Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-f32-rvv-u1v.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/rvv.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 #include <assert.h>
11 #include <riscv_vector.h>
12 #include "xnnpack/vbinary.h"
13
14
15 void xnn_qs8_vmulc_minmax_fp32_ukernel__rvv_u1v(
16 size_t batch,
17 const int8_t* input_a,
18 const int8_t* input_b,
19 int8_t* output,
20 const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
21 {
22 assert(batch != 0);
23 assert(batch % sizeof(int8_t) == 0);
24 assert(input_a != NULL);
25 assert(input_b != NULL);
26 assert(output != NULL);
27
28 const int32_t a_zero_point = params->scalar.a_zero_point;
29 const float scale = params->scalar.scale;
30 const float output_min_less_zero_point = (int32_t) params->scalar.output_min - (int32_t) params->scalar.output_zero_point;
31 const float output_max_less_zero_point = (int32_t) params->scalar.output_max - (int32_t) params->scalar.output_zero_point;
32 const float magic_bias = 12582912.0f;
33 const int32_t magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) params->scalar.output_zero_point;
34
35 const int32_t vb = (int32_t) *input_b - params->scalar.b_zero_point;
36 do {
37 const size_t n = __riscv_vsetvl_e8m1(batch);
38
39 vint8m1_t in_a_i8v = __riscv_vle8_v_i8m1(input_a, n); input_a += n;
40 vint16m2_t a_i16v = __riscv_vwsub_vx_i16m2(in_a_i8v, a_zero_point, n);
41
42 vint32m4_t acc_i32v = __riscv_vwmul_vx_i32m4(a_i16v, vb, n);
43 vfloat32m4_t acc_f32v = __riscv_vfcvt_f_x_v_f32m4(acc_i32v, n);
44 acc_f32v = __riscv_vfmul_vf_f32m4(acc_f32v, scale, n);
45 acc_f32v = __riscv_vfmin_vf_f32m4(__riscv_vfmax_vf_f32m4(acc_f32v, output_min_less_zero_point, n), output_max_less_zero_point, n);
46 acc_f32v = __riscv_vfadd_vf_f32m4(acc_f32v, magic_bias, n);
47
48 vint32m4_t out_i32v = __riscv_vfcvt_x_f_v_i32m4(acc_f32v, n);
49 out_i32v = __riscv_vsub_vx_i32m4(out_i32v, magic_bias_less_output_zero_point, n);
50 vint16m2_t out_i16v = __riscv_vncvt_x_x_w_i16m2(out_i32v, n);
51 vint8m1_t out_i8v = __riscv_vncvt_x_x_w_i8m1(out_i16v, n);
52 __riscv_vse8_v_i8m1(output, out_i8v, n); output += n;
53
54 batch -= n;
55 } while (batch != 0);
56 }