
 # AArch32 assembly
 - name: xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x8__asm_aarch32_neon_mlal_lane_cortex_a7