Copybara import of the project:

build_srcs.bzl

@@ -94,6 +94,7 @@ LOGGING_SRCS = [
 ]
 
 MICROKERNEL_DEFS = [
+    "src/bf16-f32-vcvt/bf16-f32-vcvt.inc",
     "src/f16-avgpool/f16-avgpool-minmax.inc",
     "src/f16-dwconv/f16-dwconv-minmax.inc",
     "src/f16-f32-vcvt/f16-f32-vcvt.inc",
@@ -156,6 +157,7 @@ MICROKERNEL_DEFS = [
     "src/f32-conv-hwc/f32-conv-hwc.inc",
     "src/f32-dwconv/f32-dwconv-minmax.inc",
     "src/f32-dwconv/f32-dwconv.inc",
+    "src/f32-bf16-vcvt/f32-bf16-vcvt.inc",
     "src/f32-f16-vcvt/f32-f16-vcvt.inc",
     "src/f32-maxpool/f32-maxpool-minmax.inc",
     "src/f32-qs8-vcvt/f32-qs8-vcvt.inc",

cmake/gen/scalar_microkernels.cmake

@@ -10,6 +10,7 @@
 
 
 SET(PROD_SCALAR_MICROKERNEL_SRCS
+  src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u2.c
   src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u4.c
   src/f16-qs8-vcvt/gen/f16-qs8-vcvt-scalar-imagic-u4.c
   src/f16-qu8-vcvt/gen/f16-qu8-vcvt-scalar-imagic-u4.c
@@ -25,6 +26,7 @@ SET(PROD_SCALAR_MICROKERNEL_SRCS
   src/f16-vsin/gen/f16-vsin-scalar-rational-3-2-div.c
   src/f32-argmaxpool/f32-argmaxpool-9p8x-scalar-c1.c
   src/f32-avgpool/gen/f32-avgpool-9p-minmax-scalar-u1.c
+  src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u2.c
   src/f32-conv-hwc2chw/f32-conv-hwc2chw-3x3s2p1c3x4-scalar-1x1.c
   src/f32-dwconv/gen/f32-dwconv-3p1c-minmax-scalar-acc2.c
   src/f32-dwconv/gen/f32-dwconv-3p1c-scalar-acc2.c
@@ -248,6 +250,9 @@ SET(PROD_SCALAR_MICROKERNEL_SRCS
 
 SET(NON_PROD_SCALAR_MICROKERNEL_SRCS
   src/bf16-f32-gemm/bf16-f32-gemm-1x4c2-minmax-scalar.c
+  src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u1.c
+  src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u3.c
+  src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u4.c
   src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u1.c
   src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u2.c
   src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u3.c
@@ -273,6 +278,9 @@ SET(NON_PROD_SCALAR_MICROKERNEL_SRCS
   src/f16-rminmax/gen/f16-rminmax-scalar-u3-acc3.c
   src/f16-rminmax/gen/f16-rminmax-scalar-u4-acc2.c
   src/f16-rminmax/gen/f16-rminmax-scalar-u4-acc4.c
+  src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u1.c
+  src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u3.c
+  src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u4.c
   src/f32-conv-hwc/f32-conv-hwc-3x3s2p0p1c3x4-scalar-1x1.c
   src/f32-conv-hwc/f32-conv-hwc-3x3s2p1c3x4-scalar-1x1.c
   src/f32-dwconv/gen/f32-dwconv-3p1c-minmax-scalar.c

gen/scalar_microkernels.bzl

@@ -6,6 +6,7 @@
 #
 
 PROD_SCALAR_MICROKERNEL_SRCS = [
+    "src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u2.c",
     "src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u4.c",
     "src/f16-qs8-vcvt/gen/f16-qs8-vcvt-scalar-imagic-u4.c",
     "src/f16-qu8-vcvt/gen/f16-qu8-vcvt-scalar-imagic-u4.c",
@@ -21,6 +22,7 @@ PROD_SCALAR_MICROKERNEL_SRCS = [
     "src/f16-vsin/gen/f16-vsin-scalar-rational-3-2-div.c",
     "src/f32-argmaxpool/f32-argmaxpool-9p8x-scalar-c1.c",
     "src/f32-avgpool/gen/f32-avgpool-9p-minmax-scalar-u1.c",
+    "src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u2.c",
     "src/f32-conv-hwc2chw/f32-conv-hwc2chw-3x3s2p1c3x4-scalar-1x1.c",
     "src/f32-dwconv/gen/f32-dwconv-3p1c-minmax-scalar-acc2.c",
     "src/f32-dwconv/gen/f32-dwconv-3p1c-scalar-acc2.c",
@@ -245,6 +247,9 @@ PROD_SCALAR_MICROKERNEL_SRCS = [
 
 NON_PROD_SCALAR_MICROKERNEL_SRCS = [
     "src/bf16-f32-gemm/bf16-f32-gemm-1x4c2-minmax-scalar.c",
+    "src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u1.c",
+    "src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u3.c",
+    "src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u4.c",
     "src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u1.c",
     "src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u2.c",
     "src/f16-f32-vcvt/gen/f16-f32-vcvt-scalar-u3.c",
@@ -270,6 +275,9 @@ NON_PROD_SCALAR_MICROKERNEL_SRCS = [
     "src/f16-rminmax/gen/f16-rminmax-scalar-u3-acc3.c",
     "src/f16-rminmax/gen/f16-rminmax-scalar-u4-acc2.c",
     "src/f16-rminmax/gen/f16-rminmax-scalar-u4-acc4.c",
+    "src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u1.c",
+    "src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u3.c",
+    "src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u4.c",
     "src/f32-conv-hwc/f32-conv-hwc-3x3s2p0p1c3x4-scalar-1x1.c",
     "src/f32-conv-hwc/f32-conv-hwc-3x3s2p1c3x4-scalar-1x1.c",
     "src/f32-dwconv/gen/f32-dwconv-3p1c-minmax-scalar.c",

scripts/generate-bf16-f32-vcvt.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+# Copyright 2021 Google LLC
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+#################################### Scalar ###################################
+tools/xngen src/bf16-f32-vcvt/scalar.c.in -D BATCH_TILE=1 -o src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u1.c &
+tools/xngen src/bf16-f32-vcvt/scalar.c.in -D BATCH_TILE=2 -o src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u2.c &
+tools/xngen src/bf16-f32-vcvt/scalar.c.in -D BATCH_TILE=3 -o src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u3.c &
+tools/xngen src/bf16-f32-vcvt/scalar.c.in -D BATCH_TILE=4 -o src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u4.c &
+
+wait

scripts/generate-f32-bf16-vcvt.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+# Copyright 2021 Google LLC
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+#################################### Scalar ###################################
+tools/xngen src/f32-bf16-vcvt/scalar.c.in -D BATCH_TILE=1 -o src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u1.c &
+tools/xngen src/f32-bf16-vcvt/scalar.c.in -D BATCH_TILE=2 -o src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u2.c &
+tools/xngen src/f32-bf16-vcvt/scalar.c.in -D BATCH_TILE=3 -o src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u3.c &
+tools/xngen src/f32-bf16-vcvt/scalar.c.in -D BATCH_TILE=4 -o src/f32-bf16-vcvt/gen/f32-bf16-vcvt-scalar-u4.c &
+
+wait

src/bf16-f32-vcvt/bf16-f32-vcvt.inc

@@ -0,0 +1,10 @@
+// clang-format off
+// Copyright 2026 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+XNN_UKERNEL(xnn_arch_none, xnn_bf16_f32_vcvt_ukernel__scalar_u1, 1, false, xnn_bfloat16, float, void, NULL)
+XNN_UKERNEL(xnn_arch_none, xnn_bf16_f32_vcvt_ukernel__scalar_u2, 2, false, xnn_bfloat16, float, void, NULL)
+XNN_UKERNEL(xnn_arch_none, xnn_bf16_f32_vcvt_ukernel__scalar_u3, 3, false, xnn_bfloat16, float, void, NULL)
+XNN_UKERNEL(xnn_arch_none, xnn_bf16_f32_vcvt_ukernel__scalar_u4, 4, false, xnn_bfloat16, float, void, NULL)

src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u1.c

@@ -0,0 +1,40 @@
+// clang-format off
+// Auto-generated file. Do not edit!
+//   Template: src/bf16-f32-vcvt/scalar.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "src/xnnpack/common.h"
+#include "src/xnnpack/math.h"
+#include "src/xnnpack/vcvt.h"
+
+
+void xnn_bf16_f32_vcvt_ukernel__scalar_u1(
+    size_t batch,
+    const xnn_bfloat16* input,
+    float* output,
+    const void* params)
+{
+  assert(batch != 0);
+  assert(batch % sizeof(xnn_bfloat16) == 0);
+  assert(input != NULL);
+  assert(output != NULL);
+
+  const xnn_bfloat16* i = input;
+  float* o = output;
+  do {
+    const xnn_bfloat16 vh = *i++;
+
+    *o++ = xnn_bfloat16_to_float(vh);
+
+    batch -= sizeof(xnn_bfloat16);
+  } while (batch != 0);
+}

src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u2.c

@@ -0,0 +1,47 @@
+// clang-format off
+// Auto-generated file. Do not edit!
+//   Template: src/bf16-f32-vcvt/scalar.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "src/xnnpack/common.h"
+#include "src/xnnpack/math.h"
+#include "src/xnnpack/vcvt.h"
+
+
+void xnn_bf16_f32_vcvt_ukernel__scalar_u2(
+    size_t batch,
+    const xnn_bfloat16* input,
+    float* output,
+    const void* params)
+{
+  assert(batch != 0);
+  assert(batch % sizeof(xnn_bfloat16) == 0);
+  assert(input != NULL);
+  assert(output != NULL);
+
+  const xnn_bfloat16* i = input;
+  float* o = output;
+  for (; batch >= 2 * sizeof(xnn_bfloat16); batch -= 2 * sizeof(xnn_bfloat16)) {
+    const xnn_bfloat16 vh0 = i[0];
+    const xnn_bfloat16 vh1 = i[1];
+    i += 2;
+
+    o[0] = xnn_bfloat16_to_float(vh0);
+    o[1] = xnn_bfloat16_to_float(vh1);
+    o += 2;
+  }
+  if XNN_UNLIKELY(batch != 0) {
+    const xnn_bfloat16 vh = *i;
+
+    *o = xnn_bfloat16_to_float(vh);
+  }
+}

src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u3.c

@@ -0,0 +1,53 @@
+// clang-format off
+// Auto-generated file. Do not edit!
+//   Template: src/bf16-f32-vcvt/scalar.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "src/xnnpack/common.h"
+#include "src/xnnpack/math.h"
+#include "src/xnnpack/vcvt.h"
+
+
+void xnn_bf16_f32_vcvt_ukernel__scalar_u3(
+    size_t batch,
+    const xnn_bfloat16* input,
+    float* output,
+    const void* params)
+{
+  assert(batch != 0);
+  assert(batch % sizeof(xnn_bfloat16) == 0);
+  assert(input != NULL);
+  assert(output != NULL);
+
+  const xnn_bfloat16* i = input;
+  float* o = output;
+  for (; batch >= 3 * sizeof(xnn_bfloat16); batch -= 3 * sizeof(xnn_bfloat16)) {
+    const xnn_bfloat16 vh0 = i[0];
+    const xnn_bfloat16 vh1 = i[1];
+    const xnn_bfloat16 vh2 = i[2];
+    i += 3;
+
+    o[0] = xnn_bfloat16_to_float(vh0);
+    o[1] = xnn_bfloat16_to_float(vh1);
+    o[2] = xnn_bfloat16_to_float(vh2);
+    o += 3;
+  }
+  if XNN_UNLIKELY(batch != 0) {
+    do {
+      const xnn_bfloat16 vh = *i++;
+
+      *o++ = xnn_bfloat16_to_float(vh);
+
+      batch -= sizeof(xnn_bfloat16);
+    } while (batch != 0);
+  }
+}

src/bf16-f32-vcvt/gen/bf16-f32-vcvt-scalar-u4.c

@@ -0,0 +1,55 @@
+// clang-format off
+// Auto-generated file. Do not edit!
+//   Template: src/bf16-f32-vcvt/scalar.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "src/xnnpack/common.h"
+#include "src/xnnpack/math.h"
+#include "src/xnnpack/vcvt.h"
+
+
+void xnn_bf16_f32_vcvt_ukernel__scalar_u4(
+    size_t batch,
+    const xnn_bfloat16* input,
+    float* output,
+    const void* params)
+{
+  assert(batch != 0);
+  assert(batch % sizeof(xnn_bfloat16) == 0);
+  assert(input != NULL);
+  assert(output != NULL);
+
+  const xnn_bfloat16* i = input;
+  float* o = output;
+  for (; batch >= 4 * sizeof(xnn_bfloat16); batch -= 4 * sizeof(xnn_bfloat16)) {
+    const xnn_bfloat16 vh0 = i[0];
+    const xnn_bfloat16 vh1 = i[1];
+    const xnn_bfloat16 vh2 = i[2];
+    const xnn_bfloat16 vh3 = i[3];
+    i += 4;
+
+    o[0] = xnn_bfloat16_to_float(vh0);
+    o[1] = xnn_bfloat16_to_float(vh1);
+    o[2] = xnn_bfloat16_to_float(vh2);
+    o[3] = xnn_bfloat16_to_float(vh3);
+    o += 4;
+  }
+  if XNN_UNLIKELY(batch != 0) {
+    do {
+      const xnn_bfloat16 vh = *i++;
+
+      *o++ = xnn_bfloat16_to_float(vh);
+
+      batch -= sizeof(xnn_bfloat16);
+    } while (batch != 0);
+  }
+}

src/bf16-f32-vcvt/scalar.c.in

@@ -0,0 +1,63 @@
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+$assert BATCH_TILE >= 1
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "src/xnnpack/common.h"
+#include "src/xnnpack/math.h"
+#include "src/xnnpack/vcvt.h"
+
+
+void xnn_bf16_f32_vcvt_ukernel__scalar_u${BATCH_TILE}(
+    size_t batch,
+    const xnn_bfloat16* input,
+    float* output,
+    const void* params)
+{
+  assert(batch != 0);
+  assert(batch % sizeof(xnn_bfloat16) == 0);
+  assert(input != NULL);
+  assert(output != NULL);
+
+  const xnn_bfloat16* i = input;
+  float* o = output;
+  $if BATCH_TILE > 1:
+    for (; batch >= ${BATCH_TILE} * sizeof(xnn_bfloat16); batch -= ${BATCH_TILE} * sizeof(xnn_bfloat16)) {
+      $for N in range(BATCH_TILE):
+        const xnn_bfloat16 vh${N} = i[${N}];
+      i += ${BATCH_TILE};
+
+      $for N in range(BATCH_TILE):
+        o[${N}] = xnn_bfloat16_to_float(vh${N});
+      o += ${BATCH_TILE};
+    }
+  $if BATCH_TILE == 1:
+    do {
+      const xnn_bfloat16 vh = *i++;
+
+      *o++ = xnn_bfloat16_to_float(vh);
+
+      batch -= sizeof(xnn_bfloat16);
+    } while (batch != 0);
+  $elif BATCH_TILE == 2:
+    if XNN_UNLIKELY(batch != 0) {
+      const xnn_bfloat16 vh = *i;
+
+      *o = xnn_bfloat16_to_float(vh);
+    }
+  $else:
+    if XNN_UNLIKELY(batch != 0) {
+      do {
+        const xnn_bfloat16 vh = *i++;
+
+        *o++ = xnn_bfloat16_to_float(vh);
+
+        batch -= sizeof(xnn_bfloat16);
+      } while (batch != 0);
+    }
+}