Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x4v-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_3x4v__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 <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34
35 const int8_t* a0 = a;
36 float* c0 = c;
37 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
38 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr < 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
44 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
45 if XNN_UNPREDICTABLE(mr <= 2) {
46 a2 = a1;
47 c2 = c1;
48 }
49
50 const size_t nr = __riscv_vsetvlmax_e32m4();
51 size_t vl = nr;
52 kc = round_up_po2(kc, 2);
53 do {
54 if XNN_UNLIKELY(nc < nr) {
55 vl = __riscv_vsetvl_e32m4(nc);
56 }
57 nc = nc - vl;
58
59 vint32m4_t vksum = __riscv_vle32_v_i32m4((const int32_t*)w, vl);
60 w = (const int32_t*) w + nr;
61 const int32_t vinput_zero_point0 = quantization_params[0].zero_point;
62 const int32_t vinput_zero_point1 = quantization_params[1].zero_point;
63 const int32_t vinput_zero_point2 = quantization_params[2].zero_point;
64 vint32m4_t vacc0 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point0, vl);
65 vint32m4_t vacc1 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point1, vl);
66 vint32m4_t vacc2 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point2, vl);
67
68 size_t k = kc;
69 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
70 const int8_t va0c0 = a0[0];
71 const int8_t va0c1 = a0[1];
72 a0 += 2;
73 const int8_t va1c0 = a1[0];
74 const int8_t va1c1 = a1[1];
75 a1 += 2;
76 const int8_t va2c0 = a2[0];
77 const int8_t va2c1 = a2[1];
78 a2 += 2;
79 const vint8m1_t vbi = __riscv_vle8_v_i8m1((const int8_t*) w, vl);
80 w = (const uint8_t*) w + nr;
81 const vint8m1_t vbc0 = __riscv_vsll_vx_i8m1(vbi, 4, vl);
82 const vint8m1_t vbc1 = __riscv_vand_vx_i8m1(vbi, 0xF0, vl);
83
84 vint16m2_t va0bc0 = __riscv_vwmul_vx_i16m2(vbc0, va0c0, vl);
85 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0bc0, vl);
86 vint16m2_t va0bc1 = __riscv_vwmul_vx_i16m2(vbc1, va0c1, vl);
87 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0bc1, vl);
88 vint16m2_t va1bc0 = __riscv_vwmul_vx_i16m2(vbc0, va1c0, vl);
89 vacc1 = __riscv_vwadd_wv_i32m4(vacc1, va1bc0, vl);
90 vint16m2_t va1bc1 = __riscv_vwmul_vx_i16m2(vbc1, va1c1, vl);
91 vacc1 = __riscv_vwadd_wv_i32m4(vacc1, va1bc1, vl);
92 vint16m2_t va2bc0 = __riscv_vwmul_vx_i16m2(vbc0, va2c0, vl);
93 vacc2 = __riscv_vwadd_wv_i32m4(vacc2, va2bc0, vl);
94 vint16m2_t va2bc1 = __riscv_vwmul_vx_i16m2(vbc1, va2c1, vl);
95 vacc2 = __riscv_vwadd_wv_i32m4(vacc2, va2bc1, vl);
96 }
97 vacc0 = __riscv_vsra_vx_i32m4(vacc0, 4, vl);
98 vacc1 = __riscv_vsra_vx_i32m4(vacc1, 4, vl);
99 vacc2 = __riscv_vsra_vx_i32m4(vacc2, 4, vl);
100 // i32 -> f32
101 vfloat32m4_t vout0 = __riscv_vfcvt_f_x_v_f32m4(vacc0, vl);
102 vfloat32m4_t vout1 = __riscv_vfcvt_f_x_v_f32m4(vacc1, vl);
103 vfloat32m4_t vout2 = __riscv_vfcvt_f_x_v_f32m4(vacc2, vl);
104
105 // vout * input_scale
106 const float vinput_scale0 = quantization_params[0].inv_scale;
107 const float vinput_scale1 = quantization_params[1].inv_scale;
108 const float vinput_scale2 = quantization_params[2].inv_scale;
109 vout0 = __riscv_vfmul_vf_f32m4(vout0, vinput_scale0, vl);
110 vout1 = __riscv_vfmul_vf_f32m4(vout1, vinput_scale1, vl);
111 vout2 = __riscv_vfmul_vf_f32m4(vout2, vinput_scale2, vl);
112
113 const vfloat32m4_t vfilter_output_scale = __riscv_vle32_v_f32m4((const float*) w, vl);
114 w = (const float*) w + nr;
115 vout0 = __riscv_vfmul_vv_f32m4(vout0, vfilter_output_scale, vl);
116 vout1 = __riscv_vfmul_vv_f32m4(vout1, vfilter_output_scale, vl);
117 vout2 = __riscv_vfmul_vv_f32m4(vout2, vfilter_output_scale, vl);
118
119 const vfloat32m4_t vbias = __riscv_vle32_v_f32m4((const float*) w, vl);
120 w = (const float*) w + nr;
121 vout0 = __riscv_vfadd_vv_f32m4(vout0, vbias, vl);
122 vout1 = __riscv_vfadd_vv_f32m4(vout1, vbias, vl);
123 vout2 = __riscv_vfadd_vv_f32m4(vout2, vbias, vl);
124
125 const float vmin = params->scalar.min;
126 vout0 = __riscv_vfmax_vf_f32m4(vout0, vmin, vl);
127 vout1 = __riscv_vfmax_vf_f32m4(vout1, vmin, vl);
128 vout2 = __riscv_vfmax_vf_f32m4(vout2, vmin, vl);
129 const float vmax = params->scalar.max;
130 vout0 = __riscv_vfmin_vf_f32m4(vout0, vmax, vl);
131 vout1 = __riscv_vfmin_vf_f32m4(vout1, vmax, vl);
132 vout2 = __riscv_vfmin_vf_f32m4(vout2, vmax, vl);
133
134 // store 3 x vl results to c
135 __riscv_vse32_v_f32m4(c0, vout0, vl);
136 c0 = (float*) ((uintptr_t) c0 + cn_stride);
137 __riscv_vse32_v_f32m4(c1, vout1, vl);
138 c1 = (float*) ((uintptr_t) c1 + cn_stride);
139 __riscv_vse32_v_f32m4(c2, vout2, vl);
140 c2 = (float*) ((uintptr_t) c2 + cn_stride);
141
142 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
143 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
144 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
145 } while (nc != 0);
146 }