| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/rvv.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2024 SiFive, Inc. |
| 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 <riscv_vector.h> |
| 16 |
|
| 17 |
#include "xnnpack/packw.h" |
| 18 |
|
| 19 |
void xnn_x32_packw_gemm_goi_ukernel_x1v__rvv_u2( |
| 20 |
size_t g, |
| 21 |
size_t nc, |
| 22 |
size_t kc, |
| 23 |
size_t nr, |
| 24 |
size_t kr, |
| 25 |
size_t sr, |
| 26 |
const uint32_t* weights, |
| 27 |
const uint32_t* bias, |
| 28 |
const void* scale, |
| 29 |
uint32_t* packed_weights, |
| 30 |
size_t extra_bytes, |
| 31 |
const void* params) |
| 32 |
{ |
| 33 |
assert(g != 0); |
| 34 |
assert(nc != 0); |
| 35 |
assert(kc != 0); |
| 36 |
assert(nr == __riscv_vsetvlmax_e32m1()); |
| 37 |
assert(kr == 1); |
| 38 |
assert(sr == 1); |
| 39 |
assert(weights != NULL); |
| 40 |
assert(packed_weights != NULL); |
| 41 |
|
| 42 |
uint32_t* out = packed_weights; |
| 43 |
const uint32_t* b = bias; |
| 44 |
size_t kc_bstride = kc * 4; |
| 45 |
|
| 46 |
do { |
| 47 |
const uint32_t* w0 = weights; |
| 48 |
size_t n = nc; |
| 49 |
// NC main loop: process multiple of NR |
| 50 |
for (;n >= nr; n -= nr) { |
| 51 |
// Pack nr bias at begining of tile |
| 52 |
size_t vlmax = __riscv_vsetvlmax_e32m1(); |
| 53 |
vuint32m1_t v_bias; |
| 54 |
if XNN_LIKELY(b != NULL) { |
| 55 |
v_bias = __riscv_vle32_v_u32m1(b, vlmax); b += nr; |
| 56 |
} else { |
| 57 |
v_bias = __riscv_vmv_v_x_u32m1(0, vlmax); |
| 58 |
} |
| 59 |
__riscv_vse32_v_u32m1(out, v_bias, vlmax); out += nr; |
| 60 |
|
| 61 |
uint32_t* out0 = out; |
| 62 |
size_t k = kc; |
| 63 |
// vlsseg2, LMUL must <= 4 |
| 64 |
vlmax = __riscv_vsetvlmax_e32m1(); |
| 65 |
// Pack 2 x nr weights |
| 66 |
for (; k >= 2; k -= 2) { |
| 67 |
uint32_t* out1 = out0 + nr; |
| 68 |
// 1 load & store can pack 2 x nr weights |
| 69 |
vuint32m1x2_t v_w_m1x2 = __riscv_vlsseg2e32_v_u32m1x2(w0, kc_bstride, vlmax); |
| 70 |
vuint32m1_t v_w0 = __riscv_vget_v_u32m1x2_u32m1(v_w_m1x2, 0); |
| 71 |
__riscv_vse32_v_u32m1(out0, v_w0, vlmax); |
| 72 |
vuint32m1_t v_w1 = __riscv_vget_v_u32m1x2_u32m1(v_w_m1x2, 1); |
| 73 |
__riscv_vse32_v_u32m1(out1, v_w1, vlmax); |
| 74 |
out0 = out1 + vlmax; |
| 75 |
w0 += 2; |
| 76 |
} |
| 77 |
vlmax = __riscv_vsetvlmax_e32m1(); |
| 78 |
// Pack nr weights |
| 79 |
for (; k >= 1; k -= 1) { |
| 80 |
vuint32m1_t v_w = __riscv_vlse32_v_u32m1(w0, kc_bstride, vlmax); |
| 81 |
__riscv_vse32_v_u32m1(out0, v_w, vlmax); |
| 82 |
out0 += vlmax; |
| 83 |
w0 += 1; |
| 84 |
} |
| 85 |
out = (uint32_t*) ((uintptr_t) out0 + extra_bytes); |
| 86 |
w0 += (nr - 1) * kc; |
| 87 |
} |
| 88 |
// NC remainder: process n < NR |
| 89 |
if (n > 0) { |
| 90 |
// Pack nr bias at begining of tile |
| 91 |
size_t vl = __riscv_vsetvl_e32m1(n); |
| 92 |
vuint32m1_t v_bias; |
| 93 |
if XNN_LIKELY(b != NULL) { |
| 94 |
v_bias = __riscv_vle32_v_u32m1(b, vl); b += vl; |
| 95 |
} else { |
| 96 |
v_bias = __riscv_vmv_v_x_u32m1(0, vl); |
| 97 |
} |
| 98 |
__riscv_vse32_v_u32m1(out, v_bias, vl); out += nr; |
| 99 |
|
| 100 |
size_t vlmax; |
| 101 |
uint32_t* out0 = out; |
| 102 |
size_t k = kc; |
| 103 |
// vlsseg2, LMUL must <= 4 |
| 104 |
vlmax = __riscv_vsetvlmax_e32m1(); |
| 105 |
// Pack 2 x n weights |
| 106 |
for (; k >= 2; k -= 2) { |
| 107 |
uint32_t* out1 = out0 + nr; |
| 108 |
// 1 load & store can pack 2 x nr weights |
| 109 |
size_t vl = __riscv_vsetvl_e32m1(n); |
| 110 |
vuint32m1x2_t v_w_m1x2 = __riscv_vlsseg2e32_v_u32m1x2(w0, kc_bstride, vl); |
| 111 |
vuint32m1_t v_w0 = __riscv_vget_v_u32m1x2_u32m1(v_w_m1x2, 0); |
| 112 |
__riscv_vse32_v_u32m1(out0, v_w0, vl); |
| 113 |
vuint32m1_t v_w1 = __riscv_vget_v_u32m1x2_u32m1(v_w_m1x2, 1); |
| 114 |
__riscv_vse32_v_u32m1(out1, v_w1, vl); |
| 115 |
out0 = out1 + vlmax; |
| 116 |
w0 += 2; |
| 117 |
} |
| 118 |
vlmax = __riscv_vsetvlmax_e32m1(); |
| 119 |
vl = __riscv_vsetvl_e32m1(n); |
| 120 |
// Pack n weights |
| 121 |
for (; k >= 1; k -= 1) { |
| 122 |
vuint32m1_t v_w = __riscv_vlse32_v_u32m1(w0, kc_bstride, vl); |
| 123 |
__riscv_vse32_v_u32m1(out0, v_w, vl); |
| 124 |
out0 += vlmax; |
| 125 |
w0 += 1; |
| 126 |
} |
| 127 |
out = (uint32_t*) ((uintptr_t) out0 + extra_bytes); |
| 128 |
w0 += (nr - 1) * kc; |
| 129 |
} |
| 130 |
weights += nc * kc; |
| 131 |
} while (--g != 0); |
| 132 |
} |