src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c - platform/external/XNNPACK - Gitiles

 // Auto-generated file. Do not edit!
 //   Template: src/f32-dwconv2d-chw/3x3p1-wasmsimd-loadsplat.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 <wasm_simd128.h>


 #include <xnnpack/dwconv.h>
 #include <xnnpack/math.h>


 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4(
     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 == 1);

   const v128_t vmask = wasm_v128_load(params->scalar.mask);
   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);

   const v128_t vw0123 = wasm_v128_load(weights);
   const v128_t vw4567 = wasm_v128_load(weights + 4);
   const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
   const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
   const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
   const v128_t vk01 = wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
   const v128_t vk02 = wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
   const v128_t vk10 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
   const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
   const v128_t vk12 = wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
   const v128_t vk20 = wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
   const v128_t vk21 = wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0);
   const v128_t vk22 = wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1);

   const v128_t vzero = wasm_f32x4_splat(0.0f);

   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));

   const float* i0 = zero;
   const float* i1 = input;
   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 + input_width);
   float* o2 = (float*) ((uintptr_t) o1 + input_width);

   size_t output_height = input_height;
   do {
     if XNN_UNPREDICTABLE(output_height < 2) {
       i2 = zero;
       o1 = o0;
     }
     if XNN_UNPREDICTABLE(output_height < 3) {
       i3 = zero;
       o2 = o1;
     }
     if XNN_UNPREDICTABLE(output_height < 4) {
       i4 = zero;
     }

     v128_t vi0x0123 = vzero;
     v128_t vi1x0123 = vzero;
     v128_t vi2x0123 = vzero;
     v128_t vi3x0123 = vzero;
     v128_t vi4x0123 = vzero;

     v128_t vi0x4567 = wasm_v128_load(i0);
     i0 += 4;
     v128_t vi1x4567 = wasm_v128_load(i1);
     i1 += 4;
     v128_t vi2x4567 = wasm_v128_load(i2);
     i2 += 4;
     v128_t vi3x4567 = wasm_v128_load(i3);
     i3 += 4;
     v128_t vi4x4567 = wasm_v128_load(i4);
     i4 += 4;

     size_t w = input_width;
     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
       const v128_t vi0x89AB = wasm_v128_load(i0);
       i0 += 4;
       const v128_t vi1x89AB = wasm_v128_load(i1);
       i1 += 4;
       const v128_t vi2x89AB = wasm_v128_load(i2);
       i2 += 4;
       const v128_t vi3x89AB = wasm_v128_load(i3);
       i3 += 4;
       const v128_t vi4x89AB = wasm_v128_load(i4);
       i4 += 4;

       v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
       v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
       v128_t vo2p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi2x4567, vk01));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));

       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);

       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk10));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));

       vi0x0123 = vi0x4567;
       vi1x0123 = vi1x4567;
       vi2x0123 = vi2x4567;
       vi3x0123 = vi3x4567;
       vi4x0123 = vi4x4567;

       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);

       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk02));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk22));

       vi0x4567 = vi0x89AB;
       vi1x4567 = vi1x89AB;
       vi2x4567 = vi2x89AB;
       vi3x4567 = vi3x89AB;
       vi4x4567 = vi4x89AB;


       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));

       wasm_v128_store(o2, vo2);
       o2 += 4;
       wasm_v128_store(o1, vo1);
       o1 += 4;
       wasm_v128_store(o0, vo0);
       o0 += 4;
     }
     // Always process the last block of 1..4 pixels.
     assert(w >= 1 * sizeof(float));
     assert(w <= 4 * sizeof(float));
     {
       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
       vi4x4567 = wasm_v128_and(vmask, vi4x4567);

       v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
       v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
       v128_t vo2p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi2x4567, vk01));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));

       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);

       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk10));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));

       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);

       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk02));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));
       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk22));


       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));

       if XNN_LIKELY(w == 4 * sizeof(float)) {
         wasm_v128_store(o2, vo2);
         o2 += 4;
         wasm_v128_store(o1, vo1);
         o1 += 4;
         wasm_v128_store(o0, vo0);
         o0 += 4;
       } else {
         if (w & (2 * sizeof(float))) {
           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0);
           o2 += 2;
           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0);
           o1 += 2;
           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0);
           o0 += 2;

           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
         }
         if (w & (1 * sizeof(float))) {
           *o2 = wasm_f32x4_extract_lane(vo2, 0);
           o2 += 1;
           *o1 = wasm_f32x4_extract_lane(vo1, 0);
           o1 += 1;
           *o0 = wasm_f32x4_extract_lane(vo0, 0);
           o0 += 1;
         }
       }
     }

     i0 = (const float*) ((uintptr_t) i3 - input_decrement);
     i1 = (const float*) ((uintptr_t) i4 - input_decrement);
     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 = o2;
     o1 = (float*) ((uintptr_t) o0 + input_width);
     o2 = (float*) ((uintptr_t) o1 + input_width);

     output_height = doz(output_height, 3);
   } while (output_height != 0);
 }
	// Auto-generated file. Do not edit!
	// Template: src/f32-dwconv2d-chw/3x3p1-wasmsimd-loadsplat.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 <wasm_simd128.h>


	#include <xnnpack/dwconv.h>
	#include <xnnpack/math.h>



	void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4(
	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 == 1);

	const v128_t vmask = wasm_v128_load(params->scalar.mask);
	const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
	const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);

	const v128_t vw0123 = wasm_v128_load(weights);
	const v128_t vw4567 = wasm_v128_load(weights + 4);
	const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
	const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
	const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
	const v128_t vk01 = wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
	const v128_t vk02 = wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
	const v128_t vk10 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
	const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
	const v128_t vk12 = wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
	const v128_t vk20 = wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
	const v128_t vk21 = wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0);
	const v128_t vk22 = wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1);

	const v128_t vzero = wasm_f32x4_splat(0.0f);

	const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));

	const float* i0 = zero;
	const float* i1 = input;
	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 + input_width);
	float* o2 = (float*) ((uintptr_t) o1 + input_width);

	size_t output_height = input_height;
	do {
	if XNN_UNPREDICTABLE(output_height < 2) {
	i2 = zero;
	o1 = o0;
	}
	if XNN_UNPREDICTABLE(output_height < 3) {
	i3 = zero;
	o2 = o1;
	}
	if XNN_UNPREDICTABLE(output_height < 4) {
	i4 = zero;
	}

	v128_t vi0x0123 = vzero;
	v128_t vi1x0123 = vzero;
	v128_t vi2x0123 = vzero;
	v128_t vi3x0123 = vzero;
	v128_t vi4x0123 = vzero;

	v128_t vi0x4567 = wasm_v128_load(i0);
	i0 += 4;
	v128_t vi1x4567 = wasm_v128_load(i1);
	i1 += 4;
	v128_t vi2x4567 = wasm_v128_load(i2);
	i2 += 4;
	v128_t vi3x4567 = wasm_v128_load(i3);
	i3 += 4;
	v128_t vi4x4567 = wasm_v128_load(i4);
	i4 += 4;

	size_t w = input_width;
	for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
	const v128_t vi0x89AB = wasm_v128_load(i0);
	i0 += 4;
	const v128_t vi1x89AB = wasm_v128_load(i1);
	i1 += 4;
	const v128_t vi2x89AB = wasm_v128_load(i2);
	i2 += 4;
	const v128_t vi3x89AB = wasm_v128_load(i3);
	i3 += 4;
	const v128_t vi4x89AB = wasm_v128_load(i4);
	i4 += 4;

	v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
	v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
	v128_t vo2p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi2x4567, vk01));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));

	const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
	const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
	const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
	const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
	const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);

	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk10));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));

	vi0x0123 = vi0x4567;
	vi1x0123 = vi1x4567;
	vi2x0123 = vi2x4567;
	vi3x0123 = vi3x4567;
	vi4x0123 = vi4x4567;

	const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
	const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
	const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
	const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
	const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);

	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk02));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk22));

	vi0x4567 = vi0x89AB;
	vi1x4567 = vi1x89AB;
	vi2x4567 = vi2x89AB;
	vi3x4567 = vi3x89AB;
	vi4x4567 = vi4x89AB;


	v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
	v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
	v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
	vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
	vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
	vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));

	wasm_v128_store(o2, vo2);
	o2 += 4;
	wasm_v128_store(o1, vo1);
	o1 += 4;
	wasm_v128_store(o0, vo0);
	o0 += 4;
	}
	// Always process the last block of 1..4 pixels.
	assert(w >= 1 * sizeof(float));
	assert(w <= 4 * sizeof(float));
	{
	vi0x4567 = wasm_v128_and(vmask, vi0x4567);
	vi1x4567 = wasm_v128_and(vmask, vi1x4567);
	vi2x4567 = wasm_v128_and(vmask, vi2x4567);
	vi3x4567 = wasm_v128_and(vmask, vi3x4567);
	vi4x4567 = wasm_v128_and(vmask, vi4x4567);

	v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
	v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
	v128_t vo2p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi2x4567, vk01));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));

	const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
	const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
	const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
	const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
	const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);

	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk10));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));

	const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
	const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
	const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
	const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
	const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);

	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk02));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));
	vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
	vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
	vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk22));


	v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
	v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
	v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
	vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
	vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
	vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));

	if XNN_LIKELY(w == 4 * sizeof(float)) {
	wasm_v128_store(o2, vo2);
	o2 += 4;
	wasm_v128_store(o1, vo1);
	o1 += 4;
	wasm_v128_store(o0, vo0);
	o0 += 4;
	} else {
	if (w & (2 * sizeof(float))) {
	((double) o2) = wasm_f64x2_extract_lane(vo2, 0);
	o2 += 2;
	((double) o1) = wasm_f64x2_extract_lane(vo1, 0);
	o1 += 2;
	((double) o0) = wasm_f64x2_extract_lane(vo0, 0);
	o0 += 2;

	vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
	vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
	vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
	}
	if (w & (1 * sizeof(float))) {
	*o2 = wasm_f32x4_extract_lane(vo2, 0);
	o2 += 1;
	*o1 = wasm_f32x4_extract_lane(vo1, 0);
	o1 += 1;
	*o0 = wasm_f32x4_extract_lane(vo0, 0);
	o0 += 1;
	}
	}
	}

	i0 = (const float*) ((uintptr_t) i3 - input_decrement);
	i1 = (const float*) ((uintptr_t) i4 - input_decrement);
	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 = o2;
	o1 = (float*) ((uintptr_t) o0 + input_width);
	o2 = (float*) ((uintptr_t) o1 + input_width);

	output_height = doz(output_height, 3);
	} while (output_height != 0);
	}