// 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 ["QC8", "QD8", "QS8", "QU8"]
$assert DATATYPE == "QD8" or REQUANTIZATION == "FP32"
$assert DATATYPE != "QD8" or not REQUANTIZATION
$assert VARIANT in ["LD64", "LD128", "EXTENDED"]
$assert MR <= 4
#include <assert.h>

#include <wasm_simd128.h>

#include "xnnpack/gemm.h"
#include "xnnpack/math.h"


$DATATYPE_SPEC = {"QC8": "qs8_qc8w", "QD8": "qd8_f32_qc8w", "QS8": "qs8", "QU8": "qu8"}[DATATYPE]
$LOAD_SUFFIX = {"LD128": "_ld128", "LD64": "_ld64", "EXTENDED": ""}[VARIANT]
$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else ""
$REQUANTIZATION_SPEC = "_" + REQUANTIZATION.lower() if REQUANTIZATION else ""
$PARAMS_STRUCT = REQUANTIZATION.lower() + "_scalar" if REQUANTIZATION else "scalar"
$PARAMS_TYPE = {"QC8": "union xnn_qs8_qc8w_conv_minmax_params", "QD8": "union xnn_f32_minmax_params", "QS8": "union xnn_qs8_conv_minmax_params", "QU8": "union xnn_qu8_conv_minmax_params"}[DATATYPE]
$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
$OUT_T = "float" if DATATYPE == "QD8" else XINT8_T
$WASM_X16X8_LOAD8X8 = "wasm_u16x8_load8x8" if DATATYPE == "QU8" else "wasm_i16x8_load8x8"
$WASM_X8X16_NARROW_I16X8 = "wasm_u8x16_narrow_i16x8" if DATATYPE == "QU8" else "wasm_i8x16_narrow_i16x8"
$WASM_X8X16_MIN = "wasm_u8x16_min" if DATATYPE == "QU8" else "wasm_i8x16_min"
void xnn_${DATATYPE_SPEC}_gemm${GEMM_SUFFIX}_minmax${REQUANTIZATION_SPEC}_ukernel_${MR}x4c2__wasmsimd_dot16x2${LOAD_SUFFIX}(
    size_t mr,
    size_t nc,
    size_t kc,
    const ${XINT8_T}* restrict a,
    size_t a_stride,
    const void* restrict w,
    ${OUT_T}* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if DATATYPE == "QD8":
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)],
      const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
    $else:
      const ${PARAMS_TYPE} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(${XINT8_T}) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 2 * sizeof(${XINT8_T}));
  const ${XINT8_T}* a0 = a;
  ${OUT_T}* c0 = c;
  $for M in range(1, MR):
    const ${XINT8_T}* a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M-1} + a_stride);
    ${OUT_T}* c${M} = (${OUT_T}*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  $if DATATYPE == "QD8":
    const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
    const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
    XNN_FORCE_REALIZATION(vmin);
    XNN_FORCE_REALIZATION(vmax);
  $else:
    $if DATATYPE != "QC8":
      const v128_t vscale = wasm_v128_load32_splat(&params->${PARAMS_STRUCT}.scale);
      XNN_FORCE_REALIZATION(vscale);

    const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
    const int32_t output_min_less_zero_point = (int32_t) params->${PARAMS_STRUCT}.output_min - (int32_t) params->${PARAMS_STRUCT}.output_zero_point;
    const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
    const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->${PARAMS_STRUCT}.output_zero_point);
    const v128_t voutput_max = wasm_i8x16_splat(params->${PARAMS_STRUCT}.output_max);
    XNN_FORCE_REALIZATION(vmagic_bias);
    XNN_FORCE_REALIZATION(vmagic_min);
    XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
    XNN_FORCE_REALIZATION(voutput_max);

  $if DATATYPE == "QU8":
    const v128_t vb_zero_point = wasm_i16x8_splat(params->${PARAMS_STRUCT}.kernel_zero_point);
    XNN_FORCE_REALIZATION(vb_zero_point);

  do {
    $if DATATYPE == "QD8":
      const v128_t vksum0123 = wasm_v128_load(w);
      $for M in range(MR):
        const v128_t vinput_zero_point${M} = wasm_v128_load32_splat(&quantization_params[${M}].zero_point);
      $for M in range(MR):
        v128_t vacc${M}x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point${M});
    $else:
      v128_t vacc0x0123 = wasm_v128_load(w);
      $for M in range(1, MR):
        v128_t vacc${M}x0123 = vacc0x0123;
    w = (const int32_t*) w + 4;

    size_t k = kc;
    while (k >= 8 * sizeof(${XINT8_T})) {
      $for M in range(MR):
        const v128_t vxa${M} = ${WASM_X16X8_LOAD8X8}((const v128_t*) a${M});
        a${M} += 8;

      $if VARIANT == "LD128":
        $for K in range(0, 4, 2):
          $if K == 0:
            const v128_t vb${K}${K+1} = wasm_v128_load(w);
          $else:
            const v128_t vb${K}${K+1} = wasm_v128_load((const ${XINT8_T}*) w + ${K * 8});
          $if DATATYPE == "QU8":
            const v128_t vxb${K} = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb${K}${K+1}), vb_zero_point);
            const v128_t vxb${K+1} = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb${K}${K+1}), vb_zero_point);
          $else:
            const v128_t vxb${K} = wasm_i16x8_extend_low_i8x16(vb${K}${K+1});
            const v128_t vxb${K+1} = wasm_i16x8_extend_high_i8x16(vb${K}${K+1});

          $for M in range(MR):
            vacc${M}x0123 = wasm_i32x4_add(vacc${M}x0123,
              wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa${M}, vxa${M}, ${K}, ${K}, ${K}, ${K}), vxb${K}));

          $for M in range(MR):
            vacc${M}x0123 = wasm_i32x4_add(vacc${M}x0123,
              wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa${M}, vxa${M}, ${K+1}, ${K+1}, ${K+1}, ${K+1}), vxb${K+1}));
      $else:
        $for K in range(4):
          $if VARIANT == "LD64":
            $if DATATYPE == "QU8":
              $if K == 0:
                const v128_t vxb${K} = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
              $else:
                const v128_t vxb${K} = wasm_i16x8_sub(wasm_u16x8_load8x8((const ${XINT8_T}*) w + ${K * 8}), vb_zero_point);
            $else:
              $if K == 0:
                const v128_t vxb${K} = wasm_i16x8_load8x8(w);
              $else:
                const v128_t vxb${K} = wasm_i16x8_load8x8((const ${XINT8_T}*) w + ${K * 8});
          $elif VARIANT == "EXTENDED":
            $if K == 0:
              const v128_t vxb${K} = wasm_v128_load(w);
            $else:
              const v128_t vxb${K} = wasm_v128_load((const int16_t*) w + ${K * 8});

          $for M in range(MR):
            vacc${M}x0123 = wasm_i32x4_add(vacc${M}x0123,
              wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa${M}, vxa${M}, ${K}, ${K}, ${K}, ${K}), vxb${K}));

      $if VARIANT == "EXTENDED":
        w = (const int16_t*) w + 32;
      $else:
        w = (const ${XINT8_T}*) w + 32;
      k -= 8 * sizeof(${XINT8_T});
    }
    if (k != 0) {
      $for M in range(MR):
        const v128_t vxa${M} = ${WASM_X16X8_LOAD8X8}(a${M});
        a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} + k);

      $if VARIANT == "EXTENDED":
        const v128_t vxb0 = wasm_v128_load(w);
        w = (const int16_t*) w + 8;
      $else:
        $if DATATYPE == "QU8":
          const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
        $else:
          const v128_t vxb0 = wasm_i16x8_load8x8(w);
        w = (const ${XINT8_T}*) w + 8;

      $for M in range(MR):
        vacc${M}x0123 = wasm_i32x4_add(vacc${M}x0123,
          wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa${M}, vxa${M}, 0, 0, 0, 0), vxb0));

      if (k > 2 * sizeof(${XINT8_T})) {
        $if VARIANT == "EXTENDED":
          const v128_t vxb1 = wasm_v128_load(w);
          w = (const int16_t*) w + 8;
        $else:
          $if DATATYPE == "QU8":
            const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
          $else:
            const v128_t vxb1 = wasm_i16x8_load8x8(w);
          w = (const ${XINT8_T}*) w + 8;

        $for M in range(MR):
          vacc${M}x0123 = wasm_i32x4_add(vacc${M}x0123,
            wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa${M}, vxa${M}, 1, 1, 1, 1), vxb1));

        if (k > 4 * sizeof(${XINT8_T})) {
          $if VARIANT == "EXTENDED":
            const v128_t vxb2 = wasm_v128_load(w);
            w = (const int16_t*) w + 8;
          $else:
            $if DATATYPE == "QU8":
              const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
            $else:
              const v128_t vxb2 = wasm_i16x8_load8x8(w);
            w = (const ${XINT8_T}*) w + 8;

          $for M in range(MR):
            vacc${M}x0123 = wasm_i32x4_add(vacc${M}x0123,
              wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa${M}, vxa${M}, 2, 2, 2, 2), vxb2));
        }
      }
    }

    $for M in range(MR):
      vacc${M}x0123 = wasm_f32x4_convert_i32x4(vacc${M}x0123);

    $if DATATYPE == "QD8":
      $for M in range(MR):
        const v128_t vinput_scale${M} = wasm_v128_load32_splat(&quantization_params[${M}].inv_scale);

      $for M in range(MR):
        vacc${M}x0123 = wasm_f32x4_mul(vacc${M}x0123, vinput_scale${M});

      const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
      $for M in range(MR):
        vacc${M}x0123 = wasm_f32x4_mul(vacc${M}x0123, vfilter_output_scale0123);
      w = (const float*) w + 4;

      const v128_t vbias0123 = wasm_v128_load(w);
      $for M in range(MR):
        vacc${M}x0123 = wasm_f32x4_add(vacc${M}x0123, vbias0123);
      w = (const float*) w + 4;

      $for M in range(MR):
        vacc${M}x0123 = wasm_f32x4_pmax(vacc${M}x0123, vmin);

      $for M in range(MR):
        vacc${M}x0123 = wasm_f32x4_pmin(vacc${M}x0123, vmax);

      if XNN_LIKELY(nc >= 4) {
        $for M in range(MR):
          wasm_v128_store(c${M}, vacc${M}x0123);

        $for M in range(MR):
          a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);

        $for M in range(MR):
          c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

        nc -= 4;
      } else {
        if (nc & 2) {
          $for M in range(MR):
            wasm_v128_store64_lane(c${M}, vacc${M}x0123, 0);
            vacc${M}x0123 = wasm_v64x2_shuffle(vacc${M}x0123, vacc${M}x0123, 1, 1);
            c${M} += 2;
        }
        if (nc & 1) {
          $for M in range(MR):
            wasm_v128_store32_lane(c${M}, vacc${M}x0123, 0);
        }
        nc = 0;
      }
    $else:
      $if DATATYPE == "QC8":
        const v128_t vscale0123 = wasm_v128_load(w);
        w = (const float*) w + 4;
        $for M in range(MR):
          vacc${M}x0123 = wasm_f32x4_mul(vacc${M}x0123, vscale0123);
      $else:
        $for M in range(MR):
          vacc${M}x0123 = wasm_f32x4_mul(vacc${M}x0123, vscale);

      $for M in range(MR):
        vacc${M}x0123 = wasm_f32x4_add(vacc${M}x0123, vmagic_bias);

      $for M in range(MR):
        vacc${M}x0123 = wasm_i32x4_max(vacc${M}x0123, vmagic_min);

      $for M in range(MR):
        vacc${M}x0123 = wasm_i32x4_sub(vacc${M}x0123, vmagic_bias_less_output_zero_point);

      $for M in range(0, MR, 2):
        v128_t vacc${M}${min(M+1, MR-1)}x0123 = wasm_i16x8_narrow_i32x4(vacc${M}x0123, vacc${min(M+1, MR-1)}x0123);

      $if MR > 2:
        v128_t vacc = ${WASM_X8X16_NARROW_I16X8}(vacc0${min(1, MR-1)}x0123, vacc${min(2, MR-1)}${min(3, MR-1)}x0123);
      $else:
        v128_t vacc = ${WASM_X8X16_NARROW_I16X8}(vacc0${min(1, MR-1)}x0123, vacc0${min(1, MR-1)}x0123);

      vacc = ${WASM_X8X16_MIN}(vacc, voutput_max);

      if (nc >= 4) {
        $for M in range(MR):
          wasm_v128_store32_lane(c${M}, vacc, ${M});

        $for M in range(MR):
          c${M} = (${XINT8_T}*) ((uintptr_t) c${M} + cn_stride);

        $for M in range(MR):
          a${M} = (const ${XINT8_T}*) ((uintptr_t) a${M} - kc);

        nc -= 4;
      } else {
        if (nc & 2) {
          $for M in range(MR):
            wasm_v128_store16_lane(c${M}, vacc, ${2 * M});
            c${M} += 2;

          vacc = wasm_u32x4_shr(vacc, 16);
        }
        if (nc & 1) {
          $for M in range(MR):
            wasm_v128_store8_lane(c${M}, vacc, ${4 * M});
        }

        nc = 0;
      }
  } while (nc != 0);
}