src/f32-dwconv-chw/5x5s2p2-neonfma-1x4-acc2.c - platform/external/XNNPACK - Gitiles

 // Copyright 2019 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 <arm_neon.h>

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


 void xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma_1x4_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 >= 1);
   assert(padding_top <= 2);

   const uint32x4_t vmask_even = vld1q_u32(params->neon.mask_even);
   const uint32x4_t vmask_odd = vld1q_u32(params->neon.mask_odd);
   const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
   const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);

   const float32x4_t vw0123 = vld1q_f32(weights);
   const float32x4_t vw4567 = vld1q_f32(weights + 4);
   const float32x4_t vw89AB = vld1q_f32(weights + 8);
   const float32x4_t vwCDEF = vld1q_f32(weights + 12);
   const float32x4_t vwGHIJ = vld1q_f32(weights + 16);
   const float32x4_t vwKLMN = vld1q_f32(weights + 20);
   const float32x2_t vwOP   = vld1_f32(weights + 24);

   const uint32_t padding_top_less_1 = padding_top - 1;
   const size_t input_decrement = round_down_po2(input_width - 1 * sizeof(float), 4 * sizeof(float)) + 4 * sizeof(float);

   const float* i0 = zero;
   const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
   if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
     i1 = zero;
   }
   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);

   size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
   size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
   do {
     if XNN_UNPREDICTABLE(padded_input_height <= 6) {
       i4 = zero;
     }
     if XNN_UNPREDICTABLE(padded_input_height < 6) {
       i3 = zero;
     }

     float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
     float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
     float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
     float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
     float32x4_t vi4x0123 = vmovq_n_f32(0.0f);
     float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
     float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
     float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
     float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;

     size_t w = input_width;
     for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
       float32x4_t vo468Ap0 = vdupq_laneq_f32(vw0123, 0);

       float32x4_t vi0x89AB;
       float32x4_t vi1x89AB;
       float32x4_t vi2x89AB;
       float32x4_t vi3x89AB;
       float32x4_t vi4x89AB;

       vi0x89AB = vld1q_f32(i0); i0 += 4;
       vi1x89AB = vld1q_f32(i1); i1 += 4;
       vi2x89AB = vld1q_f32(i2); i2 += 4;
       vi3x89AB = vld1q_f32(i3); i3 += 4;
       vi4x89AB = vld1q_f32(i4); i4 += 4;

       float32x4_t vi0xCDEF;
       float32x4_t vi1xCDEF;
       float32x4_t vi2xCDEF;
       float32x4_t vi3xCDEF;
       float32x4_t vi4xCDEF;

       vi0xCDEF = vld1q_f32(i0); i0 += 4;
       vi1xCDEF = vld1q_f32(i1); i1 += 4;
       vi2xCDEF = vld1q_f32(i2); i2 += 4;
       vi3xCDEF = vld1q_f32(i3); i3 += 4;
       vi4xCDEF = vld1q_f32(i4); i4 += 4;

       float32x4_t vi0x468A = vuzp1q_f32(vi0x4567, vi0x89AB);
       float32x4_t vi0x579B = vuzp2q_f32(vi0x4567, vi0x89AB);
       float32x4_t vi1x468A = vuzp1q_f32(vi1x4567, vi1x89AB);
       float32x4_t vi1x579B = vuzp2q_f32(vi1x4567, vi1x89AB);
       float32x4_t vi2x468A = vuzp1q_f32(vi2x4567, vi2x89AB);
       float32x4_t vi2x579B = vuzp2q_f32(vi2x4567, vi2x89AB);
       float32x4_t vi3x468A = vuzp1q_f32(vi3x4567, vi3x89AB);
       float32x4_t vi3x579B = vuzp2q_f32(vi3x4567, vi3x89AB);
       float32x4_t vi4x468A = vuzp1q_f32(vi4x4567, vi4x89AB);
       float32x4_t vi4x579B = vuzp2q_f32(vi4x4567, vi4x89AB);

       // middle tap
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x468A, vw0123, 3);
       float32x4_t vo468Ap1 = vmulq_laneq_f32(vi1x468A, vw89AB, 0);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x468A, vwCDEF, 1);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x468A, vwGHIJ, 2);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x468A, vwKLMN, 3);

       // one left
       const float32x4_t vi0x3579 = vextq_f32(vi0x0123, vi0x579B, 3);
       const float32x4_t vi1x3579 = vextq_f32(vi1x0123, vi1x579B, 3);
       const float32x4_t vi2x3579 = vextq_f32(vi2x0123, vi2x579B, 3);
       const float32x4_t vi3x3579 = vextq_f32(vi3x0123, vi3x579B, 3);
       const float32x4_t vi4x3579 = vextq_f32(vi4x0123, vi4x579B, 3);

       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x3579, vw0123, 2);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x3579, vw4567, 3);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x3579, vwCDEF, 0);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x3579, vwGHIJ, 1);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x3579, vwKLMN, 2);

       // two left
       // getting the vector to use for the far left tap is annoying
       // as we can't ext anything we currently have to get it.
       // To do this, we get a bit ugly.  Interpret the float 32x4
       // vector as int 64x2. Then left shift by 32.  Interpret
       // again as float 32x4.  Now the right most bits are what we
       // want them to be for the following ext.
       const float32x4_t vi0x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi0x0123), 32));
       const float32x4_t vi1x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi1x0123), 32));
       const float32x4_t vi2x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi2x0123), 32));
       const float32x4_t vi3x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi3x0123), 32));
       const float32x4_t vi4x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi4x0123), 32));

       const float32x4_t vi0x2468 = vextq_f32(vi0x0012, vi0x468A, 3);
       const float32x4_t vi1x2468 = vextq_f32(vi1x0012, vi1x468A, 3);
       const float32x4_t vi2x2468 = vextq_f32(vi2x0012, vi2x468A, 3);
       const float32x4_t vi3x2468 = vextq_f32(vi3x0012, vi3x468A, 3);
       const float32x4_t vi4x2468 = vextq_f32(vi4x0012, vi4x468A, 3);

       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x2468, vw0123, 1);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x2468, vw4567, 2);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x2468, vw89AB, 3);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x2468, vwGHIJ, 0);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x2468, vwKLMN, 1);

       vi0x0123 = vi0x89AB;
       vi1x0123 = vi1x89AB;
       vi2x0123 = vi2x89AB;
       vi3x0123 = vi3x89AB;
       vi4x0123 = vi4x89AB;

       // one right
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x579B, vw4567, 0);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x579B, vw89AB, 1);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x579B, vwCDEF, 2);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x579B, vwGHIJ, 3);
       vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x579B, vwOP, 0);

       // two right
       const float32x4_t vi0x68AC = vextq_f32(vi0x468A, vi0xCDEF, 1);
       const float32x4_t vi1x68AC = vextq_f32(vi1x468A, vi1xCDEF, 1);
       const float32x4_t vi2x68AC = vextq_f32(vi2x468A, vi2xCDEF, 1);
       const float32x4_t vi3x68AC = vextq_f32(vi3x468A, vi3xCDEF, 1);
       const float32x4_t vi4x68AC = vextq_f32(vi4x468A, vi4xCDEF, 1);

       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x68AC, vw4567, 1);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x68AC, vw89AB, 2);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x68AC, vwCDEF, 3);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x68AC, vwKLMN, 0);
       vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x68AC, vwOP, 1);

       vi0x4567 = vi0xCDEF;
       vi1x4567 = vi1xCDEF;
       vi2x4567 = vi2xCDEF;
       vi3x4567 = vi3xCDEF;
       vi4x4567 = vi4xCDEF;

       float32x4_t vo0 = vaddq_f32(vo468Ap0, vo468Ap1);

       vo0 = vmaxq_f32(vo0, vmin);
       vo0 = vminq_f32(vo0, vmax);

       size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
       if XNN_LIKELY(w_tmp >= 4) {
         vst1q_f32(output, vo0); output += 4;
       } else {
         float* output_tmp = output;
         float32x2_t vo0_lo = vget_low_f32(vo0);
         if (w_tmp & 2) {
           vst1_f32(output_tmp, vo0_lo); output_tmp += 2;
           vo0_lo = vget_high_f32(vo0);
         }
         if (w_tmp & 1) {
           vst1_lane_f32(output_tmp, vo0_lo, 0);
         }
       }
     }

     {
       float32x4_t vo468Ap0 = vdupq_laneq_f32(vw0123, 0);

       float32x4_t vi0x89AB;
       float32x4_t vi1x89AB;
       float32x4_t vi2x89AB;
       float32x4_t vi3x89AB;
       float32x4_t vi4x89AB;

       if XNN_LIKELY(w > 4 * sizeof(float)) {
         vi0x89AB = vld1q_f32(i0); i0 += 4;
         vi1x89AB = vld1q_f32(i1); i1 += 4;
         vi2x89AB = vld1q_f32(i2); i2 += 4;
         vi3x89AB = vld1q_f32(i3); i3 += 4;
         vi4x89AB = vld1q_f32(i4); i4 += 4;
       } else {
         vi0x89AB = vmovq_n_f32(0.0f);
         vi1x89AB = vmovq_n_f32(0.0f);
         vi2x89AB = vmovq_n_f32(0.0f);
         vi3x89AB = vmovq_n_f32(0.0f);
         vi4x89AB = vmovq_n_f32(0.0f);
       }

       float32x4_t vi0xCDEF;
       float32x4_t vi1xCDEF;
       float32x4_t vi2xCDEF;
       float32x4_t vi3xCDEF;
       float32x4_t vi4xCDEF;

       if XNN_LIKELY(w > 8 * sizeof(float)) {
         vi0xCDEF = vld1q_f32(i0); i0 += 4;
         vi1xCDEF = vld1q_f32(i1); i1 += 4;
         vi2xCDEF = vld1q_f32(i2); i2 += 4;
         vi3xCDEF = vld1q_f32(i3); i3 += 4;
         vi4xCDEF = vld1q_f32(i4); i4 += 4;
       } else {
         vi0xCDEF = vmovq_n_f32(0.0f);
         vi1xCDEF = vmovq_n_f32(0.0f);
         vi2xCDEF = vmovq_n_f32(0.0f);
         vi3xCDEF = vmovq_n_f32(0.0f);
         vi4xCDEF = vmovq_n_f32(0.0f);
       }
       float32x4_t vi0x468A = vuzp1q_f32(vi0x4567, vi0x89AB);
       float32x4_t vi0x579B = vuzp2q_f32(vi0x4567, vi0x89AB);
       float32x4_t vi1x468A = vuzp1q_f32(vi1x4567, vi1x89AB);
       float32x4_t vi1x579B = vuzp2q_f32(vi1x4567, vi1x89AB);
       float32x4_t vi2x468A = vuzp1q_f32(vi2x4567, vi2x89AB);
       float32x4_t vi2x579B = vuzp2q_f32(vi2x4567, vi2x89AB);
       float32x4_t vi3x468A = vuzp1q_f32(vi3x4567, vi3x89AB);
       float32x4_t vi3x579B = vuzp2q_f32(vi3x4567, vi3x89AB);
       float32x4_t vi4x468A = vuzp1q_f32(vi4x4567, vi4x89AB);
       float32x4_t vi4x579B = vuzp2q_f32(vi4x4567, vi4x89AB);

       vi0x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x468A)));
       vi1x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x468A)));
       vi2x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x468A)));
       vi3x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x468A)));
       vi4x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x468A)));

       vi0x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x579B)));
       vi1x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x579B)));
       vi2x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x579B)));
       vi3x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x579B)));
       vi4x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x579B)));

       // middle tap
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x468A, vw0123, 3);
       float32x4_t vo468Ap1 = vmulq_laneq_f32(vi1x468A, vw89AB, 0);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x468A, vwCDEF, 1);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x468A, vwGHIJ, 2);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x468A, vwKLMN, 3);

       // one left
       const float32x4_t vi0x3579 = vextq_f32(vi0x0123, vi0x579B, 3);
       const float32x4_t vi1x3579 = vextq_f32(vi1x0123, vi1x579B, 3);
       const float32x4_t vi2x3579 = vextq_f32(vi2x0123, vi2x579B, 3);
       const float32x4_t vi3x3579 = vextq_f32(vi3x0123, vi3x579B, 3);
       const float32x4_t vi4x3579 = vextq_f32(vi4x0123, vi4x579B, 3);

       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x3579, vw0123, 2);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x3579, vw4567, 3);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x3579, vwCDEF, 0);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x3579, vwGHIJ, 1);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x3579, vwKLMN, 2);

       // two left
       // getting the vector to use for the far left tap is annoying
       // as we can't ext anything we currently have to get it.
       // To do this, we get a bit ugly.  Interpret the float 32x4
       // vector as int 64x2. Then left shift by 32.  Interpret
       // again as float 32x4.  Now the right most bits are what we
       // want them to be for the following ext.
       const float32x4_t vi0x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi0x0123), 32));
       const float32x4_t vi1x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi1x0123), 32));
       const float32x4_t vi2x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi2x0123), 32));
       const float32x4_t vi3x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi3x0123), 32));
       const float32x4_t vi4x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi4x0123), 32));

       const float32x4_t vi0x2468 = vextq_f32(vi0x0012, vi0x468A, 3);
       const float32x4_t vi1x2468 = vextq_f32(vi1x0012, vi1x468A, 3);
       const float32x4_t vi2x2468 = vextq_f32(vi2x0012, vi2x468A, 3);
       const float32x4_t vi3x2468 = vextq_f32(vi3x0012, vi3x468A, 3);
       const float32x4_t vi4x2468 = vextq_f32(vi4x0012, vi4x468A, 3);

       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x2468, vw0123, 1);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x2468, vw4567, 2);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x2468, vw89AB, 3);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x2468, vwGHIJ, 0);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x2468, vwKLMN, 1);

       vi0x0123 = vi0x89AB;
       vi1x0123 = vi1x89AB;
       vi2x0123 = vi2x89AB;
       vi3x0123 = vi3x89AB;
       vi4x0123 = vi4x89AB;

       // one right
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x579B, vw4567, 0);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x579B, vw89AB, 1);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x579B, vwCDEF, 2);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x579B, vwGHIJ, 3);
       vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x579B, vwOP, 0);

       // two right
       const float32x4_t vi0x68AC = vextq_f32(vi0x468A, vi0xCDEF, 1);
       const float32x4_t vi1x68AC = vextq_f32(vi1x468A, vi1xCDEF, 1);
       const float32x4_t vi2x68AC = vextq_f32(vi2x468A, vi2xCDEF, 1);
       const float32x4_t vi3x68AC = vextq_f32(vi3x468A, vi3xCDEF, 1);
       const float32x4_t vi4x68AC = vextq_f32(vi4x468A, vi4xCDEF, 1);

       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x68AC, vw4567, 1);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x68AC, vw89AB, 2);
       vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x68AC, vwCDEF, 3);
       vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x68AC, vwKLMN, 0);
       vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x68AC, vwOP, 1);

       vi0x4567 = vi0xCDEF;
       vi1x4567 = vi1xCDEF;
       vi2x4567 = vi2xCDEF;
       vi3x4567 = vi3xCDEF;
       vi4x4567 = vi4xCDEF;

       float32x4_t vo0 = vaddq_f32(vo468Ap0, vo468Ap1);

       vo0 = vmaxq_f32(vo0, vmin);
       vo0 = vminq_f32(vo0, vmax);

       size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
       if XNN_LIKELY(w_tmp >= 4) {
         vst1q_f32(output, vo0); output += 4;
       } else {
         float32x2_t vo0_lo = vget_low_f32(vo0);
         if (w_tmp & 2) {
           vst1_f32(output, vo0_lo); output += 2;
           vo0_lo = vget_high_f32(vo0);
         }
         if (w_tmp & 1) {
           vst1_lane_f32(output, vo0_lo, 0); output += 1;
         }
       }
     }

     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
     i2 = (const float*) ((uintptr_t) i4 - input_decrement);
     i3 = (const float*) ((uintptr_t) i2 + input_width);
     i4 = (const float*) ((uintptr_t) i3 + input_width);

     output_height -= 1;
     padded_input_height -= 2;
   } while (output_height != 0);
 }
	// Copyright 2019 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 <arm_neon.h>

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


	void xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma_1x4_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 >= 1);
	assert(padding_top <= 2);

	const uint32x4_t vmask_even = vld1q_u32(params->neon.mask_even);
	const uint32x4_t vmask_odd = vld1q_u32(params->neon.mask_odd);
	const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
	const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);

	const float32x4_t vw0123 = vld1q_f32(weights);
	const float32x4_t vw4567 = vld1q_f32(weights + 4);
	const float32x4_t vw89AB = vld1q_f32(weights + 8);
	const float32x4_t vwCDEF = vld1q_f32(weights + 12);
	const float32x4_t vwGHIJ = vld1q_f32(weights + 16);
	const float32x4_t vwKLMN = vld1q_f32(weights + 20);
	const float32x2_t vwOP = vld1_f32(weights + 24);

	const uint32_t padding_top_less_1 = padding_top - 1;
	const size_t input_decrement = round_down_po2(input_width - 1 * sizeof(float), 4 * sizeof(float)) + 4 * sizeof(float);

	const float* i0 = zero;
	const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
	const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
	if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
	i1 = zero;
	}
	const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
	const float* i4 = (const float*) ((uintptr_t) i3 + input_width);

	size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
	size_t output_height = (padded_input_height - 5 /* kernel size / + 2 / subsampling */) / 2;
	do {
	if XNN_UNPREDICTABLE(padded_input_height <= 6) {
	i4 = zero;
	}
	if XNN_UNPREDICTABLE(padded_input_height < 6) {
	i3 = zero;
	}

	float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
	float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
	float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
	float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
	float32x4_t vi4x0123 = vmovq_n_f32(0.0f);
	float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
	float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
	float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
	float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
	float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;

	size_t w = input_width;
	for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
	float32x4_t vo468Ap0 = vdupq_laneq_f32(vw0123, 0);

	float32x4_t vi0x89AB;
	float32x4_t vi1x89AB;
	float32x4_t vi2x89AB;
	float32x4_t vi3x89AB;
	float32x4_t vi4x89AB;

	vi0x89AB = vld1q_f32(i0); i0 += 4;
	vi1x89AB = vld1q_f32(i1); i1 += 4;
	vi2x89AB = vld1q_f32(i2); i2 += 4;
	vi3x89AB = vld1q_f32(i3); i3 += 4;
	vi4x89AB = vld1q_f32(i4); i4 += 4;

	float32x4_t vi0xCDEF;
	float32x4_t vi1xCDEF;
	float32x4_t vi2xCDEF;
	float32x4_t vi3xCDEF;
	float32x4_t vi4xCDEF;

	vi0xCDEF = vld1q_f32(i0); i0 += 4;
	vi1xCDEF = vld1q_f32(i1); i1 += 4;
	vi2xCDEF = vld1q_f32(i2); i2 += 4;
	vi3xCDEF = vld1q_f32(i3); i3 += 4;
	vi4xCDEF = vld1q_f32(i4); i4 += 4;

	float32x4_t vi0x468A = vuzp1q_f32(vi0x4567, vi0x89AB);
	float32x4_t vi0x579B = vuzp2q_f32(vi0x4567, vi0x89AB);
	float32x4_t vi1x468A = vuzp1q_f32(vi1x4567, vi1x89AB);
	float32x4_t vi1x579B = vuzp2q_f32(vi1x4567, vi1x89AB);
	float32x4_t vi2x468A = vuzp1q_f32(vi2x4567, vi2x89AB);
	float32x4_t vi2x579B = vuzp2q_f32(vi2x4567, vi2x89AB);
	float32x4_t vi3x468A = vuzp1q_f32(vi3x4567, vi3x89AB);
	float32x4_t vi3x579B = vuzp2q_f32(vi3x4567, vi3x89AB);
	float32x4_t vi4x468A = vuzp1q_f32(vi4x4567, vi4x89AB);
	float32x4_t vi4x579B = vuzp2q_f32(vi4x4567, vi4x89AB);

	// middle tap
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x468A, vw0123, 3);
	float32x4_t vo468Ap1 = vmulq_laneq_f32(vi1x468A, vw89AB, 0);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x468A, vwCDEF, 1);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x468A, vwGHIJ, 2);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x468A, vwKLMN, 3);

	// one left
	const float32x4_t vi0x3579 = vextq_f32(vi0x0123, vi0x579B, 3);
	const float32x4_t vi1x3579 = vextq_f32(vi1x0123, vi1x579B, 3);
	const float32x4_t vi2x3579 = vextq_f32(vi2x0123, vi2x579B, 3);
	const float32x4_t vi3x3579 = vextq_f32(vi3x0123, vi3x579B, 3);
	const float32x4_t vi4x3579 = vextq_f32(vi4x0123, vi4x579B, 3);

	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x3579, vw0123, 2);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x3579, vw4567, 3);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x3579, vwCDEF, 0);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x3579, vwGHIJ, 1);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x3579, vwKLMN, 2);

	// two left
	// getting the vector to use for the far left tap is annoying
	// as we can't ext anything we currently have to get it.
	// To do this, we get a bit ugly. Interpret the float 32x4
	// vector as int 64x2. Then left shift by 32. Interpret
	// again as float 32x4. Now the right most bits are what we
	// want them to be for the following ext.
	const float32x4_t vi0x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi0x0123), 32));
	const float32x4_t vi1x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi1x0123), 32));
	const float32x4_t vi2x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi2x0123), 32));
	const float32x4_t vi3x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi3x0123), 32));
	const float32x4_t vi4x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi4x0123), 32));

	const float32x4_t vi0x2468 = vextq_f32(vi0x0012, vi0x468A, 3);
	const float32x4_t vi1x2468 = vextq_f32(vi1x0012, vi1x468A, 3);
	const float32x4_t vi2x2468 = vextq_f32(vi2x0012, vi2x468A, 3);
	const float32x4_t vi3x2468 = vextq_f32(vi3x0012, vi3x468A, 3);
	const float32x4_t vi4x2468 = vextq_f32(vi4x0012, vi4x468A, 3);

	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x2468, vw0123, 1);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x2468, vw4567, 2);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x2468, vw89AB, 3);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x2468, vwGHIJ, 0);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x2468, vwKLMN, 1);

	vi0x0123 = vi0x89AB;
	vi1x0123 = vi1x89AB;
	vi2x0123 = vi2x89AB;
	vi3x0123 = vi3x89AB;
	vi4x0123 = vi4x89AB;

	// one right
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x579B, vw4567, 0);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x579B, vw89AB, 1);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x579B, vwCDEF, 2);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x579B, vwGHIJ, 3);
	vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x579B, vwOP, 0);

	// two right
	const float32x4_t vi0x68AC = vextq_f32(vi0x468A, vi0xCDEF, 1);
	const float32x4_t vi1x68AC = vextq_f32(vi1x468A, vi1xCDEF, 1);
	const float32x4_t vi2x68AC = vextq_f32(vi2x468A, vi2xCDEF, 1);
	const float32x4_t vi3x68AC = vextq_f32(vi3x468A, vi3xCDEF, 1);
	const float32x4_t vi4x68AC = vextq_f32(vi4x468A, vi4xCDEF, 1);

	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x68AC, vw4567, 1);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x68AC, vw89AB, 2);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x68AC, vwCDEF, 3);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x68AC, vwKLMN, 0);
	vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x68AC, vwOP, 1);

	vi0x4567 = vi0xCDEF;
	vi1x4567 = vi1xCDEF;
	vi2x4567 = vi2xCDEF;
	vi3x4567 = vi3xCDEF;
	vi4x4567 = vi4xCDEF;

	float32x4_t vo0 = vaddq_f32(vo468Ap0, vo468Ap1);

	vo0 = vmaxq_f32(vo0, vmin);
	vo0 = vminq_f32(vo0, vmax);

	size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
	if XNN_LIKELY(w_tmp >= 4) {
	vst1q_f32(output, vo0); output += 4;
	} else {
	float* output_tmp = output;
	float32x2_t vo0_lo = vget_low_f32(vo0);
	if (w_tmp & 2) {
	vst1_f32(output_tmp, vo0_lo); output_tmp += 2;
	vo0_lo = vget_high_f32(vo0);
	}
	if (w_tmp & 1) {
	vst1_lane_f32(output_tmp, vo0_lo, 0);
	}
	}
	}

	{
	float32x4_t vo468Ap0 = vdupq_laneq_f32(vw0123, 0);

	float32x4_t vi0x89AB;
	float32x4_t vi1x89AB;
	float32x4_t vi2x89AB;
	float32x4_t vi3x89AB;
	float32x4_t vi4x89AB;

	if XNN_LIKELY(w > 4 * sizeof(float)) {
	vi0x89AB = vld1q_f32(i0); i0 += 4;
	vi1x89AB = vld1q_f32(i1); i1 += 4;
	vi2x89AB = vld1q_f32(i2); i2 += 4;
	vi3x89AB = vld1q_f32(i3); i3 += 4;
	vi4x89AB = vld1q_f32(i4); i4 += 4;
	} else {
	vi0x89AB = vmovq_n_f32(0.0f);
	vi1x89AB = vmovq_n_f32(0.0f);
	vi2x89AB = vmovq_n_f32(0.0f);
	vi3x89AB = vmovq_n_f32(0.0f);
	vi4x89AB = vmovq_n_f32(0.0f);
	}

	float32x4_t vi0xCDEF;
	float32x4_t vi1xCDEF;
	float32x4_t vi2xCDEF;
	float32x4_t vi3xCDEF;
	float32x4_t vi4xCDEF;

	if XNN_LIKELY(w > 8 * sizeof(float)) {
	vi0xCDEF = vld1q_f32(i0); i0 += 4;
	vi1xCDEF = vld1q_f32(i1); i1 += 4;
	vi2xCDEF = vld1q_f32(i2); i2 += 4;
	vi3xCDEF = vld1q_f32(i3); i3 += 4;
	vi4xCDEF = vld1q_f32(i4); i4 += 4;
	} else {
	vi0xCDEF = vmovq_n_f32(0.0f);
	vi1xCDEF = vmovq_n_f32(0.0f);
	vi2xCDEF = vmovq_n_f32(0.0f);
	vi3xCDEF = vmovq_n_f32(0.0f);
	vi4xCDEF = vmovq_n_f32(0.0f);
	}
	float32x4_t vi0x468A = vuzp1q_f32(vi0x4567, vi0x89AB);
	float32x4_t vi0x579B = vuzp2q_f32(vi0x4567, vi0x89AB);
	float32x4_t vi1x468A = vuzp1q_f32(vi1x4567, vi1x89AB);
	float32x4_t vi1x579B = vuzp2q_f32(vi1x4567, vi1x89AB);
	float32x4_t vi2x468A = vuzp1q_f32(vi2x4567, vi2x89AB);
	float32x4_t vi2x579B = vuzp2q_f32(vi2x4567, vi2x89AB);
	float32x4_t vi3x468A = vuzp1q_f32(vi3x4567, vi3x89AB);
	float32x4_t vi3x579B = vuzp2q_f32(vi3x4567, vi3x89AB);
	float32x4_t vi4x468A = vuzp1q_f32(vi4x4567, vi4x89AB);
	float32x4_t vi4x579B = vuzp2q_f32(vi4x4567, vi4x89AB);

	vi0x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x468A)));
	vi1x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x468A)));
	vi2x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x468A)));
	vi3x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x468A)));
	vi4x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x468A)));

	vi0x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x579B)));
	vi1x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x579B)));
	vi2x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x579B)));
	vi3x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x579B)));
	vi4x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x579B)));

	// middle tap
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x468A, vw0123, 3);
	float32x4_t vo468Ap1 = vmulq_laneq_f32(vi1x468A, vw89AB, 0);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x468A, vwCDEF, 1);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x468A, vwGHIJ, 2);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x468A, vwKLMN, 3);

	// one left
	const float32x4_t vi0x3579 = vextq_f32(vi0x0123, vi0x579B, 3);
	const float32x4_t vi1x3579 = vextq_f32(vi1x0123, vi1x579B, 3);
	const float32x4_t vi2x3579 = vextq_f32(vi2x0123, vi2x579B, 3);
	const float32x4_t vi3x3579 = vextq_f32(vi3x0123, vi3x579B, 3);
	const float32x4_t vi4x3579 = vextq_f32(vi4x0123, vi4x579B, 3);

	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x3579, vw0123, 2);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x3579, vw4567, 3);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x3579, vwCDEF, 0);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x3579, vwGHIJ, 1);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x3579, vwKLMN, 2);

	// two left
	// getting the vector to use for the far left tap is annoying
	// as we can't ext anything we currently have to get it.
	// To do this, we get a bit ugly. Interpret the float 32x4
	// vector as int 64x2. Then left shift by 32. Interpret
	// again as float 32x4. Now the right most bits are what we
	// want them to be for the following ext.
	const float32x4_t vi0x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi0x0123), 32));
	const float32x4_t vi1x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi1x0123), 32));
	const float32x4_t vi2x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi2x0123), 32));
	const float32x4_t vi3x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi3x0123), 32));
	const float32x4_t vi4x0012 = vreinterpretq_f32_u64(vshlq_n_u64(vreinterpretq_u64_f32(vi4x0123), 32));

	const float32x4_t vi0x2468 = vextq_f32(vi0x0012, vi0x468A, 3);
	const float32x4_t vi1x2468 = vextq_f32(vi1x0012, vi1x468A, 3);
	const float32x4_t vi2x2468 = vextq_f32(vi2x0012, vi2x468A, 3);
	const float32x4_t vi3x2468 = vextq_f32(vi3x0012, vi3x468A, 3);
	const float32x4_t vi4x2468 = vextq_f32(vi4x0012, vi4x468A, 3);

	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x2468, vw0123, 1);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x2468, vw4567, 2);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x2468, vw89AB, 3);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x2468, vwGHIJ, 0);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi4x2468, vwKLMN, 1);

	vi0x0123 = vi0x89AB;
	vi1x0123 = vi1x89AB;
	vi2x0123 = vi2x89AB;
	vi3x0123 = vi3x89AB;
	vi4x0123 = vi4x89AB;

	// one right
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x579B, vw4567, 0);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x579B, vw89AB, 1);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x579B, vwCDEF, 2);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x579B, vwGHIJ, 3);
	vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x579B, vwOP, 0);

	// two right
	const float32x4_t vi0x68AC = vextq_f32(vi0x468A, vi0xCDEF, 1);
	const float32x4_t vi1x68AC = vextq_f32(vi1x468A, vi1xCDEF, 1);
	const float32x4_t vi2x68AC = vextq_f32(vi2x468A, vi2xCDEF, 1);
	const float32x4_t vi3x68AC = vextq_f32(vi3x468A, vi3xCDEF, 1);
	const float32x4_t vi4x68AC = vextq_f32(vi4x468A, vi4xCDEF, 1);

	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi0x68AC, vw4567, 1);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi1x68AC, vw89AB, 2);
	vo468Ap0 = vfmaq_laneq_f32(vo468Ap0, vi2x68AC, vwCDEF, 3);
	vo468Ap1 = vfmaq_laneq_f32(vo468Ap1, vi3x68AC, vwKLMN, 0);
	vo468Ap0 = vfmaq_lane_f32( vo468Ap0, vi4x68AC, vwOP, 1);

	vi0x4567 = vi0xCDEF;
	vi1x4567 = vi1xCDEF;
	vi2x4567 = vi2xCDEF;
	vi3x4567 = vi3xCDEF;
	vi4x4567 = vi4xCDEF;

	float32x4_t vo0 = vaddq_f32(vo468Ap0, vo468Ap1);

	vo0 = vmaxq_f32(vo0, vmin);
	vo0 = vminq_f32(vo0, vmax);

	size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
	if XNN_LIKELY(w_tmp >= 4) {
	vst1q_f32(output, vo0); output += 4;
	} else {
	float32x2_t vo0_lo = vget_low_f32(vo0);
	if (w_tmp & 2) {
	vst1_f32(output, vo0_lo); output += 2;
	vo0_lo = vget_high_f32(vo0);
	}
	if (w_tmp & 1) {
	vst1_lane_f32(output, vo0_lo, 0); output += 1;
	}
	}
	}

	i0 = (const float*) ((uintptr_t) i2 - input_decrement);
	i1 = (const float*) ((uintptr_t) i3 - input_decrement);
	i2 = (const float*) ((uintptr_t) i4 - input_decrement);
	i3 = (const float*) ((uintptr_t) i2 + input_width);
	i4 = (const float*) ((uintptr_t) i3 + input_width);

	output_height -= 1;
	padded_input_height -= 2;
	} while (output_height != 0);
	}