Add ifdef for xnn_enable_avx512f and xnn_enable_avx512skx

- build with --define=xnn_enable_avx512f=false etc PiperOrigin-RevId: 688667461
google · Oct 22, 2024 · e71b737 · e71b737
1 parent 31205d0
commit e71b737
Show file tree

Hide file tree

Showing 88 changed files with 514 additions and 276 deletions.
diff --git a/bench/f16-rmax.cc b/bench/f16-rmax.cc
@@ -128,7 +128,7 @@ static void f16_rmax(
     ->UseRealTime();
 #endif  // XNN_ENABLE_AVX512FP16 && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
-#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+#if XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
   BENCHMARK_CAPTURE(f16_rmax, avx512skx_u16,
                     xnn_f16_rmax_ukernel__avx512skx_u16,
                     /*init_params=*/nullptr,
@@ -159,7 +159,9 @@ static void f16_rmax(
                     benchmark::utils::CheckAVX512SKX)
     ->Apply(benchmark::utils::ReductionParameters<xnn_float16>)
     ->UseRealTime();
+#endif  // XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
   BENCHMARK_CAPTURE(f16_rmax, f16c_u32,
                     xnn_f16_rmax_ukernel__f16c_u32,
                     /*init_params=*/nullptr,

diff --git a/bench/f16-rmin.cc b/bench/f16-rmin.cc
@@ -128,7 +128,7 @@ static void f16_rmin(
     ->UseRealTime();
 #endif  // XNN_ENABLE_AVX512FP16 && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
-#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+#if XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
   BENCHMARK_CAPTURE(f16_rmin, avx512skx_u16,
                     xnn_f16_rmin_ukernel__avx512skx_u16,
                     /*init_params=*/nullptr,
@@ -159,7 +159,7 @@ static void f16_rmin(
                     benchmark::utils::CheckAVX512SKX)
     ->Apply(benchmark::utils::ReductionParameters<xnn_float16>)
     ->UseRealTime();
-#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+#endif  // XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 BENCHMARK_CAPTURE(f16_rmin, scalar_u1,
                   xnn_f16_rmin_ukernel__scalar_u1)

diff --git a/bench/f16-rminmax.cc b/bench/f16-rminmax.cc
@@ -128,7 +128,7 @@ static void f16_rminmax(
     ->UseRealTime();
 #endif  // XNN_ENABLE_AVX512FP16 && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
-#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+#if XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
   BENCHMARK_CAPTURE(f16_rminmax, avx512skx_u16,
                     xnn_f16_rminmax_ukernel__avx512skx_u16,
                     /*init_params=*/nullptr,
@@ -159,7 +159,7 @@ static void f16_rminmax(
                     benchmark::utils::CheckAVX512SKX)
     ->Apply(benchmark::utils::ReductionParameters<xnn_float16>)
     ->UseRealTime();
-#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+#endif  // XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 BENCHMARK_CAPTURE(f16_rminmax, scalar_u1,
                   xnn_f16_rminmax_ukernel__scalar_u1)

diff --git a/bench/f32-dwconv.cc b/bench/f32-dwconv.cc
@@ -825,9 +825,69 @@ static void f32_dwconv(
   BENCHMARK_DWCONV(f32_dwconv_8f8m9l4c4s4r__neon_acc2)
   BENCHMARK_DWCONV(f32_dwconv_8f8m9l8c4s4r__neon)
   BENCHMARK_DWCONV(f32_dwconv_8f8m9l8c4s4r__neon_acc2)
-
 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
 
+#if XNN_ENABLE_AVX512F && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+  static void f32_dwconv_25p16c__avx512f(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_25p16c__avx512f,
+               xnn_init_f32_minmax_scalar_params,
+               16 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
+  }
+  static void f32_dwconv_25p16c__avx512f_acc2(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_25p16c__avx512f_acc2,
+               xnn_init_f32_minmax_scalar_params,
+               16 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
+  }
+  static void f32_dwconv_25p32c__avx512f(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_25p32c__avx512f,
+               xnn_init_f32_minmax_scalar_params,
+               32 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
+  }
+  static void f32_dwconv_25p32c__avx512f_acc2(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_25p32c__avx512f_acc2,
+               xnn_init_f32_minmax_scalar_params,
+               32 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
+  }
+
+  static void f32_dwconv_5f5m5l16c16s1r__avx512f(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_5f5m5l16c16s1r__avx512f, xnn_init_f32_minmax_scalar_params,
+               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
+               16 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
+  }
+  static void f32_dwconv_5f5m5l16c16s1r__avx512f_acc2(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_5f5m5l16c16s1r__avx512f_acc2, xnn_init_f32_minmax_scalar_params,
+               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
+               16 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
+  }
+  static void f32_dwconv_5f5m5l32c16s1r__avx512f(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_5f5m5l32c16s1r__avx512f, xnn_init_f32_minmax_scalar_params,
+               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
+               32 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
+  }
+  static void f32_dwconv_5f5m5l32c16s1r__avx512f_acc2(benchmark::State& state, const char* net) {
+    f32_dwconv(state,
+               xnn_f32_dwconv_minmax_ukernel_5f5m5l32c16s1r__avx512f_acc2, xnn_init_f32_minmax_scalar_params,
+               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
+               32 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
+  }
+
+  BENCHMARK_DWCONV(f32_dwconv_25p16c__avx512f)
+  BENCHMARK_DWCONV(f32_dwconv_25p16c__avx512f_acc2)
+  BENCHMARK_DWCONV(f32_dwconv_25p32c__avx512f)
+  BENCHMARK_DWCONV(f32_dwconv_25p32c__avx512f_acc2)
+
+  BENCHMARK_DWCONV(f32_dwconv_5f5m5l16c16s1r__avx512f)
+  BENCHMARK_DWCONV(f32_dwconv_5f5m5l16c16s1r__avx512f_acc2)
+  BENCHMARK_DWCONV(f32_dwconv_5f5m5l32c16s1r__avx512f)
+  BENCHMARK_DWCONV(f32_dwconv_5f5m5l32c16s1r__avx512f_acc2)
+#endif  // XNN_ENABLE_AVX512F && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
   static void f32_dwconv_4p4c__sse(benchmark::State& state, const char* net) {
@@ -1195,56 +1255,6 @@ static void f32_dwconv(
                benchmark::utils::CheckFMA3);
   }
 
-  static void f32_dwconv_25p16c__avx512f(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_25p16c__avx512f,
-               xnn_init_f32_minmax_scalar_params,
-               16 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
-  }
-  static void f32_dwconv_25p16c__avx512f_acc2(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_25p16c__avx512f_acc2,
-               xnn_init_f32_minmax_scalar_params,
-               16 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
-  }
-  static void f32_dwconv_25p32c__avx512f(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_25p32c__avx512f,
-               xnn_init_f32_minmax_scalar_params,
-               32 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
-  }
-  static void f32_dwconv_25p32c__avx512f_acc2(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_25p32c__avx512f_acc2,
-               xnn_init_f32_minmax_scalar_params,
-               32 /* channel tile */, 25 /* primary tile */, benchmark::utils::CheckAVX512F);
-  }
-
-  static void f32_dwconv_5f5m5l16c16s1r__avx512f(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_5f5m5l16c16s1r__avx512f, xnn_init_f32_minmax_scalar_params,
-               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
-               16 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
-  }
-  static void f32_dwconv_5f5m5l16c16s1r__avx512f_acc2(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_5f5m5l16c16s1r__avx512f_acc2, xnn_init_f32_minmax_scalar_params,
-               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
-               16 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
-  }
-  static void f32_dwconv_5f5m5l32c16s1r__avx512f(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_5f5m5l32c16s1r__avx512f, xnn_init_f32_minmax_scalar_params,
-               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
-               32 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
-  }
-  static void f32_dwconv_5f5m5l32c16s1r__avx512f_acc2(benchmark::State& state, const char* net) {
-    f32_dwconv(state,
-               xnn_f32_dwconv_minmax_ukernel_5f5m5l32c16s1r__avx512f_acc2, xnn_init_f32_minmax_scalar_params,
-               5 /* first pass tile */, 5 /* middle pass tile */, 5 /* last pass tile */,
-               32 /* channel tile */, 16 /* channel subtile */, 1 /* channel round */, benchmark::utils::CheckAVX512F);
-  }
-
   BENCHMARK_DWCONV(f32_dwconv_4p4c__sse)
   BENCHMARK_DWCONV(f32_dwconv_9p4c__sse)
   BENCHMARK_DWCONV(f32_dwconv_25p4c__sse)
@@ -1308,19 +1318,8 @@ static void f32_dwconv(
   BENCHMARK_DWCONV(f32_dwconv_7f6m6l16c8s4r__fma3_acc2)
   BENCHMARK_DWCONV(f32_dwconv_7f6m6l32c8s4r__fma3)
   BENCHMARK_DWCONV(f32_dwconv_7f6m6l32c8s4r__fma3_acc2)
-
-  BENCHMARK_DWCONV(f32_dwconv_25p16c__avx512f)
-  BENCHMARK_DWCONV(f32_dwconv_25p16c__avx512f_acc2)
-  BENCHMARK_DWCONV(f32_dwconv_25p32c__avx512f)
-  BENCHMARK_DWCONV(f32_dwconv_25p32c__avx512f_acc2)
-
-  BENCHMARK_DWCONV(f32_dwconv_5f5m5l16c16s1r__avx512f)
-  BENCHMARK_DWCONV(f32_dwconv_5f5m5l16c16s1r__avx512f_acc2)
-  BENCHMARK_DWCONV(f32_dwconv_5f5m5l32c16s1r__avx512f)
-  BENCHMARK_DWCONV(f32_dwconv_5f5m5l32c16s1r__avx512f_acc2)
 #endif  // XNN_ARCH_X88 || XNN_ARCH_X86_64
 
-
 #if XNN_ARCH_WASM
   static void f32_dwconv_9p1c__wasm(benchmark::State& state, const char* net) {
     f32_dwconv(state,

diff --git a/bench/f32-gemm.cc b/bench/f32-gemm.cc
@@ -1306,7 +1306,7 @@ static void ruy_st(benchmark::State& state, const char* net)
   BENCHMARK_GEMM(f32_gemm_8x8s4__neonfma)
 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
 
-#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+#if XNN_ENABLE_AVX512F && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
   static void f32_gemm_1x16__avx512f_broadcast(benchmark::State& state, const char* net) {
     GEMMBenchmark(state,
       xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast,
@@ -1350,6 +1350,15 @@ static void ruy_st(benchmark::State& state, const char* net)
       benchmark::utils::CheckAVX512F);
   }
 
+  BENCHMARK_GEMM(f32_gemm_1x16__avx512f_broadcast)
+  BENCHMARK_GEMM(f32_gemm_4x16__avx512f_broadcast)
+  BENCHMARK_GEMM(f32_gemm_5x16__avx512f_broadcast)
+  BENCHMARK_GEMM(f32_gemm_6x16__avx512f_broadcast)
+  BENCHMARK_GEMM(f32_gemm_7x16__avx512f_broadcast)
+  BENCHMARK_GEMM(f32_gemm_8x16__avx512f_broadcast)
+#endif  // XNN_ENABLE_AVX512F && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
   static void f32_gemm_1x8__fma3_broadcast(benchmark::State& state, const char* net) {
     GEMMBenchmark(state,
       xnn_f32_gemm_minmax_ukernel_1x8__fma3_broadcast,
@@ -1638,13 +1647,6 @@ static void ruy_st(benchmark::State& state, const char* net)
       /*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1);
   }
 
-  BENCHMARK_GEMM(f32_gemm_1x16__avx512f_broadcast)
-  BENCHMARK_GEMM(f32_gemm_4x16__avx512f_broadcast)
-  BENCHMARK_GEMM(f32_gemm_5x16__avx512f_broadcast)
-  BENCHMARK_GEMM(f32_gemm_6x16__avx512f_broadcast)
-  BENCHMARK_GEMM(f32_gemm_7x16__avx512f_broadcast)
-  BENCHMARK_GEMM(f32_gemm_8x16__avx512f_broadcast)
-
   BENCHMARK_GEMM(f32_gemm_1x8__fma3_broadcast)
   BENCHMARK_GEMM(f32_gemm_4x8__fma3_broadcast)
   BENCHMARK_GEMM(f32_gemm_5x8__fma3_broadcast)

diff --git a/bench/f32-igemm.cc b/bench/f32-igemm.cc
@@ -682,6 +682,56 @@ static void f32_igemm(benchmark::State& state,
   BENCHMARK_CONV(f32_igemm_8x8s4__neonfma)
 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
 
+#if XNN_ENABLE_AVX512F && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+  static void f32_igemm_1x16__avx512f_broadcast(benchmark::State& state, const char* net) {
+    f32_igemm(state,
+      xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast,
+      xnn_init_f32_minmax_scalar_params,
+      /*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
+      benchmark::utils::CheckAVX512F);
+  }
+  static void f32_igemm_4x16__avx512f_broadcast(benchmark::State& state, const char* net) {
+    f32_igemm(state,
+      xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast,
+      xnn_init_f32_minmax_scalar_params,
+      /*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
+      benchmark::utils::CheckAVX512F);
+  }
+  static void f32_igemm_5x16__avx512f_broadcast(benchmark::State& state, const char* net) {
+    f32_igemm(state,
+      xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast,
+      xnn_init_f32_minmax_scalar_params,
+      /*mr=*/5, /*nr=*/16, /*kr=*/1, /*sr=*/1,
+      benchmark::utils::CheckAVX512F);
+  }
+  static void f32_igemm_6x16__avx512f_broadcast(benchmark::State& state, const char* net) {
+    f32_igemm(state,
+      xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast,
+      xnn_init_f32_minmax_scalar_params,
+      /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
+      benchmark::utils::CheckAVX512F);
+  }
+  static void f32_igemm_7x16__avx512f_broadcast(benchmark::State& state, const char* net) {
+    f32_igemm(state,
+      xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast,
+      xnn_init_f32_minmax_scalar_params,
+      /*mr=*/7, /*nr=*/16, /*kr=*/1, /*sr=*/1,
+      benchmark::utils::CheckAVX512F);
+  }
+  static void f32_igemm_8x16__avx512f_broadcast(benchmark::State& state, const char* net) {
+    f32_igemm(state,
+      xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast,
+      xnn_init_f32_minmax_scalar_params,
+      /*mr=*/8, /*nr=*/16, /*kr=*/1, /*sr=*/1,
+      benchmark::utils::CheckAVX512F);
+  }
+  BENCHMARK_CONV(f32_igemm_1x16__avx512f_broadcast)
+  BENCHMARK_CONV(f32_igemm_4x16__avx512f_broadcast)
+  BENCHMARK_CONV(f32_igemm_5x16__avx512f_broadcast)
+  BENCHMARK_CONV(f32_igemm_6x16__avx512f_broadcast)
+  BENCHMARK_CONV(f32_igemm_7x16__avx512f_broadcast)
+  BENCHMARK_CONV(f32_igemm_8x16__avx512f_broadcast)
+#endif  // XNN_ENABLE_AVX512F && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
 
 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
   static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) {
@@ -904,48 +954,6 @@ static void f32_igemm(benchmark::State& state,
       /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
       benchmark::utils::CheckFMA3);
   }
-  static void f32_igemm_1x16__avx512f_broadcast(benchmark::State& state, const char* net) {
-    f32_igemm(state,
-      xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast,
-      xnn_init_f32_minmax_scalar_params,
-      /*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
-      benchmark::utils::CheckAVX512F);
-  }
-  static void f32_igemm_4x16__avx512f_broadcast(benchmark::State& state, const char* net) {
-    f32_igemm(state,
-      xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast,
-      xnn_init_f32_minmax_scalar_params,
-      /*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
-      benchmark::utils::CheckAVX512F);
-  }
-  static void f32_igemm_5x16__avx512f_broadcast(benchmark::State& state, const char* net) {
-    f32_igemm(state,
-      xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast,
-      xnn_init_f32_minmax_scalar_params,
-      /*mr=*/5, /*nr=*/16, /*kr=*/1, /*sr=*/1,
-      benchmark::utils::CheckAVX512F);
-  }
-  static void f32_igemm_6x16__avx512f_broadcast(benchmark::State& state, const char* net) {
-    f32_igemm(state,
-      xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast,
-      xnn_init_f32_minmax_scalar_params,
-      /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
-      benchmark::utils::CheckAVX512F);
-  }
-  static void f32_igemm_7x16__avx512f_broadcast(benchmark::State& state, const char* net) {
-    f32_igemm(state,
-      xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast,
-      xnn_init_f32_minmax_scalar_params,
-      /*mr=*/7, /*nr=*/16, /*kr=*/1, /*sr=*/1,
-      benchmark::utils::CheckAVX512F);
-  }
-  static void f32_igemm_8x16__avx512f_broadcast(benchmark::State& state, const char* net) {
-    f32_igemm(state,
-      xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast,
-      xnn_init_f32_minmax_scalar_params,
-      /*mr=*/8, /*nr=*/16, /*kr=*/1, /*sr=*/1,
-      benchmark::utils::CheckAVX512F);
-  }
 
   BENCHMARK_CONV(f32_igemm_1x8__sse_load1)
   BENCHMARK_CONV(f32_igemm_3x8__sse_load1)
@@ -980,12 +988,6 @@ static void f32_igemm(benchmark::State& state,
   BENCHMARK_CONV(f32_igemm_6x16__fma3_broadcast)
   BENCHMARK_CONV(f32_igemm_5x16__fma3_broadcast_prfm)
   BENCHMARK_CONV(f32_igemm_6x16__fma3_broadcast_prfm)
-  BENCHMARK_CONV(f32_igemm_1x16__avx512f_broadcast)
-  BENCHMARK_CONV(f32_igemm_4x16__avx512f_broadcast)
-  BENCHMARK_CONV(f32_igemm_5x16__avx512f_broadcast)
-  BENCHMARK_CONV(f32_igemm_6x16__avx512f_broadcast)
-  BENCHMARK_CONV(f32_igemm_7x16__avx512f_broadcast)
-  BENCHMARK_CONV(f32_igemm_8x16__avx512f_broadcast)
 #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
 
 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD