src/qs8-dwconv/unipass-wasmsimd-mul16.c.in

// 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.

$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert KERNEL_TILE >= 2
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/dwconv.h>


void xnn_qs8_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__wasmsimd_mul16(
    size_t channels,
    size_t output_width,
    const int8_t** input,
    const void* weights,
    int8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const int8_t* zero,
    const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(channels != 0);
  assert(output_width != 0);

  do {
    $for K in range(KERNEL_TILE):
      const int8_t* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const int8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
      $for C in range(4, CHANNEL_TILE, 4):
        v128_t vacc${ABC[C:C+4]} = wasm_v128_load((const void*) ((uintptr_t) w + ${C} * sizeof(int32_t)));

      $for K in range(KERNEL_TILE):

        $for C in range(0, CHANNEL_TILE, 8):
          $if C == 0:
            const v128_t vi${K}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${K});
          $else:
            const v128_t vi${K}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${K} + ${C});
          const v128_t vk${K}x${ABC[C:C+8]} = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t)));
        i${K} += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 8):
          const v128_t vprod${K}x${ABC[C:C+8]} = wasm_i16x8_mul(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});

        $for C in range(0, CHANNEL_TILE, 8):
          vacc${ABC[C:C+4]} = wasm_i32x4_add(vacc${ABC[C:C+4]}, wasm_i32x4_widen_low_i16x8(vprod${K}x${ABC[C:C+8]}));
          vacc${ABC[C+4:C+8]} = wasm_i32x4_add(vacc${ABC[C+4:C+8]}, wasm_i32x4_widen_high_i16x8(vprod${K}x${ABC[C:C+8]}));

      w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t));

      $for C in range(0, CHANNEL_TILE, 4):
        const v128_t vsign${ABC[C:C+4]} = wasm_i32x4_shr(vacc${ABC[C:C+4]}, 31);

      $for C in range(0, CHANNEL_TILE, 4):
        const v128_t vacc${ABC[C:C+2]} = wasm_v32x4_shuffle(vacc${ABC[C:C+4]}, vsign${ABC[C:C+4]}, 0, 4, 1, 5);
        const v128_t vacc${ABC[C+2:C+4]} = wasm_v32x4_shuffle(vacc${ABC[C:C+4]}, vsign${ABC[C:C+4]}, 2, 6, 3, 7);

      const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
      const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
      $for C in range(0, CHANNEL_TILE, 2):
        const v128_t vprod${ABC[C:C+2]} = wasm_i64x2_add(wasm_i64x2_mul(vacc${ABC[C:C+2]}, vmultiplier), vrounding);

      $for C in range(0, CHANNEL_TILE, 4):
        const v128_t vq31prod${ABC[C:C+4]} = wasm_v32x4_shuffle(vprod${ABC[C:C+2]}, vprod${ABC[C+2:C+4]}, 1, 3, 5, 7);

      const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
      $for C in range(0, CHANNEL_TILE, 4):
        const v128_t vrem${ABC[C:C+4]} = wasm_i32x4_add(wasm_v128_and(vq31prod${ABC[C:C+4]}, vremainder_mask), wasm_i32x4_shr(vq31prod${ABC[C:C+4]}, 31));

      const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
      const int32_t vshift = params->wasmsimd.shift;
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${ABC[C:C+4]}, vshift), wasm_i32x4_gt(vrem${ABC[C:C+4]}, vthreshold));

      const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
      $for C in range(0, CHANNEL_TILE, 8):
        v128_t vout${ABC[C:C+8]} = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc${ABC[C:C+4]}, vacc${ABC[C+4:C+8]}), voutput_zero_point);

      const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
      const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
      $for C in range(0, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          v128_t vout${ABC[C:C+16]} = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout${ABC[C:C+8]}, vout${ABC[C+8:C+16]}), voutput_min), voutput_max);
        $else:
          v128_t vout${ABC[C:C+8]}${ABC[C:C+8]} = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout${ABC[C:C+8]}, vout${ABC[C:C+8]}), voutput_min), voutput_max);

      $if CHANNEL_TILE > 8:
        wasm_v128_store(output, vout${ABC[0:16]});
      $else:
        *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
      $for C in range(16, CHANNEL_TILE, 16):
        $if C + 8 < CHANNEL_TILE:
          wasm_v128_store(output + ${C}, vout${ABC[C:C+16]});
        $else:
          *((double*) (output + ${C})) = wasm_f64x2_extract_lane(vout${ABC[C:C+8]}${ABC[C:C+8]}, 0);
      output += ${CHANNEL_TILE};
    }
    if XNN_UNLIKELY(c != 0) {
      $if CHANNEL_TILE > 8:
        const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t));
      ${"do " if CHANNEL_TILE > 8 else ""}{
        v128_t vacc${ABC[0:4]} = wasm_v128_load(w);
        v128_t vacc${ABC[4:8]} = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));

        $for K in range(KERNEL_TILE):

          const v128_t vi${K}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${K});
          $if CHANNEL_TILE > 8:
            $if K == 0:
              const v128_t vk${K}x${ABC[0:8]} = wasm_i16x8_load_8x8(k);
            $else:
              const v128_t vk${K}x${ABC[0:8]} = wasm_i16x8_load_8x8((const void*) (k + ${K * CHANNEL_TILE}));
          $else:
            const v128_t vk${K}x${ABC[0:8]} = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t)));
          $if CHANNEL_TILE > 8:
            i${K} += 8;

          const v128_t vprod${K}x${ABC[0:8]} = wasm_i16x8_mul(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]});

          vacc${ABC[0:4]} = wasm_i32x4_add(vacc${ABC[0:4]}, wasm_i32x4_widen_low_i16x8(vprod${K}x${ABC[0:8]}));
          vacc${ABC[4:8]} = wasm_i32x4_add(vacc${ABC[4:8]}, wasm_i32x4_widen_high_i16x8(vprod${K}x${ABC[0:8]}));

        $if CHANNEL_TILE > 8:
          w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
          k += 8;

      const v128_t vsign${ABC[0:4]} = wasm_i32x4_shr(vacc${ABC[0:4]}, 31);
      const v128_t vsign${ABC[4:8]} = wasm_i32x4_shr(vacc${ABC[4:8]}, 31);

      const v128_t vacc${ABC[0:2]} = wasm_v32x4_shuffle(vacc${ABC[0:4]}, vsign${ABC[0:4]}, 0, 4, 1, 5);
      const v128_t vacc${ABC[2:4]} = wasm_v32x4_shuffle(vacc${ABC[0:4]}, vsign${ABC[0:4]}, 2, 6, 3, 7);
      const v128_t vacc${ABC[4:6]} = wasm_v32x4_shuffle(vacc${ABC[4:8]}, vsign${ABC[4:8]}, 0, 4, 1, 5);
      const v128_t vacc${ABC[6:8]} = wasm_v32x4_shuffle(vacc${ABC[4:8]}, vsign${ABC[4:8]}, 2, 6, 3, 7);

      const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
      const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
      const v128_t vprod${ABC[0:2]} = wasm_i64x2_add(wasm_i64x2_mul(vacc${ABC[0:2]}, vmultiplier), vrounding);
      const v128_t vprod${ABC[2:4]} = wasm_i64x2_add(wasm_i64x2_mul(vacc${ABC[2:4]}, vmultiplier), vrounding);
      const v128_t vprod${ABC[4:6]} = wasm_i64x2_add(wasm_i64x2_mul(vacc${ABC[4:6]}, vmultiplier), vrounding);
      const v128_t vprod${ABC[6:8]} = wasm_i64x2_add(wasm_i64x2_mul(vacc${ABC[6:8]}, vmultiplier), vrounding);

      const v128_t vq31prod${ABC[0:4]} = wasm_v32x4_shuffle(vprod${ABC[0:2]}, vprod${ABC[2:4]}, 1, 3, 5, 7);
      const v128_t vq31prod${ABC[4:8]} = wasm_v32x4_shuffle(vprod${ABC[4:6]}, vprod${ABC[6:8]}, 1, 3, 5, 7);

      const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
      const v128_t vrem${ABC[0:4]} = wasm_i32x4_add(wasm_v128_and(vq31prod${ABC[0:4]}, vremainder_mask), wasm_i32x4_shr(vq31prod${ABC[0:4]}, 31));
      const v128_t vrem${ABC[4:8]} = wasm_i32x4_add(wasm_v128_and(vq31prod${ABC[4:8]}, vremainder_mask), wasm_i32x4_shr(vq31prod${ABC[4:8]}, 31));

      const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
      const int32_t vshift = params->wasmsimd.shift;
      vacc${ABC[0:4]} = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${ABC[0:4]}, vshift), wasm_i32x4_gt(vrem${ABC[0:4]}, vthreshold));
      vacc${ABC[4:8]} = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod${ABC[4:8]}, vshift), wasm_i32x4_gt(vrem${ABC[4:8]}, vthreshold));

      const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
      v128_t vout${ABC[0:8]} = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);

      const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
      const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
      v128_t vout${ABC[0:8]}${ABC[0:8]} = wasm_i8x16_min(wasm_i8x16_max(wasm_i8x16_narrow_i16x8(vout${ABC[0:8]}, vout${ABC[0:8]}), voutput_min), voutput_max);

      $if CHANNEL_TILE > 8:
        if XNN_LIKELY(c >= 8) {
          *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          output += 8;
          c -= 8;
        } else {
          if (c & 4) {
            *((float*) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
            output += 4;
          }
          if (c & 2) {
            *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
            output += 2;
          }
          if (c & 1) {
            *output = (int8_t) wasm_i8x16_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
            output += 1;
          }
          c = 0;
        }
      $else:
        if (c & 4) {
          *((float*) output) = wasm_f32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
          output += 4;
        }
        if (c & 2) {
          *((uint16_t*) output) = (uint16_t) wasm_i16x8_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
          output += 2;
        }
        if (c & 1) {
          *output = (int8_t) wasm_i8x16_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          output += 1;
        }
      }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
    }

    output = (int8_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}