Update XNNPACK revision to fcbf55a (#2223)

Summary:
Update the revision of the XNNPACK library dependency to fcbf55a. Note that this is part of a synchronized update of XNNPACK for ExecuTorch, PyTorch, and Meta internal.

bypass-github-export-checks
bypass-github-pytorch-ci-checks
bypass-github-executorch-ci-checks
allow-large-files

Pull Request resolved: #2223

Test Plan: Tested via internal and external CI for PyTorch and ExecuTorch. Also note that the full changeset was tested via D54463495.

Reviewed By: mcr229

Differential Revision: D54499911

Pulled By: GregoryComer

fbshipit-source-id: b9e2d3bbb44e63f7c3a422c215b8a7915adce46d

Author: GregoryComer

backends/xnnpack/runtime/XNNExecutor.cpp

@@ -32,15 +32,7 @@ Error XNNExecutor::set_external_input(
     ET_LOG(Error, "Input dim mismatch between tensor and shape struct");
   }
 
-#ifdef ENABLE_DYNAMIC_QUANTIZATION
-  externals_.emplace_back(xnn_external_value{
-      id,
-      input->mutable_data_ptr(),
-      static_cast<size_t>(shape->num_dims),
-      shape->dim});
-#else
   externals_.emplace_back(xnn_external_value{id, input->mutable_data_ptr()});
-#endif
   return Error::Ok;
 }
 

backends/xnnpack/runtime/XNNExecutor.h

@@ -124,16 +124,8 @@ class XNNExecutor {
         outputs.size());
 
     for (int i = 0; i < outputs.size(); i++) {
-#ifdef ENABLE_DYNAMIC_QUANTIZATION
-      externals_.emplace_back(xnn_external_value{
-          output_ids_[i],
-          outputs[i]->mutable_data_ptr<float>(),
-          static_cast<size_t>(output_shapes[i].num_dims),
-          output_shapes[i].dim});
-#else
       externals_.emplace_back(xnn_external_value{
           output_ids_[i], outputs[i]->mutable_data_ptr<float>()});
-#endif
     }
 
     return Error::Ok;

backends/xnnpack/third-party/XNNPACK

@@ -36,6 +36,7 @@
     "PROD_AVX512F_MICROKERNEL_SRCS": "defined(__i386__) || defined(__i686__) || defined(__x86_64__)",
     "PROD_AVX512SKX_MICROKERNEL_SRCS": "defined(__i386__) || defined(__i686__) || defined(__x86_64__)",
     "PROD_AVX512VBMI_MICROKERNEL_SRCS": "defined(__i386__) || defined(__i686__) || defined(__x86_64__)",
+    "PROD_AVX512VNNIGFNI_MICROKERNEL_SRCS": "defined(__i386__) || defined(__i686__) || defined(__x86_64__)",
     "PROD_AVX512VNNI_MICROKERNEL_SRCS": "defined(__i386__) || defined(__i686__) || defined(__x86_64__)",
     "PROD_RVV_MICROKERNEL_SRCS": "defined(__riscv) || defined(__riscv__)",
     "PROD_AVXVNNI_MICROKERNEL_SRCS": "defined(__i386__) || defined(__i686__) || defined(__x86_64__)",
@@ -80,6 +81,7 @@
     "PROD_AVX512F_MICROKERNEL_SRCS",
     "PROD_AVX512SKX_MICROKERNEL_SRCS",
     "PROD_AVX512VBMI_MICROKERNEL_SRCS",
+    "PROD_AVX512VNNIGFNI_MICROKERNEL_SRCS",
     "PROD_AVX512VNNI_MICROKERNEL_SRCS",
     "PROD_RVV_MICROKERNEL_SRCS",
     "PROD_AVXVNNI_MICROKERNEL_SRCS",

backends/xnnpack/third-party/generate-xnnpack-wrappers.py

@@ -17,6 +17,7 @@ load(
     "PROD_AVX512F_MICROKERNEL_SRCS",
     "PROD_AVX512SKX_MICROKERNEL_SRCS",
     "PROD_AVX512VBMI_MICROKERNEL_SRCS",
+    "PROD_AVX512VNNIGFNI_MICROKERNEL_SRCS",
     "PROD_AVX512VNNI_MICROKERNEL_SRCS",
     "PROD_AVXVNNI_MICROKERNEL_SRCS",
     "PROD_AVX_MICROKERNEL_SRCS",
@@ -1125,9 +1126,45 @@ def define_xnnpack():
         ],
     )
 
+    AVX512VNNIGFNI_COMPILER_FLAGS = AVX512VNNI_COMPILER_FLAGS + [
+        "-mgfni",
+    ]
+
+    # @lint-ignore BUCKLINT: native and fb_native are explicitly forbidden in fbcode.
+    native.cxx_library(
+        name = "ukernels_avx512vnnigfni",
+        srcs = select({
+            "DEFAULT": PROD_AVX512VNNIGFNI_MICROKERNEL_SRCS,
+            "ovr_config//cpu:arm32": DEFAULT_DUMMY_SRC,
+            "ovr_config//cpu:arm64": DEFAULT_DUMMY_SRC,
+        }),
+        headers = subdir_glob([
+            ("XNNPACK/src", "**/*.h"),
+            ("XNNPACK/src", "**/*.c"),
+        ]),
+        header_namespace = "",
+        compiler_flags = [
+            "-O2",
+            "-Wno-error=missing-braces",  # required since the SGX toolchain does not have this by default
+        ] + select({
+            "DEFAULT": AVX512VNNIGFNI_COMPILER_FLAGS,
+            "ovr_config//cpu:arm32": [],
+            "ovr_config//cpu:arm64": [],
+        }),
+        preferred_linkage = "static",
+        preprocessor_flags = [
+            "-DXNN_LOG_LEVEL=0",
+        ],
+        exported_deps = [
+            ":interface",
+        ],
+    )
+
     AVXVNNI_COMPILER_FLAGS = [
         "-mavx2",
         "-mavxvnni",
+        "-mf16c",
+        "-mfma",
     ]
 
     # @lint-ignore BUCKLINT: native and fb_native are explicitly forbidden in fbcode.
@@ -1180,6 +1217,7 @@ def define_xnnpack():
         ":ukernels_ssse3",
         ":ukernels_xop",
         ":ukernels_avx512vbmi",
+        ":ukernels_avx512vnnigfni",
         ":ukernels_avx512vnni",
         ":ukernels_avxvnni",
     ]