Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x4v-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_4x4v__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 <= 4);
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 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
50 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 4) {
52 a3 = a2;
53 c3 = c2;
54 }
55
56 const size_t nr = __riscv_vsetvlmax_e32m4();
57 size_t vl = nr;
58 kc = round_up_po2(kc, 2);
59 do {
60 if XNN_UNLIKELY(nc < nr) {
61 vl = __riscv_vsetvl_e32m4(nc);
62 }
63 nc = nc - vl;
64
65 vint32m4_t vksum = __riscv_vle32_v_i32m4((const int32_t*)w, vl);
66 w = (const int32_t*) w + nr;
67 const int32_t vinput_zero_point0 = quantization_params[0].zero_point;
68 const int32_t vinput_zero_point1 = quantization_params[1].zero_point;
69 const int32_t vinput_zero_point2 = quantization_params[2].zero_point;
70 const int32_t vinput_zero_point3 = quantization_params[3].zero_point;
71 vint32m4_t vacc0 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point0, vl);
72 vint32m4_t vacc1 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point1, vl);
73 vint32m4_t vacc2 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point2, vl);
74 vint32m4_t vacc3 = __riscv_vmul_vx_i32m4(vksum, vinput_zero_point3, vl);
75
76 size_t k = kc;
77 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
78 const int8_t va0c0 = a0[0];
79 const int8_t va0c1 = a0[1];
80 a0 += 2;
81 const int8_t va1c0 = a1[0];
82 const int8_t va1c1 = a1[1];
83 a1 += 2;
84 const int8_t va2c0 = a2[0];
85 const int8_t va2c1 = a2[1];
86 a2 += 2;
87 const int8_t va3c0 = a3[0];
88 const int8_t va3c1 = a3[1];
89 a3 += 2;
90 const vint8m1_t vbi = __riscv_vle8_v_i8m1((const int8_t*) w, vl);
91 w = (const uint8_t*) w + nr;
92 const vint8m1_t vbc0 = __riscv_vsll_vx_i8m1(vbi, 4, vl);
93 const vint8m1_t vbc1 = __riscv_vand_vx_i8m1(vbi, 0xF0, vl);
94
95 vint16m2_t va0bc0 = __riscv_vwmul_vx_i16m2(vbc0, va0c0, vl);
96 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0bc0, vl);
97 vint16m2_t va0bc1 = __riscv_vwmul_vx_i16m2(vbc1, va0c1, vl);
98 vacc0 = __riscv_vwadd_wv_i32m4(vacc0, va0bc1, vl);
99 vint16m2_t va1bc0 = __riscv_vwmul_vx_i16m2(vbc0, va1c0, vl);
100 vacc1 = __riscv_vwadd_wv_i32m4(vacc1, va1bc0, vl);
101 vint16m2_t va1bc1 = __riscv_vwmul_vx_i16m2(vbc1, va1c1, vl);
102 vacc1 = __riscv_vwadd_wv_i32m4(vacc1, va1bc1, vl);
103 vint16m2_t va2bc0 = __riscv_vwmul_vx_i16m2(vbc0, va2c0, vl);
104 vacc2 = __riscv_vwadd_wv_i32m4(vacc2, va2bc0, vl);
105 vint16m2_t va2bc1 = __riscv_vwmul_vx_i16m2(vbc1, va2c1, vl);
106 vacc2 = __riscv_vwadd_wv_i32m4(vacc2, va2bc1, vl);
107 vint16m2_t va3bc0 = __riscv_vwmul_vx_i16m2(vbc0, va3c0, vl);
108 vacc3 = __riscv_vwadd_wv_i32m4(vacc3, va3bc0, vl);
109 vint16m2_t va3bc1 = __riscv_vwmul_vx_i16m2(vbc1, va3c1, vl);
110 vacc3 = __riscv_vwadd_wv_i32m4(vacc3, va3bc1, vl);
111 }
112 vacc0 = __riscv_vsra_vx_i32m4(vacc0, 4, vl);
113 vacc1 = __riscv_vsra_vx_i32m4(vacc1, 4, vl);
114 vacc2 = __riscv_vsra_vx_i32m4(vacc2, 4, vl);
115 vacc3 = __riscv_vsra_vx_i32m4(vacc3, 4, vl);
116 // i32 -> f32
117 vfloat32m4_t vout0 = __riscv_vfcvt_f_x_v_f32m4(vacc0, vl);
118 vfloat32m4_t vout1 = __riscv_vfcvt_f_x_v_f32m4(vacc1, vl);
119 vfloat32m4_t vout2 = __riscv_vfcvt_f_x_v_f32m4(vacc2, vl);
120 vfloat32m4_t vout3 = __riscv_vfcvt_f_x_v_f32m4(vacc3, vl);
121
122 // vout * input_scale
123 const float vinput_scale0 = quantization_params[0].inv_scale;
124 const float vinput_scale1 = quantization_params[1].inv_scale;
125 const float vinput_scale2 = quantization_params[2].inv_scale;
126 const float vinput_scale3 = quantization_params[3].inv_scale;
127 vout0 = __riscv_vfmul_vf_f32m4(vout0, vinput_scale0, vl);
128 vout1 = __riscv_vfmul_vf_f32m4(vout1, vinput_scale1, vl);
129 vout2 = __riscv_vfmul_vf_f32m4(vout2, vinput_scale2, vl);
130 vout3 = __riscv_vfmul_vf_f32m4(vout3, vinput_scale3, vl);
131
132 const vfloat32m4_t vfilter_output_scale = __riscv_vle32_v_f32m4((const float*) w, vl);
133 w = (const float*) w + nr;
134 vout0 = __riscv_vfmul_vv_f32m4(vout0, vfilter_output_scale, vl);
135 vout1 = __riscv_vfmul_vv_f32m4(vout1, vfilter_output_scale, vl);
136 vout2 = __riscv_vfmul_vv_f32m4(vout2, vfilter_output_scale, vl);
137 vout3 = __riscv_vfmul_vv_f32m4(vout3, vfilter_output_scale, vl);
138
139 const vfloat32m4_t vbias = __riscv_vle32_v_f32m4((const float*) w, vl);
140 w = (const float*) w + nr;
141 vout0 = __riscv_vfadd_vv_f32m4(vout0, vbias, vl);
142 vout1 = __riscv_vfadd_vv_f32m4(vout1, vbias, vl);
143 vout2 = __riscv_vfadd_vv_f32m4(vout2, vbias, vl);
144 vout3 = __riscv_vfadd_vv_f32m4(vout3, vbias, vl);
145
146 const float vmin = params->scalar.min;
147 vout0 = __riscv_vfmax_vf_f32m4(vout0, vmin, vl);
148 vout1 = __riscv_vfmax_vf_f32m4(vout1, vmin, vl);
149 vout2 = __riscv_vfmax_vf_f32m4(vout2, vmin, vl);
150 vout3 = __riscv_vfmax_vf_f32m4(vout3, vmin, vl);
151 const float vmax = params->scalar.max;
152 vout0 = __riscv_vfmin_vf_f32m4(vout0, vmax, vl);
153 vout1 = __riscv_vfmin_vf_f32m4(vout1, vmax, vl);
154 vout2 = __riscv_vfmin_vf_f32m4(vout2, vmax, vl);
155 vout3 = __riscv_vfmin_vf_f32m4(vout3, vmax, vl);
156
157 // store 4 x vl results to c
158 __riscv_vse32_v_f32m4(c0, vout0, vl);
159 c0 = (float*) ((uintptr_t) c0 + cn_stride);
160 __riscv_vse32_v_f32m4(c1, vout1, vl);
161 c1 = (float*) ((uintptr_t) c1 + cn_stride);
162 __riscv_vse32_v_f32m4(c2, vout2, vl);
163 c2 = (float*) ((uintptr_t) c2 + cn_stride);
164 __riscv_vse32_v_f32m4(c3, vout3, vl);
165 c3 = (float*) ((uintptr_t) c3 + cn_stride);
166
167 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
168 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
169 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
170 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
171 } while (nc != 0);
172 }