xref: /aosp_15_r20/external/XNNPACK/src/s8-ibilinear/wasmsimd-mul32.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Copyright 2021 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert DATATYPE in ["S8", "U8"]
7$assert CHANNEL_TILE % 8 == 0
8$assert CHANNEL_TILE >= 8
9$assert PIXEL_TILE == 1
10$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
11#include <assert.h>
12
13#include <wasm_simd128.h>
14
15#include <xnnpack/common.h>
16#include <xnnpack/ibilinear.h>
17
18
19$XINT8_T = {"S8": "int8_t", "U8": "uint8_t"}[DATATYPE]
20$WASM_X16X8_LOAD_8X8 = {"S8": "wasm_i16x8_load8x8", "U8": "wasm_u16x8_load8x8"}[DATATYPE]
21$WASM_X32X4_SHR = {"S8": "wasm_i32x4_shr", "U8": "wasm_u32x4_shr"}[DATATYPE]
22$WASM_X8X16_NARROW_I16X8 = {"S8": "wasm_i8x16_narrow_i16x8", "U8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
23void xnn_${DATATYPE.lower()}_ibilinear_ukernel__wasmsimd_mul32_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}(
24    size_t output_pixels,
25    size_t channels,
26    const ${XINT8_T}**restrict input,
27    size_t input_offset,
28    const int16_t*restrict weights,
29    ${XINT8_T}*restrict output,
30    size_t output_increment) XNN_OOB_READS
31{
32  assert(output_pixels != 0);
33  assert(channels != 0);
34
35  do {
36    const ${XINT8_T}* i0 = (const ${XINT8_T}*) ((uintptr_t) input[0] + input_offset);
37    const ${XINT8_T}* i1 = (const ${XINT8_T}*) ((uintptr_t) input[1] + input_offset);
38    const ${XINT8_T}* i2 = (const ${XINT8_T}*) ((uintptr_t) input[2] + input_offset);
39    const ${XINT8_T}* i3 = (const ${XINT8_T}*) ((uintptr_t) input[3] + input_offset);
40    input += 4;
41
42    const v128_t valphah = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights));
43    const v128_t valphav = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights + 1));
44    weights += 2;
45
46    const v128_t vrounding = wasm_i32x4_const_splat(0x00200000);
47
48    size_t c = channels;
49    $if CHANNEL_TILE > 8:
50      for (; c >= ${CHANNEL_TILE} * sizeof(${XINT8_T}); c -= ${CHANNEL_TILE} * sizeof(${XINT8_T})) {
51        const v128_t vtl${ABC[0:8]} = ${WASM_X16X8_LOAD_8X8}(i0);
52        const v128_t vtr${ABC[0:8]} = ${WASM_X16X8_LOAD_8X8}(i1);
53        const v128_t vbl${ABC[0:8]} = ${WASM_X16X8_LOAD_8X8}(i2);
54        const v128_t vbr${ABC[0:8]} = ${WASM_X16X8_LOAD_8X8}(i3);
55        $for C in range(8, CHANNEL_TILE, 8):
56          const v128_t vtl${ABC[C:C+8]} = ${WASM_X16X8_LOAD_8X8}(i0 + ${C});
57          const v128_t vtr${ABC[C:C+8]} = ${WASM_X16X8_LOAD_8X8}(i1 + ${C});
58          const v128_t vbl${ABC[C:C+8]} = ${WASM_X16X8_LOAD_8X8}(i2 + ${C});
59          const v128_t vbr${ABC[C:C+8]} = ${WASM_X16X8_LOAD_8X8}(i3 + ${C});
60        i0 += ${CHANNEL_TILE};
61        i1 += ${CHANNEL_TILE};
62        i2 += ${CHANNEL_TILE};
63        i3 += ${CHANNEL_TILE};
64
65        $for C in range(0, CHANNEL_TILE, 8):
66          const v128_t vtd${ABC[C:C+8]} = wasm_i16x8_sub(vtr${ABC[C:C+8]}, vtl${ABC[C:C+8]});
67          const v128_t vbd${ABC[C:C+8]} = wasm_i16x8_sub(vbr${ABC[C:C+8]}, vbl${ABC[C:C+8]});
68          const v128_t vdl${ABC[C:C+8]} = wasm_i16x8_sub(vbl${ABC[C:C+8]}, vtl${ABC[C:C+8]});
69          const v128_t vdd${ABC[C:C+8]} = wasm_i16x8_sub(vbd${ABC[C:C+8]}, vtd${ABC[C:C+8]});
70
71        $for C in range(0, CHANNEL_TILE, 8):
72          const v128_t vt${ABC[C:C+4]} = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl${ABC[C:C+8]}), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd${ABC[C:C+8]}), valphah));
73          const v128_t vt${ABC[C+4:C+8]} = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl${ABC[C:C+8]}), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd${ABC[C:C+8]}), valphah));
74          const v128_t vd${ABC[C:C+4]} = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl${ABC[C:C+8]}), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd${ABC[C:C+8]}), valphah));
75          const v128_t vd${ABC[C+4:C+8]} = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl${ABC[C:C+8]}), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd${ABC[C:C+8]}), valphah));
76
77        $for C in range(0, CHANNEL_TILE, 4):
78          v128_t vacc${ABC[C:C+4]} = wasm_i32x4_mul(vd${ABC[C:C+4]}, valphav);
79
80        $for C in range(0, CHANNEL_TILE, 4):
81          vacc${ABC[C:C+4]} = wasm_i32x4_add(wasm_i32x4_shl(vt${ABC[C:C+4]}, 11), vacc${ABC[C:C+4]});
82
83        $for C in range(0, CHANNEL_TILE, 4):
84          vacc${ABC[C:C+4]} = ${WASM_X32X4_SHR}(wasm_i16x8_add(vacc${ABC[C:C+4]}, vrounding), 22);
85
86        $for C in range(0, CHANNEL_TILE, 8):
87          const v128_t vacc${ABC[C:C+8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[C:C+4]}, vacc${ABC[C+4:C+8]});
88
89        $for C in range(0, CHANNEL_TILE, 16):
90          $if C + 8 < CHANNEL_TILE:
91            const v128_t vo${ABC[C:C+16]} = ${WASM_X8X16_NARROW_I16X8}(vacc${ABC[C:C+8]}, vacc${ABC[C+8:C+16]});
92          $else:
93            const v128_t vo${ABC[C:C+8]} = ${WASM_X8X16_NARROW_I16X8}(vacc${ABC[C:C+8]}, vacc${ABC[C:C+8]});
94
95        wasm_v128_store(output, vo${ABC[0:16]});
96        $for C in range(16, CHANNEL_TILE, 16):
97          $if C + 8 < CHANNEL_TILE:
98            wasm_v128_store(output + ${C}, vo${ABC[C:C+16]});
99          $else:
100            *((double*) (output + ${C})) = wasm_f64x2_extract_lane(vo${ABC[C:C+8]}, 0);
101        output += ${CHANNEL_TILE};
102      }
103    for (; c >= 8 * sizeof(${XINT8_T}); c -= 8 * sizeof(${XINT8_T})) {
104      const v128_t vtl01234567 = ${WASM_X16X8_LOAD_8X8}(i0);
105      i0 += 8;
106      const v128_t vtr01234567 = ${WASM_X16X8_LOAD_8X8}(i1);
107      i1 += 8;
108      const v128_t vbl01234567 = ${WASM_X16X8_LOAD_8X8}(i2);
109      i2 += 8;
110      const v128_t vbr01234567 = ${WASM_X16X8_LOAD_8X8}(i3);
111      i3 += 8;
112
113      const v128_t vtd01234567 = wasm_i16x8_sub(vtr01234567, vtl01234567);
114      const v128_t vbd01234567 = wasm_i16x8_sub(vbr01234567, vbl01234567);
115      const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
116      const v128_t vdd01234567 = wasm_i16x8_sub(vbd01234567, vtd01234567);
117
118      const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd01234567), valphah));
119      const v128_t vt4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd01234567), valphah));
120      const v128_t vd0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd01234567), valphah));
121      const v128_t vd4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd01234567), valphah));
122
123      v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
124      v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
125
126      vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
127      vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
128
129      vacc0123 = ${WASM_X32X4_SHR}(wasm_i16x8_add(vacc0123, vrounding), 22);
130      vacc4567 = ${WASM_X32X4_SHR}(wasm_i16x8_add(vacc4567, vrounding), 22);
131
132      const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
133
134      const v128_t vo01234567 = ${WASM_X8X16_NARROW_I16X8}(vacc01234567, vacc01234567);
135
136      *((double*) output) = wasm_f64x2_extract_lane(vo01234567, 0);
137      output += 8;
138    }
139    if XNN_UNLIKELY(c != 0) {
140      const v128_t vtl01234567 = ${WASM_X16X8_LOAD_8X8}(i0);
141      const v128_t vtr01234567 = ${WASM_X16X8_LOAD_8X8}(i1);
142      const v128_t vbl01234567 = ${WASM_X16X8_LOAD_8X8}(i2);
143      const v128_t vbr01234567 = ${WASM_X16X8_LOAD_8X8}(i3);
144
145      const v128_t vtd01234567 = wasm_i16x8_sub(vtr01234567, vtl01234567);
146      const v128_t vbd01234567 = wasm_i16x8_sub(vbr01234567, vbl01234567);
147      const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
148      const v128_t vdd01234567 = wasm_i16x8_sub(vbd01234567, vtd01234567);
149
150      const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd01234567), valphah));
151      const v128_t vt4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd01234567), valphah));
152      const v128_t vd0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd01234567), valphah));
153      const v128_t vd4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd01234567), valphah));
154
155      v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
156      v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
157
158      vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
159      vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
160
161      vacc0123 = ${WASM_X32X4_SHR}(wasm_i16x8_add(vacc0123, vrounding), 22);
162      vacc4567 = ${WASM_X32X4_SHR}(wasm_i16x8_add(vacc4567, vrounding), 22);
163
164      const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
165
166      v128_t vo01234567 = ${WASM_X8X16_NARROW_I16X8}(vacc01234567, vacc01234567);
167
168      if (c & (4 * sizeof(${XINT8_T}))) {
169        *((float*) output) = wasm_f32x4_extract_lane(vo01234567, 0);
170        output += 4;
171        vo01234567 = wasm_u64x2_shr(vo01234567, 32);
172      }
173      uint32_t vo0123 = (uint32_t) wasm_i32x4_extract_lane(vo01234567, 0);
174      if (c & (2 * sizeof(${XINT8_T}))) {
175        *((uint16_t*) output) = (uint16_t) vo0123;
176        output += 2;
177        vo0123 >>= 16;
178      }
179      if (c & (1 * sizeof(${XINT8_T}))) {
180        *output++ = (uint8_t) vo0123;
181      }
182    }
183
184    output = (${XINT8_T}*) ((uintptr_t) output + output_increment);
185  } while (--output_pixels != 0);
186}
187