Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4v-minmax-rvv.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #include <assert.h>
11
12 #include <riscv_vector.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/math.h"
16
17 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x4v__rvv(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34
35 const int8_t* a0 = a;
36 float* c0 = c;
37
38 const size_t nr = __riscv_vsetvlmax_e32m4();
39 size_t vl = nr;
40 kc = round_up_po2(kc, 2);
41 do {
42 if XNN_UNLIKELY(nc < nr) {
43 vl = __riscv_vsetvl_e32m4(nc);
44 }
45 nc = nc - vl;
46
47 vint32m4_t vksum = __riscv_vle32_v_i32m4((const int32_t*)w, vl);
48 w = (const int32_t*) w + nr;
49 const int32_t vinput_zero_point0 = quantization_params[0].zero_point;
50 vint32m4_t vacc0 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point0, vl);
51
52 size_t k = kc;
53 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
54 const int8_t va0c0 = a0[0];
55 const int8_t va0c1 = a0[1];
56 a0 += 2;
57 const vint8m1_t vbi = __riscv_vle8_v_i8m1((const int8_t*) w, vl);
58 w = (const uint8_t*) w + nr;
59 const vint8m1_t vbc0 = __riscv_vsll_vx_i8m1(vbi, 4, vl);
60 const vint8m1_t vbc1 = __riscv_vand_vx_i8m1(vbi, 0xF0, vl);
61
62 vint16m2_t va0bc0 = __riscv_vwmul_vx_i16m2(vbc0, va0c0, vl);
63 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0bc0, vl);
64 vint16m2_t va0bc1 = __riscv_vwmul_vx_i16m2(vbc1, va0c1, vl);
65 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0bc1, vl);
66 }
67 vacc0 = __riscv_vsra_vx_i32m4(vacc0, 4, vl);
68 // i32 -> f32
69 vfloat32m4_t vout0 = __riscv_vfcvt_f_x_v_f32m4(vacc0, vl);
70
71 // vout * input_scale
72 const float vinput_scale0 = quantization_params[0].inv_scale;
73 vout0 = __riscv_vfmul_vf_f32m4(vout0, vinput_scale0, vl);
74
75 const vfloat32m4_t vfilter_output_scale = __riscv_vle32_v_f32m4((const float*) w, vl);
76 w = (const float*) w + nr;
77 vout0 = __riscv_vfmul_vv_f32m4(vout0, vfilter_output_scale, vl);
78
79 const vfloat32m4_t vbias = __riscv_vle32_v_f32m4((const float*) w, vl);
80 w = (const float*) w + nr;
81 vout0 = __riscv_vfadd_vv_f32m4(vout0, vbias, vl);
82
83 const float vmin = params->scalar.min;
84 vout0 = __riscv_vfmax_vf_f32m4(vout0, vmin, vl);
85 const float vmax = params->scalar.max;
86 vout0 = __riscv_vfmin_vf_f32m4(vout0, vmax, vl);
87
88 // store 1 x vl results to c
89 __riscv_vse32_v_f32m4(c0, vout0, vl);
90 c0 = (float*) ((uintptr_t) c0 + cn_stride);
91
92 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
93 } while (nc != 0);
94 }