src/f32-dwconv2d-chw/3x3s2p1-wasmsimd-loadsplat.c.in - platform/external/XNNPACK - Git at Google

 // 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.

 $assert ROW_TILE >= 1
 $assert ACCUMULATORS >= 1
 #include <assert.h>

 #include <wasm_simd128.h>

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


 $ARCH_SUFFIX = "_x86" if X86 else "_arm"

 void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd${ARCH_SUFFIX}_loadsplat_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
     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)]) XNN_OOB_READS
 {
   assert(input_height != 0);
   assert(input_width != 0);
   assert(input_width % sizeof(float) == 0);
   assert(padding_top >= 0);
   assert(padding_top <= 1);

   const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even);
   const v128_t vmask_odd  = wasm_v128_load(params->scalar.mask_odd);
   const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
   const v128_t vmin = wasm_v128_load32_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_v128_load64_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 size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float));
   $if ROW_TILE > 1:
     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;
   }
   $for M in range(2, 1 + 2 * ROW_TILE):
     const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);

   float* o0 = output;
   $for M in range(1, ROW_TILE):
     float* o${M} = (float*) ((uintptr_t) o${M-1} + 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 {
     $for M in range(2, 1 + 2 * ROW_TILE):
       if XNN_UNPREDICTABLE(padded_input_height < ${2 + M}) {
         i${M} = zero;
         $if M % 2 == 1:
           o${(M - 1) // 2} = o${(M - 1) // 2 - 1};
       }

     $for M in range(1 + 2 * ROW_TILE):
       v128_t vi${M}x1357 = wasm_f32x4_const_splat(0.0f);

     size_t w = input_width;
     for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) {
       $for M in range(ROW_TILE):
         v128_t vo${M}p0 = vbias;

       $for M in range(1 + 2 * ROW_TILE):
         const v128_t vi${M}x89AB = wasm_v128_load(i${M});
         const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4);
         i${M} += 8;

       $for M in range(1 + 2 * ROW_TILE):
         const v128_t vi${M}x8ACE = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6);
         const v128_t vi${M}x9BDF = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7);

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 1:
           v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, vk01);
         $else:
           vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, vk01));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 2:
           v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11);
         $else:
           vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 3:
           v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21);
         $else:
           vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21));

       $for M in range(1 + 2 * ROW_TILE):
         const v128_t vi${M}x7BDF = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6);
         vi${M}x1357 = vi${M}x9BDF;

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 4:
           v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M}x7BDF, vk00);
         $else:
           vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x7BDF, vk00));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 5:
           v128_t vo${M}p5 = wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10);
         $else:
           vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 6:
           v128_t vo${M}p6 = wasm_f32x4_mul(vi${2*M+2}x7BDF, vk11);
         $else:
           vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x7BDF, vk20));

       $for M in range(ROW_TILE):
         vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, vk02));

       $for M in range(ROW_TILE):
         vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, vk12));

       $for M in range(ROW_TILE):
         vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, vk22));

       $if ACCUMULATORS > 1:
         $ACC_SLICE = 1
         $while ACC_SLICE < ACCUMULATORS:
           $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
             $if A + ACC_SLICE < ACCUMULATORS:
               $for M in range(ROW_TILE):
                 vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
           $ACC_SLICE *= 2

       $if X86:
         $for M in range(ROW_TILE):
           v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0);
         $for M in range(ROW_TILE):
           vo${M} = wasm_f32x4_pmin(vmax, vo${M});
       $else:
         $for M in range(ROW_TILE):
           v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
         $for M in range(ROW_TILE):
           vo${M} = wasm_f32x4_min(vo${M}, vmax);

       $for M in reversed(range(ROW_TILE)):
         wasm_v128_store(o${M}, vo${M}); o${M} += 4;
     }
     // Last block has 0-7 pixels to process.
     assert(w < 8 * sizeof(float));
     if XNN_LIKELY(w != 0) {
       $for M in range(ROW_TILE):
         v128_t vo${M}p0 = vbias;

       $for M in range(1 + 2 * ROW_TILE):
         const v128_t vi${M}x89AB = wasm_v128_load(i${M});
         const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4);

       $for M in range(1 + 2 * ROW_TILE):
         const v128_t vi${M}x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6));
         const v128_t vi${M}x9BDF = wasm_v128_and(vmask_odd,  wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 1:
           v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, vk01);
         $else:
           vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, vk01));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 2:
           v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11);
         $else:
           vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 3:
           v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21);
         $else:
           vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21));

       $for M in range(1 + 2 * ROW_TILE):
         const v128_t vi${M}x7BDF = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6);

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 4:
           v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M}x7BDF, vk00);
         $else:
           vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x7BDF, vk00));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 5:
           v128_t vo${M}p5 = wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10);
         $else:
           vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10));

       $for M in range(ROW_TILE):
         $if ACCUMULATORS > 6:
           v128_t vo${M}p6 = wasm_f32x4_mul(vi${2*M+2}x7BDF, vk11);
         $else:
           vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x7BDF, vk20));

       $for M in range(ROW_TILE):
         vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, vk02));

       $for M in range(ROW_TILE):
         vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, vk12));

       $for M in range(ROW_TILE):
         vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, vk22));

       $if ACCUMULATORS > 1:
         $ACC_SLICE = 1
         $while ACC_SLICE < ACCUMULATORS:
           $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
             $if A + ACC_SLICE < ACCUMULATORS:
               $for M in range(ROW_TILE):
                 vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
           $ACC_SLICE *= 2

       $if X86:
         $for M in range(ROW_TILE):
           v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0);
         $for M in range(ROW_TILE):
           vo${M} = wasm_f32x4_pmin(vmax, vo${M});
       $else:
         $for M in range(ROW_TILE):
           v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
         $for M in range(ROW_TILE):
           vo${M} = wasm_f32x4_min(vo${M}, vmax);

       w += 1 * sizeof(float);
       if (w & (8 * sizeof(float))) {
         $for M in reversed(range(ROW_TILE)):
           wasm_v128_store(o${M}, vo${M}); o${M} += 4;
       } else {
         if (w & (4 * sizeof(float))) {
           $for M in reversed(range(ROW_TILE)):
             *((double*) o${M}) = wasm_f64x2_extract_lane(vo${M}, 0); o${M} += 2;

           $for M in range(ROW_TILE):
             vo${M} = wasm_v32x4_shuffle(vo${M}, vo${M}, 2, 3, 0, 1);
         }
         if (w & (2 * sizeof(float))) {
           $for M in reversed(range(ROW_TILE)):
             *o${M} = wasm_f32x4_extract_lane(vo${M}, 0); o${M} += 1;
         }
       }
     }

     i0 = (const float*) ((uintptr_t) i${2 * ROW_TILE} - input_decrement);
     $for M in range(1, 1 + 2 * ROW_TILE):
       i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);

     $if ROW_TILE > 1:
       o0 = o${ROW_TILE - 1};
       $for M in range(1, ROW_TILE):
         o${M} = (float*) ((uintptr_t) o${M-1} + output_width);

     $if ROW_TILE > 1:
       output_height = doz(output_height, ${ROW_TILE});
       padded_input_height = doz(padded_input_height, ${ROW_TILE * 2});
     $else:
       output_height -= 1;
       padded_input_height -= 2;
   } while (output_height != 0);
 }
	// 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.

	$assert ROW_TILE >= 1
	$assert ACCUMULATORS >= 1
	#include <assert.h>

	#include <wasm_simd128.h>

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


	$ARCH_SUFFIX = "_x86" if X86 else "_arm"

	void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd${ARCH_SUFFIX}_loadsplat_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
	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)]) XNN_OOB_READS
	{
	assert(input_height != 0);
	assert(input_width != 0);
	assert(input_width % sizeof(float) == 0);
	assert(padding_top >= 0);
	assert(padding_top <= 1);

	const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even);
	const v128_t vmask_odd = wasm_v128_load(params->scalar.mask_odd);
	const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
	const v128_t vmin = wasm_v128_load32_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_v128_load64_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 size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width / 2 /* subsampling / sizeof(float));
	$if ROW_TILE > 1:
	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;
	}
	$for M in range(2, 1 + 2 * ROW_TILE):
	const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);

	float* o0 = output;
	$for M in range(1, ROW_TILE):
	float* o${M} = (float*) ((uintptr_t) o${M-1} + 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 {
	$for M in range(2, 1 + 2 * ROW_TILE):
	if XNN_UNPREDICTABLE(padded_input_height < ${2 + M}) {
	i${M} = zero;
	$if M % 2 == 1:
	o${(M - 1) // 2} = o${(M - 1) // 2 - 1};
	}

	$for M in range(1 + 2 * ROW_TILE):
	v128_t vi${M}x1357 = wasm_f32x4_const_splat(0.0f);

	size_t w = input_width;
	for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) {
	$for M in range(ROW_TILE):
	v128_t vo${M}p0 = vbias;

	$for M in range(1 + 2 * ROW_TILE):
	const v128_t vi${M}x89AB = wasm_v128_load(i${M});
	const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4);
	i${M} += 8;

	$for M in range(1 + 2 * ROW_TILE):
	const v128_t vi${M}x8ACE = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6);
	const v128_t vi${M}x9BDF = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7);

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 1:
	v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, vk01);
	$else:
	vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, vk01));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 2:
	v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11);
	$else:
	vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 3:
	v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21);
	$else:
	vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21));

	$for M in range(1 + 2 * ROW_TILE):
	const v128_t vi${M}x7BDF = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6);
	vi${M}x1357 = vi${M}x9BDF;

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 4:
	v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M}x7BDF, vk00);
	$else:
	vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x7BDF, vk00));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 5:
	v128_t vo${M}p5 = wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10);
	$else:
	vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 6:
	v128_t vo${M}p6 = wasm_f32x4_mul(vi${2*M+2}x7BDF, vk11);
	$else:
	vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x7BDF, vk20));

	$for M in range(ROW_TILE):
	vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, vk02));

	$for M in range(ROW_TILE):
	vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, vk12));

	$for M in range(ROW_TILE):
	vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, vk22));

	$if ACCUMULATORS > 1:
	$ACC_SLICE = 1
	$while ACC_SLICE < ACCUMULATORS:
	$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
	$if A + ACC_SLICE < ACCUMULATORS:
	$for M in range(ROW_TILE):
	vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
	$ACC_SLICE *= 2

	$if X86:
	$for M in range(ROW_TILE):
	v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0);
	$for M in range(ROW_TILE):
	vo${M} = wasm_f32x4_pmin(vmax, vo${M});
	$else:
	$for M in range(ROW_TILE):
	v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
	$for M in range(ROW_TILE):
	vo${M} = wasm_f32x4_min(vo${M}, vmax);

	$for M in reversed(range(ROW_TILE)):
	wasm_v128_store(o${M}, vo${M}); o${M} += 4;
	}
	// Last block has 0-7 pixels to process.
	assert(w < 8 * sizeof(float));
	if XNN_LIKELY(w != 0) {
	$for M in range(ROW_TILE):
	v128_t vo${M}p0 = vbias;

	$for M in range(1 + 2 * ROW_TILE):
	const v128_t vi${M}x89AB = wasm_v128_load(i${M});
	const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4);

	$for M in range(1 + 2 * ROW_TILE):
	const v128_t vi${M}x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6));
	const v128_t vi${M}x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 1:
	v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, vk01);
	$else:
	vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, vk01));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 2:
	v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11);
	$else:
	vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, vk11));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 3:
	v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21);
	$else:
	vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, vk21));

	$for M in range(1 + 2 * ROW_TILE):
	const v128_t vi${M}x7BDF = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6);

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 4:
	v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M}x7BDF, vk00);
	$else:
	vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x7BDF, vk00));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 5:
	v128_t vo${M}p5 = wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10);
	$else:
	vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x7BDF, vk10));

	$for M in range(ROW_TILE):
	$if ACCUMULATORS > 6:
	v128_t vo${M}p6 = wasm_f32x4_mul(vi${2*M+2}x7BDF, vk11);
	$else:
	vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x7BDF, vk20));

	$for M in range(ROW_TILE):
	vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, vk02));

	$for M in range(ROW_TILE):
	vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, vk12));

	$for M in range(ROW_TILE):
	vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, vk22));

	$if ACCUMULATORS > 1:
	$ACC_SLICE = 1
	$while ACC_SLICE < ACCUMULATORS:
	$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
	$if A + ACC_SLICE < ACCUMULATORS:
	$for M in range(ROW_TILE):
	vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
	$ACC_SLICE *= 2

	$if X86:
	$for M in range(ROW_TILE):
	v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0);
	$for M in range(ROW_TILE):
	vo${M} = wasm_f32x4_pmin(vmax, vo${M});
	$else:
	$for M in range(ROW_TILE):
	v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
	$for M in range(ROW_TILE):
	vo${M} = wasm_f32x4_min(vo${M}, vmax);

	w += 1 * sizeof(float);
	if (w & (8 * sizeof(float))) {
	$for M in reversed(range(ROW_TILE)):
	wasm_v128_store(o${M}, vo${M}); o${M} += 4;
	} else {
	if (w & (4 * sizeof(float))) {
	$for M in reversed(range(ROW_TILE)):
	((double) o${M}) = wasm_f64x2_extract_lane(vo${M}, 0); o${M} += 2;

	$for M in range(ROW_TILE):
	vo${M} = wasm_v32x4_shuffle(vo${M}, vo${M}, 2, 3, 0, 1);
	}
	if (w & (2 * sizeof(float))) {
	$for M in reversed(range(ROW_TILE)):
	*o${M} = wasm_f32x4_extract_lane(vo${M}, 0); o${M} += 1;
	}
	}
	}

	i0 = (const float) ((uintptr_t) i${2 ROW_TILE} - input_decrement);
	$for M in range(1, 1 + 2 * ROW_TILE):
	i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);

	$if ROW_TILE > 1:
	o0 = o${ROW_TILE - 1};
	$for M in range(1, ROW_TILE):
	o${M} = (float*) ((uintptr_t) o${M-1} + output_width);

	$if ROW_TILE > 1:
	output_height = doz(output_height, ${ROW_TILE});
	padded_input_height = doz(padded_input_height, ${ROW_TILE * 2});
	$else:
	output_height -= 1;
	padded_input_height -= 2;
	} while (output_height != 0);
	}