Auto-generate scalar versions of DWCONV2D CHW 3x3s2p1 micro-kernels
PiperOrigin-RevId: 338962809
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c
new file mode 100644
index 0000000..29c630b
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c
@@ -0,0 +1,136 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+ float* o0 = output;
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+
+ vo0p1 += vi0x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo0p1 += vi2x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo0p1 += vi1x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+
+ vo0p0 += vo0p1;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+
+ vo0p1 += vi0x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo0p1 += vi2x1 * vk21;
+
+ vo0p0 += vo0p1;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i1);
+ i1 = (const float*) ((uintptr_t) i2);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+
+ output_height -= 1;
+ padded_input_height -= 2;
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c
new file mode 100644
index 0000000..7c2992c
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c
@@ -0,0 +1,138 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+ float* o0 = output;
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ float vo0p2 = vi2x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+
+ vo0p0 += vi0x1 * vk01;
+ vo0p1 += vi1x1 * vk11;
+ vo0p2 += vi2x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo0p1 += vi1x2 * vk12;
+ vo0p2 += vi2x2 * vk22;
+
+ vo0p0 += vo0p1;
+ vo0p0 += vo0p2;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ float vo0p2 = vi2x0 * vk20;
+
+ vo0p0 += vi0x1 * vk01;
+ vo0p1 += vi1x1 * vk11;
+ vo0p2 += vi2x1 * vk21;
+
+ vo0p0 += vo0p1;
+ vo0p0 += vo0p2;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i1);
+ i1 = (const float*) ((uintptr_t) i2);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+
+ output_height -= 1;
+ padded_input_height -= 2;
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c
new file mode 100644
index 0000000..fd0e8e4
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c
@@ -0,0 +1,140 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+ float* o0 = output;
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ float vo0p2 = vi2x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+
+ float vo0p3 = vi0x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo0p1 += vi2x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+
+ vo0p2 += vi0x2 * vk02;
+ vo0p3 += vi1x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+
+ vo0p0 += vo0p1;
+ vo0p2 += vo0p3;
+ vo0p0 += vo0p2;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ float vo0p2 = vi2x0 * vk20;
+
+ float vo0p3 = vi0x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo0p1 += vi2x1 * vk21;
+
+ vo0p0 += vo0p1;
+ vo0p2 += vo0p3;
+ vo0p0 += vo0p2;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i1);
+ i1 = (const float*) ((uintptr_t) i2);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+
+ output_height -= 1;
+ padded_input_height -= 2;
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c
new file mode 100644
index 0000000..a1bc2e6
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c
@@ -0,0 +1,134 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+ float* o0 = output;
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+
+ vo0p0 += vi0x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo0p0 += vi1x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+
+ vo0p0 += vi0x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i1);
+ i1 = (const float*) ((uintptr_t) i2);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+
+
+ output_height -= 1;
+ padded_input_height -= 2;
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c
new file mode 100644
index 0000000..a440c7c
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c
@@ -0,0 +1,186 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+ const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+ const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
+ const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
+
+ float* o0 = output;
+ float* o1 = (float*) ((uintptr_t) o0 + output_width);
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 5) {
+ i3 = zero;
+ o1 = o0;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 6) {
+ i4 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+ float vi3x0 = 0.0f;
+ float vi4x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+ const float vi3x1 = i3[0];
+ const float vi4x1 = i4[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ float vo1p1 = vi3x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+ const float vi3x2 = i3[1];
+ i3 += 2;
+ const float vi4x2 = i4[1];
+ i4 += 2;
+
+ vo0p1 += vi0x1 * vk01;
+ vo1p1 += vi2x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo0p1 += vi2x1 * vk21;
+ vo1p1 += vi4x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+ vi3x0 = vi3x2;
+ vi4x0 = vi4x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo1p0 += vi2x2 * vk02;
+ vo0p1 += vi1x2 * vk12;
+ vo1p1 += vi3x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+ vo1p0 += vi4x2 * vk22;
+
+ vo0p0 += vo0p1;
+ vo1p0 += vo1p1;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+ const float vi3x1 = *i3++;
+ const float vi4x1 = *i4++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ float vo0p1 = vi1x0 * vk10;
+ float vo1p1 = vi3x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+
+ vo0p1 += vi0x1 * vk01;
+ vo1p1 += vi2x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo0p1 += vi2x1 * vk21;
+ vo1p1 += vi4x1 * vk21;
+
+ vo0p0 += vo0p1;
+ vo1p0 += vo1p1;
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i3);
+ i1 = (const float*) ((uintptr_t) i4);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+ i3 = (const float*) ((uintptr_t) i2 + input_width);
+ i4 = (const float*) ((uintptr_t) i3 + input_width);
+
+ o0 = o1;
+ o1 = (float*) ((uintptr_t) o0 + output_width);
+
+ output_height = doz(output_height, 2);
+ padded_input_height = doz(padded_input_height, 4);
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c
new file mode 100644
index 0000000..1e7df0d
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c
@@ -0,0 +1,182 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+ const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+ const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
+ const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
+
+ float* o0 = output;
+ float* o1 = (float*) ((uintptr_t) o0 + output_width);
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 5) {
+ i3 = zero;
+ o1 = o0;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 6) {
+ i4 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+ float vi3x0 = 0.0f;
+ float vi4x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+ const float vi3x1 = i3[0];
+ const float vi4x1 = i4[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo1p0 += vi3x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+ const float vi3x2 = i3[1];
+ i3 += 2;
+ const float vi4x2 = i4[1];
+ i4 += 2;
+
+ vo0p0 += vi0x1 * vk01;
+ vo1p0 += vi2x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+ vo1p0 += vi4x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+ vi3x0 = vi3x2;
+ vi4x0 = vi4x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo1p0 += vi2x2 * vk02;
+ vo0p0 += vi1x2 * vk12;
+ vo1p0 += vi3x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+ vo1p0 += vi4x2 * vk22;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+ const float vi3x1 = *i3++;
+ const float vi4x1 = *i4++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo1p0 += vi3x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+
+ vo0p0 += vi0x1 * vk01;
+ vo1p0 += vi2x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+ vo1p0 += vi4x1 * vk21;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i3);
+ i1 = (const float*) ((uintptr_t) i4);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+ i3 = (const float*) ((uintptr_t) i2 + input_width);
+ i4 = (const float*) ((uintptr_t) i3 + input_width);
+
+ o0 = o1;
+ o1 = (float*) ((uintptr_t) o0 + output_width);
+
+ output_height = doz(output_height, 2);
+ padded_input_height = doz(padded_input_height, 4);
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c
new file mode 100644
index 0000000..d096ce1
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c
@@ -0,0 +1,228 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+ const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+ const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
+ const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
+ const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
+ const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
+
+ float* o0 = output;
+ float* o1 = (float*) ((uintptr_t) o0 + output_width);
+ float* o2 = (float*) ((uintptr_t) o1 + output_width);
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 5) {
+ i3 = zero;
+ o1 = o0;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 6) {
+ i4 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 7) {
+ i5 = zero;
+ o2 = o1;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 8) {
+ i6 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+ float vi3x0 = 0.0f;
+ float vi4x0 = 0.0f;
+ float vi5x0 = 0.0f;
+ float vi6x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+ const float vi3x1 = i3[0];
+ const float vi4x1 = i4[0];
+ const float vi5x1 = i5[0];
+ const float vi6x1 = i6[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ float vo2p0 = vbias + vi4x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo1p0 += vi3x0 * vk10;
+ vo2p0 += vi5x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+ vo2p0 += vi6x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+ const float vi3x2 = i3[1];
+ i3 += 2;
+ const float vi4x2 = i4[1];
+ i4 += 2;
+ const float vi5x2 = i5[1];
+ i5 += 2;
+ const float vi6x2 = i6[1];
+ i6 += 2;
+
+ vo0p0 += vi0x1 * vk01;
+ vo1p0 += vi2x1 * vk01;
+ vo2p0 += vi4x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo2p0 += vi5x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+ vo1p0 += vi4x1 * vk21;
+ vo2p0 += vi6x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+ vi3x0 = vi3x2;
+ vi4x0 = vi4x2;
+ vi5x0 = vi5x2;
+ vi6x0 = vi6x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo1p0 += vi2x2 * vk02;
+ vo2p0 += vi4x2 * vk02;
+ vo0p0 += vi1x2 * vk12;
+ vo1p0 += vi3x2 * vk12;
+ vo2p0 += vi5x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+ vo1p0 += vi4x2 * vk22;
+ vo2p0 += vi6x2 * vk22;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+ float vo2 = math_max_f32(vo2p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+ vo2 = math_min_f32(vo2, vmax);
+
+ *o2++ = vo2;
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+ const float vi3x1 = *i3++;
+ const float vi4x1 = *i4++;
+ const float vi5x1 = *i5++;
+ const float vi6x1 = *i6++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ float vo2p0 = vbias + vi4x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo1p0 += vi3x0 * vk10;
+ vo2p0 += vi5x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+ vo2p0 += vi6x0 * vk20;
+
+ vo0p0 += vi0x1 * vk01;
+ vo1p0 += vi2x1 * vk01;
+ vo2p0 += vi4x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo2p0 += vi5x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+ vo1p0 += vi4x1 * vk21;
+ vo2p0 += vi6x1 * vk21;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+ float vo2 = math_max_f32(vo2p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+ vo2 = math_min_f32(vo2, vmax);
+
+ *o2++ = vo2;
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i5);
+ i1 = (const float*) ((uintptr_t) i6);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+ i3 = (const float*) ((uintptr_t) i2 + input_width);
+ i4 = (const float*) ((uintptr_t) i3 + input_width);
+ i5 = (const float*) ((uintptr_t) i4 + input_width);
+ i6 = (const float*) ((uintptr_t) i5 + input_width);
+
+ o0 = o2;
+ o1 = (float*) ((uintptr_t) o0 + output_width);
+ o2 = (float*) ((uintptr_t) o1 + output_width);
+
+ output_height = doz(output_height, 3);
+ padded_input_height = doz(padded_input_height, 6);
+ } while (output_height != 0);
+}
diff --git a/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c
new file mode 100644
index 0000000..4163ba2
--- /dev/null
+++ b/src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c
@@ -0,0 +1,274 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1(
+ size_t input_height,
+ size_t input_width,
+ const float* input,
+ const float* weights,
+ const float* zero,
+ float* output,
+ uint32_t padding_top,
+ const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
+{
+ assert(input_height != 0);
+ assert(input_width != 0);
+ assert(input_width % sizeof(float) == 0);
+ assert(padding_top >= 0);
+ assert(padding_top <= 1);
+
+ const float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+
+ const float vbias = weights[0];
+ const float vk00 = weights[1];
+ const float vk01 = weights[2];
+ const float vk02 = weights[3];
+ const float vk10 = weights[4];
+ const float vk11 = weights[5];
+ const float vk12 = weights[6];
+ const float vk20 = weights[7];
+ const float vk21 = weights[8];
+ const float vk22 = weights[9];
+
+ const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
+
+ const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
+ const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
+ if XNN_UNPREDICTABLE(padding_top != 0) {
+ i0 = zero;
+ }
+ const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
+ const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
+ const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
+ const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
+ const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
+ const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
+ const float* i8 = (const float*) ((uintptr_t) i7 + input_width);
+
+ float* o0 = output;
+ float* o1 = (float*) ((uintptr_t) o0 + output_width);
+ float* o2 = (float*) ((uintptr_t) o1 + output_width);
+ float* o3 = (float*) ((uintptr_t) o2 + output_width);
+
+ size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
+ size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
+ do {
+ if XNN_UNPREDICTABLE(padded_input_height < 4) {
+ i2 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 5) {
+ i3 = zero;
+ o1 = o0;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 6) {
+ i4 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 7) {
+ i5 = zero;
+ o2 = o1;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 8) {
+ i6 = zero;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 9) {
+ i7 = zero;
+ o3 = o2;
+ }
+ if XNN_UNPREDICTABLE(padded_input_height < 10) {
+ i8 = zero;
+ }
+
+ float vi0x0 = 0.0f;
+ float vi1x0 = 0.0f;
+ float vi2x0 = 0.0f;
+ float vi3x0 = 0.0f;
+ float vi4x0 = 0.0f;
+ float vi5x0 = 0.0f;
+ float vi6x0 = 0.0f;
+ float vi7x0 = 0.0f;
+ float vi8x0 = 0.0f;
+
+ size_t w = input_width;
+ for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) {
+ const float vi0x1 = i0[0];
+ const float vi1x1 = i1[0];
+ const float vi2x1 = i2[0];
+ const float vi3x1 = i3[0];
+ const float vi4x1 = i4[0];
+ const float vi5x1 = i5[0];
+ const float vi6x1 = i6[0];
+ const float vi7x1 = i7[0];
+ const float vi8x1 = i8[0];
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ float vo2p0 = vbias + vi4x0 * vk00;
+ float vo3p0 = vbias + vi6x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo1p0 += vi3x0 * vk10;
+ vo2p0 += vi5x0 * vk10;
+ vo3p0 += vi7x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+ vo2p0 += vi6x0 * vk20;
+ vo3p0 += vi8x0 * vk20;
+
+ const float vi0x2 = i0[1];
+ i0 += 2;
+ const float vi1x2 = i1[1];
+ i1 += 2;
+ const float vi2x2 = i2[1];
+ i2 += 2;
+ const float vi3x2 = i3[1];
+ i3 += 2;
+ const float vi4x2 = i4[1];
+ i4 += 2;
+ const float vi5x2 = i5[1];
+ i5 += 2;
+ const float vi6x2 = i6[1];
+ i6 += 2;
+ const float vi7x2 = i7[1];
+ i7 += 2;
+ const float vi8x2 = i8[1];
+ i8 += 2;
+
+ vo0p0 += vi0x1 * vk01;
+ vo1p0 += vi2x1 * vk01;
+ vo2p0 += vi4x1 * vk01;
+ vo3p0 += vi6x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo2p0 += vi5x1 * vk11;
+ vo3p0 += vi7x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+ vo1p0 += vi4x1 * vk21;
+ vo2p0 += vi6x1 * vk21;
+ vo3p0 += vi8x1 * vk21;
+
+ vi0x0 = vi0x2;
+ vi1x0 = vi1x2;
+ vi2x0 = vi2x2;
+ vi3x0 = vi3x2;
+ vi4x0 = vi4x2;
+ vi5x0 = vi5x2;
+ vi6x0 = vi6x2;
+ vi7x0 = vi7x2;
+ vi8x0 = vi8x2;
+
+ vo0p0 += vi0x2 * vk02;
+ vo1p0 += vi2x2 * vk02;
+ vo2p0 += vi4x2 * vk02;
+ vo3p0 += vi6x2 * vk02;
+ vo0p0 += vi1x2 * vk12;
+ vo1p0 += vi3x2 * vk12;
+ vo2p0 += vi5x2 * vk12;
+ vo3p0 += vi7x2 * vk12;
+ vo0p0 += vi2x2 * vk22;
+ vo1p0 += vi4x2 * vk22;
+ vo2p0 += vi6x2 * vk22;
+ vo3p0 += vi8x2 * vk22;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+ float vo2 = math_max_f32(vo2p0, vmin);
+ float vo3 = math_max_f32(vo3p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+ vo2 = math_min_f32(vo2, vmax);
+ vo3 = math_min_f32(vo3, vmax);
+
+ *o3++ = vo3;
+ *o2++ = vo2;
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+ // Potentially process the last pixel.
+ assert(w <= 1 * sizeof(float));
+ if (w != 0) {
+ const float vi0x1 = *i0++;
+ const float vi1x1 = *i1++;
+ const float vi2x1 = *i2++;
+ const float vi3x1 = *i3++;
+ const float vi4x1 = *i4++;
+ const float vi5x1 = *i5++;
+ const float vi6x1 = *i6++;
+ const float vi7x1 = *i7++;
+ const float vi8x1 = *i8++;
+
+ float vo0p0 = vbias + vi0x0 * vk00;
+ float vo1p0 = vbias + vi2x0 * vk00;
+ float vo2p0 = vbias + vi4x0 * vk00;
+ float vo3p0 = vbias + vi6x0 * vk00;
+ vo0p0 += vi1x0 * vk10;
+ vo1p0 += vi3x0 * vk10;
+ vo2p0 += vi5x0 * vk10;
+ vo3p0 += vi7x0 * vk10;
+ vo0p0 += vi2x0 * vk20;
+ vo1p0 += vi4x0 * vk20;
+ vo2p0 += vi6x0 * vk20;
+ vo3p0 += vi8x0 * vk20;
+
+ vo0p0 += vi0x1 * vk01;
+ vo1p0 += vi2x1 * vk01;
+ vo2p0 += vi4x1 * vk01;
+ vo3p0 += vi6x1 * vk01;
+ vo0p0 += vi1x1 * vk11;
+ vo1p0 += vi3x1 * vk11;
+ vo2p0 += vi5x1 * vk11;
+ vo3p0 += vi7x1 * vk11;
+ vo0p0 += vi2x1 * vk21;
+ vo1p0 += vi4x1 * vk21;
+ vo2p0 += vi6x1 * vk21;
+ vo3p0 += vi8x1 * vk21;
+
+
+ float vo0 = math_max_f32(vo0p0, vmin);
+ float vo1 = math_max_f32(vo1p0, vmin);
+ float vo2 = math_max_f32(vo2p0, vmin);
+ float vo3 = math_max_f32(vo3p0, vmin);
+
+ vo0 = math_min_f32(vo0, vmax);
+ vo1 = math_min_f32(vo1, vmax);
+ vo2 = math_min_f32(vo2, vmax);
+ vo3 = math_min_f32(vo3, vmax);
+
+ *o3++ = vo3;
+ *o2++ = vo2;
+ *o1++ = vo1;
+ *o0++ = vo0;
+ }
+
+ i0 = (const float*) ((uintptr_t) i7);
+ i1 = (const float*) ((uintptr_t) i8);
+ i2 = (const float*) ((uintptr_t) i1 + input_width);
+ i3 = (const float*) ((uintptr_t) i2 + input_width);
+ i4 = (const float*) ((uintptr_t) i3 + input_width);
+ i5 = (const float*) ((uintptr_t) i4 + input_width);
+ i6 = (const float*) ((uintptr_t) i5 + input_width);
+ i7 = (const float*) ((uintptr_t) i6 + input_width);
+ i8 = (const float*) ((uintptr_t) i7 + input_width);
+
+ o0 = o3;
+ o1 = (float*) ((uintptr_t) o0 + output_width);
+ o2 = (float*) ((uintptr_t) o1 + output_width);
+ o3 = (float*) ((uintptr_t) o2 + output_width);
+
+ output_height = doz(output_height, 4);
+ padded_input_height = doz(padded_input_height, 8);
+ } while (output_height != 0);
+}