| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/NR2-neon.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
|
| 11 |
#include <assert.h> |
| 12 |
#include <stddef.h> |
| 13 |
#include <stdint.h> |
| 14 |
|
| 15 |
#include <arm_neon.h> |
| 16 |
|
| 17 |
#include "xnnpack/packw.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_x32_packw_gemm_goi_ukernel_x2__neon_ld2lane_u2( |
| 21 |
size_t g, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
size_t nr, |
| 25 |
size_t kr, |
| 26 |
size_t sr, |
| 27 |
const uint32_t* weights, |
| 28 |
const uint32_t* bias, |
| 29 |
const void* scale, |
| 30 |
uint32_t* packed_weights, |
| 31 |
size_t extra_bytes, |
| 32 |
const void* params) |
| 33 |
{ |
| 34 |
assert(g != 0); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(nr == 2); |
| 38 |
assert(kr == 1); |
| 39 |
assert(sr == 1); |
| 40 |
assert(weights != NULL); |
| 41 |
assert(packed_weights != NULL); |
| 42 |
|
| 43 |
uint32x2x2_t v00; |
| 44 |
v00.val[0] = vdup_n_u32(0); |
| 45 |
v00.val[1] = vdup_n_u32(0); |
| 46 |
|
| 47 |
do { |
| 48 |
// NC main loop multiple of 2 |
| 49 |
const uint32_t* w0 = weights; |
| 50 |
size_t n = nc; |
| 51 |
|
| 52 |
for (; n >= 2; n -= 2) { |
| 53 |
if XNN_LIKELY(bias != NULL) { |
| 54 |
uint32x2_t vb0 = vld1_u32(bias); bias += 2; |
| 55 |
vst1_u32(packed_weights, vb0); packed_weights += 2; |
| 56 |
} else { |
| 57 |
const uint32x2_t vzero = vmov_n_u32(0); |
| 58 |
vst1_u32(packed_weights, vzero); packed_weights += 2; |
| 59 |
} |
| 60 |
|
| 61 |
const uint32_t* w1 = w0 + kc; |
| 62 |
|
| 63 |
// KC main loop multiple of 2 |
| 64 |
size_t k = kc; |
| 65 |
for (; k >= 2; k -= 2) { |
| 66 |
v00 = vld2_lane_u32(w0, v00, 0); w0 += 2; |
| 67 |
v00 = vld2_lane_u32(w1, v00, 1); w1 += 2; |
| 68 |
vst1_u32(packed_weights + 0, v00.val[0]); |
| 69 |
vst1_u32(packed_weights + 2, v00.val[1]); |
| 70 |
packed_weights += 4; |
| 71 |
} |
| 72 |
|
| 73 |
// KC remainder |
| 74 |
for (; k != 0; --k) { |
| 75 |
v00.val[0] = vld1_lane_u32(w0, v00.val[0], 0); w0 += 1; |
| 76 |
v00.val[0] = vld1_lane_u32(w1, v00.val[0], 1); w1 += 1; |
| 77 |
vst1_u32(packed_weights + 0, v00.val[0]); |
| 78 |
packed_weights += 2; |
| 79 |
} |
| 80 |
packed_weights = (uint32_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 81 |
w0 = w1; |
| 82 |
} |
| 83 |
|
| 84 |
if XNN_UNLIKELY(n != 0) { |
| 85 |
// NC remainder of 1 |
| 86 |
if XNN_LIKELY(bias != NULL) { |
| 87 |
*packed_weights = *bias++; |
| 88 |
} else { |
| 89 |
const uint32x2_t vzero = vmov_n_u32(0); |
| 90 |
vst1_u32(packed_weights + 0, vzero); |
| 91 |
} |
| 92 |
packed_weights += 2; |
| 93 |
size_t k = kc; |
| 94 |
do { |
| 95 |
*packed_weights = *w0++; |
| 96 |
packed_weights += 2; |
| 97 |
} while (--k); |
| 98 |
packed_weights = (uint32_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 99 |
} |
| 100 |
weights += nc * kc; |
| 101 |
} while (--g != 0); |
| 102 |
} |