Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-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_qc8w_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 union xnn_f32_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 do {
41 if XNN_UNLIKELY(nc < nr) {
42 vl = __riscv_vsetvl_e32m4(nc);
43 }
44 nc = nc - vl;
45
46 vint32m4_t vksum = __riscv_vle32_v_i32m4((const int32_t*)w, vl);
47 w = (const int32_t*) w + nr;
48 const int32_t vinput_zero_point0 = quantization_params[0].zero_point;
49 vint32m4_t vacc0 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point0, vl);
50
51 size_t k = kc;
52 do {
53 const int8_t va0 = *a0++;
54 const vint8m1_t vb = __riscv_vle8_v_i8m1((const int8_t*) w, vl);
55 w = (const int8_t*) w + nr;
56
57 vint16m2_t va0b = __riscv_vwmul_vx_i16m2(vb, va0, vl);
58 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0b, vl);
59
60 k -= sizeof(int8_t);
61 } while (k != 0);
62 // i32 -> f32
63 vfloat32m4_t vout0 = __riscv_vfcvt_f_x_v_f32m4(vacc0, vl);
64
65 // vout * input_scale
66 const float vinput_scale0 = quantization_params[0].inv_scale;
67 vout0 = __riscv_vfmul_vf_f32m4(vout0, vinput_scale0, vl);
68
69 const vfloat32m4_t vfilter_output_scale = __riscv_vle32_v_f32m4((const float*) w, vl);
70 w = (const float*) w + nr;
71 vout0 = __riscv_vfmul_vv_f32m4(vout0, vfilter_output_scale, vl);
72
73 const vfloat32m4_t vbias = __riscv_vle32_v_f32m4((const float*) w, vl);
74 w = (const float*) w + nr;
75 vout0 = __riscv_vfadd_vv_f32m4(vout0, vbias, vl);
76
77 const float vmin = params->scalar.min;
78 vout0 = __riscv_vfmax_vf_f32m4(vout0, vmin, vl);
79 const float vmax = params->scalar.max;
80 vout0 = __riscv_vfmin_vf_f32m4(vout0, vmax, vl);
81
82 // store 1 x vl results to c
83 __riscv_vse32_v_f32m4(c0, vout0, vl);
84 c0 = (float*) ((uintptr_t) c0 + cn_stride);
85
86 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
87 } while (nc != 0);
88 }