// Copyright 2021 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 DATATYPE in ["QS8", "QU8"]
$assert REQUANTIZATION == "FP32"
$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/vmul.h>


$PARAMS_STRUCT = REQUANTIZATION.lower() + "_wasmsimd"
$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$WASM_X16X8_LOAD8X8 = {"QS8": "wasm_i16x8_load8x8", "QU8": "wasm_u16x8_load8x8"}[DATATYPE]
$WASM_X32X4_EXTEND_LOW_X16X8 = {"QS8": "wasm_i32x4_extend_low_i16x8", "QU8": "wasm_u32x4_extend_low_u16x8"}[DATATYPE]
$WASM_X32X4_EXTEND_HIGH_X16X8 = {"QS8": "wasm_i32x4_extend_high_i16x8", "QU8": "wasm_u32x4_extend_high_u16x8"}[DATATYPE]
$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
$WASM_X8X16_MIN = {"QS8": "wasm_i8x16_min", "QU8": "wasm_u8x16_min"}[DATATYPE]
void xnn_${DATATYPE.lower()}_vmul_minmax_${REQUANTIZATION.lower()}_ukernel__wasmsimd_mul32_ld64_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* input_a,
    const ${XINT8_T}* input_b,
    ${XINT8_T}* output,
    const union xnn_${DATATYPE.lower()}_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS

{
  const v128_t va_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.a_zero_point);
  const v128_t vb_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.b_zero_point);
  const v128_t vscale = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.scale);
  const v128_t vmagic_bias = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias);
  const v128_t vmagic_min = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_min);
  const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
  const v128_t voutput_max = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.output_max);

  for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
    const v128_t va${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_a);
    const v128_t vb${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_b);
    $for N in range(8, BATCH_TILE, 8):
      const v128_t va${ABC[N:N+8]} = ${WASM_X16X8_LOAD8X8}(input_a + ${N});
      const v128_t vb${ABC[N:N+8]} = ${WASM_X16X8_LOAD8X8}(input_b + ${N});
    input_a += ${BATCH_TILE};
    input_b += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 8):
      const v128_t vxa${ABC[N:N+8]} = wasm_i16x8_sub(va${ABC[N:N+8]}, va_zero_point);
      const v128_t vxb${ABC[N:N+8]} = wasm_i16x8_sub(vb${ABC[N:N+8]}, vb_zero_point);

    $for N in range(0, BATCH_TILE, 8):
      v128_t vacc${ABC[N:N+4]} = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa${ABC[N:N+8]}), wasm_i32x4_extend_low_i16x8(vxb${ABC[N:N+8]}));
      v128_t vacc${ABC[N+4:N+8]} = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa${ABC[N:N+8]}), wasm_i32x4_extend_high_i16x8(vxb${ABC[N:N+8]}));

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = wasm_f32x4_convert_i32x4(vacc${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = wasm_f32x4_mul(vacc${ABC[N:N+4]}, vscale);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = wasm_f32x4_add(vacc${ABC[N:N+4]}, vmagic_bias);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = wasm_i32x4_max(vacc${ABC[N:N+4]}, vmagic_min);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = wasm_i32x4_sub(vacc${ABC[N:N+4]}, vmagic_bias_less_output_zero_point);

    $for N in range(0, BATCH_TILE, 8):
      v128_t vout${ABC[N:N+8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]});

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        v128_t vout${ABC[N:N+16]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
      $else:
        v128_t vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = ${WASM_X8X16_MIN}(vout${ABC[N:N+16]}, voutput_max);
      $else:
        vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_MIN}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);

    $if BATCH_TILE >= 16:
      wasm_v128_store(output, vout${ABC[0:16]});
    $else:
      *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
    $for N in range(16, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        wasm_v128_store(output + ${N}, vout${ABC[N:N+16]});
      $else:
        *((double*) output) = wasm_f64x2_extract_lane(output + ${N}, vout${ABC[N:N+8]}${ABC[N:N+8]});
    output += ${BATCH_TILE};
  }
  if XNN_UNLIKELY(n != 0) {
    ${"do " if BATCH_TILE > 8 else ""}{
      const v128_t va${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_a);
      const v128_t vb${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_b);
      $if BATCH_TILE > 8:
        input_a += 8;
        input_b += 8;

      const v128_t vxa${ABC[0:8]} = wasm_i16x8_sub(va${ABC[0:8]}, va_zero_point);
      const v128_t vxb${ABC[0:8]} = wasm_i16x8_sub(vb${ABC[0:8]}, vb_zero_point);

      v128_t vacc${ABC[0:4]} = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa${ABC[0:8]}), wasm_i32x4_extend_low_i16x8(vxb${ABC[0:8]}));
      v128_t vacc${ABC[4:8]} = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa${ABC[0:8]}), wasm_i32x4_extend_high_i16x8(vxb${ABC[0:8]}));

      vacc${ABC[0:4]} = wasm_f32x4_convert_i32x4(vacc${ABC[0:4]});
      vacc${ABC[4:8]} = wasm_f32x4_convert_i32x4(vacc${ABC[4:8]});

      vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale);
      vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale);

      vacc${ABC[0:4]} = wasm_f32x4_add(vacc${ABC[0:4]}, vmagic_bias);
      vacc${ABC[4:8]} = wasm_f32x4_add(vacc${ABC[4:8]}, vmagic_bias);

      vacc${ABC[0:4]} = wasm_i32x4_max(vacc${ABC[0:4]}, vmagic_min);
      vacc${ABC[4:8]} = wasm_i32x4_max(vacc${ABC[4:8]}, vmagic_min);

      vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
      vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);

      v128_t vout${ABC[0:8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]});
      v128_t vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[0:8]}, vout${ABC[0:8]});
      vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

      $if BATCH_TILE > 8:
        if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) {
          *((double*) output) = wasm_f64x2_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          output += 8;
          n -= 8 * sizeof(${XINT8_T});
        } else {
          if (n & (4 * sizeof(${XINT8_T}))) {
            *((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;
          }
          uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
          if (n & (2 * sizeof(${XINT8_T}))) {
            *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]};
            vout${ABC[0:4]} >>= 16;
            output += 2;
          }
          if (n & (1 * sizeof(${XINT8_T}))) {
            *output = (${XINT8_T}) vout${ABC[0:4]};
          }
          n = 0;
        }
      $else:
        if (n & (4 * sizeof(${XINT8_T}))) {
          *((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;
        }
        uint32_t vout${ABC[0:4]} = wasm_i32x4_extract_lane(vout${ABC[0:8]}${ABC[0:8]}, 0);
        if (n & (2 * sizeof(${XINT8_T}))) {
          *((uint16_t*) output) = (uint16_t) vout${ABC[0:4]};
          vout${ABC[0:4]} >>= 16;
          output += 2;
        }
        if (n & (1 * sizeof(${XINT8_T}))) {
          *output = (${XINT8_T}) vout${ABC[0:4]};
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}