src/f32-bilinear/psimd.c.in

// 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.

$assert CHANNEL_TILE % 4 == 0
$assert CHANNEL_TILE >= 4
$assert PIXEL_TILE == 1
$ABC = "0123456789ABCDEFGHIJKLMN"
#include <assert.h>

#include <psimd.h>

#include <xnnpack/bilinear.h>


void xnn_f32_bilinear_ukernel__psimd_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
    size_t output_pixels,
    size_t channels,
    const float**restrict input,
    size_t input_offset,
    const float*restrict weights,
    float*restrict output,
    size_t output_increment)
{
  assert(output_pixels != 0);
  assert(channels != 0);
  assert(channels % sizeof(float) == 0);

  do {
    const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset);
    const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset);
    const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset);
    const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset);
    input += 4;

    const psimd_f32 valphah = psimd_load_splat_f32(weights);
    const psimd_f32 valphav = psimd_load_splat_f32(weights + 1);
    weights += 2;

    size_t c = channels;
    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
      const psimd_f32 vtl${ABC[0:4]} = psimd_load_f32(i0);
      const psimd_f32 vtr${ABC[0:4]} = psimd_load_f32(i1);
      const psimd_f32 vbl${ABC[0:4]} = psimd_load_f32(i2);
      const psimd_f32 vbr${ABC[0:4]} = psimd_load_f32(i3);
      $for C in range(4, CHANNEL_TILE, 4):
        const psimd_f32 vtl${ABC[C:C+4]} = psimd_load_f32(i0 + ${C});
        const psimd_f32 vtr${ABC[C:C+4]} = psimd_load_f32(i1 + ${C});
        const psimd_f32 vbl${ABC[C:C+4]} = psimd_load_f32(i2 + ${C});
        const psimd_f32 vbr${ABC[C:C+4]} = psimd_load_f32(i3 + ${C});
      i0 += ${CHANNEL_TILE};
      i1 += ${CHANNEL_TILE};
      i2 += ${CHANNEL_TILE};
      i3 += ${CHANNEL_TILE};

      $for C in range(0, CHANNEL_TILE, 4):
        const psimd_f32 vtd${ABC[C:C+4]} = psimd_sub_f32(vtr${ABC[C:C+4]}, vtl${ABC[C:C+4]});
        const psimd_f32 vbd${ABC[C:C+4]} = psimd_sub_f32(vbr${ABC[C:C+4]}, vbl${ABC[C:C+4]});

      $for C in range(0, CHANNEL_TILE, 4):
        const psimd_f32 vt${ABC[C:C+4]} = psimd_qfma_f32(vtl${ABC[C:C+4]}, vtd${ABC[C:C+4]}, valphah);
        const psimd_f32 vb${ABC[C:C+4]} = psimd_qfma_f32(vbl${ABC[C:C+4]}, vbd${ABC[C:C+4]}, valphah);

      $for C in range(0, CHANNEL_TILE, 4):
        const psimd_f32 vd${ABC[C:C+4]} = psimd_sub_f32(vb${ABC[C:C+4]}, vt${ABC[C:C+4]});

      $for C in range(0, CHANNEL_TILE, 4):
        const psimd_f32 vo${ABC[C:C+4]} = psimd_qfma_f32(vt${ABC[C:C+4]}, vd${ABC[C:C+4]}, valphav);

      psimd_store_f32(output, vo${ABC[0:4]});
      $for C in range(4, CHANNEL_TILE, 4):
        psimd_store_f32(output + ${C}, vo${ABC[C:C+4]});
      output += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 4:
      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
        const psimd_f32 vtl0123 = psimd_load_f32(i0);
        const psimd_f32 vtr0123 = psimd_load_f32(i1);
        const psimd_f32 vbl0123 = psimd_load_f32(i2);
        const psimd_f32 vbr0123 = psimd_load_f32(i3);
        i0 += 4;
        i1 += 4;
        i2 += 4;
        i3 += 4;

        const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
        const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);

        const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
        const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);

        const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);

        const psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);

        psimd_store_f32(output, vo0123);
        output += 4;
      }
    if XNN_UNLIKELY(c != 0) {
      const psimd_f32 vtl0123 = psimd_load_f32(i0);
      const psimd_f32 vtr0123 = psimd_load_f32(i1);
      const psimd_f32 vbl0123 = psimd_load_f32(i2);
      const psimd_f32 vbr0123 = psimd_load_f32(i3);

      const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
      const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);

      const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
      const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);

      const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);

      psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);

      if (c & (2 * sizeof(float))) {
        psimd_store2_f32(output, vo0123);
        vo0123 = psimd_concat_hi_f32(vo0123, vo0123);
        output += 2;
      }
      if (c & (1 * sizeof(float))) {
        psimd_store1_f32(output, vo0123);
        output += 1;
      }
    }

    output = (float*) ((uintptr_t) output + output_increment);
  } while (--output_pixels != 0);
}