Commit 05057e5635e5ef7fb5be3156f477866cce98fbdb
1 parent
438283c1
Exists in
master
and in
1 other branch
Eliminate unnecessary VTRNs in SPLIT(16,4) NEON implementation
Also makes the ARMv8 version consistent with the older one, in terms of processing width
Showing
1 changed file
with
22 additions
and
79 deletions
Show diff stats
src/neon/gf_w16_neon.c
| ... | ... | @@ -46,7 +46,11 @@ |
| 46 | 46 | #include <stdlib.h> |
| 47 | 47 | #include "gf_w16.h" |
| 48 | 48 | |
| 49 | -#ifdef ARCH_AARCH64 | |
| 49 | +#ifndef ARCH_AARCH64 | |
| 50 | +#define vqtbl1q_u8(tbl, v) vcombine_u8(vtbl2_u8(tbl, vget_low_u8(v)), \ | |
| 51 | + vtbl2_u8(tbl, vget_high_u8(v))) | |
| 52 | +#endif | |
| 53 | + | |
| 50 | 54 | static |
| 51 | 55 | inline |
| 52 | 56 | void |
| ... | ... | @@ -56,23 +60,29 @@ neon_w16_split_4_multiply_region(gf_t *gf, uint16_t *src, uint16_t *dst, |
| 56 | 60 | { |
| 57 | 61 | unsigned i; |
| 58 | 62 | uint8_t *high = tbl + 4 * 16; |
| 59 | - uint16x8_t va0, va1, r0, r1; | |
| 60 | 63 | uint8x16_t loset, rl, rh; |
| 61 | 64 | uint8x16x2_t va; |
| 62 | 65 | |
| 66 | +#ifdef ARCH_AARCH64 | |
| 63 | 67 | uint8x16_t tbl_h[4], tbl_l[4]; |
| 64 | 68 | for (i = 0; i < 4; i++) { |
| 65 | 69 | tbl_l[i] = vld1q_u8(tbl + i*16); |
| 66 | 70 | tbl_h[i] = vld1q_u8(high + i*16); |
| 67 | 71 | } |
| 72 | +#else | |
| 73 | + uint8x8x2_t tbl_h[4], tbl_l[4]; | |
| 74 | + for (i = 0; i < 4; i++) { | |
| 75 | + tbl_l[i].val[0] = vld1_u8(tbl + i*16); | |
| 76 | + tbl_l[i].val[1] = vld1_u8(tbl + i*16 + 8); | |
| 77 | + tbl_h[i].val[0] = vld1_u8(high + i*16); | |
| 78 | + tbl_h[i].val[1] = vld1_u8(high + i*16 + 8); | |
| 79 | + } | |
| 80 | +#endif | |
| 68 | 81 | |
| 69 | 82 | loset = vdupq_n_u8(0xf); |
| 70 | 83 | |
| 71 | 84 | while (dst < d_end) { |
| 72 | - va0 = vld1q_u16(src); | |
| 73 | - va1 = vld1q_u16(src + 8); | |
| 74 | - | |
| 75 | - va = vtrnq_u8(vreinterpretq_u8_u16(va0), vreinterpretq_u8_u16(va1)); | |
| 85 | + va = vld2q_u8((uint8_t*)src); | |
| 76 | 86 | |
| 77 | 87 | rl = vqtbl1q_u8(tbl_l[0], vandq_u8(va.val[0], loset)); |
| 78 | 88 | rh = vqtbl1q_u8(tbl_h[0], vandq_u8(va.val[0], loset)); |
| ... | ... | @@ -84,88 +94,21 @@ neon_w16_split_4_multiply_region(gf_t *gf, uint16_t *src, uint16_t *dst, |
| 84 | 94 | |
| 85 | 95 | rl = veorq_u8(rl, vqtbl1q_u8(tbl_l[1], va.val[0])); |
| 86 | 96 | rh = veorq_u8(rh, vqtbl1q_u8(tbl_h[1], va.val[0])); |
| 87 | - rl = veorq_u8(rl, vqtbl1q_u8(tbl_l[3], va.val[1])); | |
| 88 | - rh = veorq_u8(rh, vqtbl1q_u8(tbl_h[3], va.val[1])); | |
| 89 | - | |
| 90 | - va = vtrnq_u8(rl, rh); | |
| 91 | - r0 = vreinterpretq_u16_u8(va.val[0]); | |
| 92 | - r1 = vreinterpretq_u16_u8(va.val[1]); | |
| 97 | + va.val[0] = veorq_u8(rl, vqtbl1q_u8(tbl_l[3], va.val[1])); | |
| 98 | + va.val[1] = veorq_u8(rh, vqtbl1q_u8(tbl_h[3], va.val[1])); | |
| 93 | 99 | |
| 94 | 100 | if (xor) { |
| 95 | - va0 = vld1q_u16(dst); | |
| 96 | - va1 = vld1q_u16(dst + 8); | |
| 97 | - r0 = veorq_u16(r0, va0); | |
| 98 | - r1 = veorq_u16(r1, va1); | |
| 101 | + uint8x16x2_t vb = vld2q_u8((uint8_t*)dst); | |
| 102 | + va.val[0] = veorq_u8(va.val[0], vb.val[0]); | |
| 103 | + va.val[1] = veorq_u8(va.val[1], vb.val[1]); | |
| 99 | 104 | } |
| 100 | - vst1q_u16(dst, r0); | |
| 101 | - vst1q_u16(dst + 8, r1); | |
| 105 | + vst2q_u8((uint8_t*)dst, va); | |
| 102 | 106 | |
| 103 | 107 | src += 16; |
| 104 | 108 | dst += 16; |
| 105 | 109 | } |
| 106 | 110 | } |
| 107 | 111 | |
| 108 | -#else /* ARCH_AARCH64 */ | |
| 109 | - | |
| 110 | -static | |
| 111 | -inline | |
| 112 | -void | |
| 113 | -neon_w16_split_4_multiply_region(gf_t *gf, uint16_t *src, uint16_t *dst, | |
| 114 | - uint16_t *d_end, uint8_t *tbl, | |
| 115 | - gf_val_32_t val, int xor) | |
| 116 | -{ | |
| 117 | - unsigned i; | |
| 118 | - uint8_t *high = tbl + 4 * 16; | |
| 119 | - uint16x8_t va, r; | |
| 120 | - uint8x8_t loset, vb, vc, rl, rh; | |
| 121 | - | |
| 122 | - uint8x8x2_t tbl_h[4], tbl_l[4]; | |
| 123 | - for (i = 0; i < 4; i++) { | |
| 124 | - tbl_l[i].val[0] = vld1_u8(tbl + i*16); | |
| 125 | - tbl_l[i].val[1] = vld1_u8(tbl + i*16 + 8); | |
| 126 | - tbl_h[i].val[0] = vld1_u8(high + i*16); | |
| 127 | - tbl_h[i].val[1] = vld1_u8(high + i*16 + 8); | |
| 128 | - } | |
| 129 | - | |
| 130 | - loset = vdup_n_u8(0xf); | |
| 131 | - | |
| 132 | - while (dst < d_end) { | |
| 133 | - va = vld1q_u16(src); | |
| 134 | - | |
| 135 | - vb = vmovn_u16(va); | |
| 136 | - vc = vshrn_n_u16(va, 8); | |
| 137 | - | |
| 138 | - rl = vtbl2_u8(tbl_l[0], vand_u8(vb, loset)); | |
| 139 | - rh = vtbl2_u8(tbl_h[0], vand_u8(vb, loset)); | |
| 140 | - vb = vshr_n_u8(vb, 4); | |
| 141 | - rl = veor_u8(rl, vtbl2_u8(tbl_l[2], vand_u8(vc, loset))); | |
| 142 | - rh = veor_u8(rh, vtbl2_u8(tbl_h[2], vand_u8(vc, loset))); | |
| 143 | - vc = vshr_n_u8(vc, 4); | |
| 144 | - rl = veor_u8(rl, vtbl2_u8(tbl_l[1], vb)); | |
| 145 | - rh = veor_u8(rh, vtbl2_u8(tbl_h[1], vb)); | |
| 146 | - rl = veor_u8(rl, vtbl2_u8(tbl_l[3], vc)); | |
| 147 | - rh = veor_u8(rh, vtbl2_u8(tbl_h[3], vc)); | |
| 148 | - | |
| 149 | - r = vmovl_u8(rl); | |
| 150 | - r = vorrq_u16(r, vshll_n_u8(rh, 8)); | |
| 151 | - | |
| 152 | - if (xor) { | |
| 153 | - va = vld1q_u16(dst); | |
| 154 | - r = veorq_u16(r, va); | |
| 155 | - } | |
| 156 | - vst1q_u16(dst, r); | |
| 157 | - | |
| 158 | - src += 8; | |
| 159 | - dst += 8; | |
| 160 | - } | |
| 161 | -} | |
| 162 | -#endif /* ARCH_AARCH64 */ | |
| 163 | - | |
| 164 | -#ifndef ARCH_AARCH64 | |
| 165 | -#define vqtbl1q_u8(tbl, v) vcombine_u8(vtbl2_u8(tbl, vget_low_u8(v)), \ | |
| 166 | - vtbl2_u8(tbl, vget_high_u8(v))) | |
| 167 | -#endif | |
| 168 | - | |
| 169 | 112 | static |
| 170 | 113 | inline |
| 171 | 114 | void | ... | ... |