Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16c8-minmax-neondot-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neondot.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
12 #include <arm_neon.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x16c8__neondot_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43
44 do {
45 const int32x4_t vinput_zero_point0 = vld1q_dup_s32(&quantization_params[0].zero_point);
46 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
47 const int32x4_t vinit0x0123 = vmulq_s32(vksum0123, vinput_zero_point0);
48 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
49 const int32x4_t vinit0x4567 = vmulq_s32(vksum4567, vinput_zero_point0);
50 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
51 const int32x4_t vinit0x89AB = vmulq_s32(vksum89AB, vinput_zero_point0);
52 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
53 const int32x4_t vinit0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point0);
54 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vget_low_u32(vreinterpretq_u32_s32(vinit0x0123))));
55 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vget_high_u32(vreinterpretq_u32_s32(vinit0x0123))));
56 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vget_low_u32(vreinterpretq_u32_s32(vinit0x4567))));
57 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vget_high_u32(vreinterpretq_u32_s32(vinit0x4567))));
58 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vget_low_u32(vreinterpretq_u32_s32(vinit0x89AB))));
59 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vget_high_u32(vreinterpretq_u32_s32(vinit0x89AB))));
60 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vget_low_u32(vreinterpretq_u32_s32(vinit0xCDEF))));
61 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vget_high_u32(vreinterpretq_u32_s32(vinit0xCDEF))));
62
63 size_t k = kc;
64 do {
65 const int8x16_t va0c01234567 = vreinterpretq_s8_u64(vld1q_dup_u64((const void*) a0)); a0 += 8;
66
67 const int8x16_t vb01c01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
68 const int8x16_t vb23c01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
69 const int8x16_t vb45c01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
70 const int8x16_t vb67c01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
71 const int8x16_t vb89c01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
72 const int8x16_t vbABc01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
73 const int8x16_t vbCDc01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
74 const int8x16_t vbEFc01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
75
76 vacc0x01 = vdotq_s32(vacc0x01, vb01c01234567, va0c01234567);
77 vacc0x23 = vdotq_s32(vacc0x23, vb23c01234567, va0c01234567);
78 vacc0x45 = vdotq_s32(vacc0x45, vb45c01234567, va0c01234567);
79 vacc0x67 = vdotq_s32(vacc0x67, vb67c01234567, va0c01234567);
80 vacc0x89 = vdotq_s32(vacc0x89, vb89c01234567, va0c01234567);
81 vacc0xAB = vdotq_s32(vacc0xAB, vbABc01234567, va0c01234567);
82 vacc0xCD = vdotq_s32(vacc0xCD, vbCDc01234567, va0c01234567);
83 vacc0xEF = vdotq_s32(vacc0xEF, vbEFc01234567, va0c01234567);
84
85 k -= 8 * sizeof(int8_t);
86 } while (k != 0);
87
88 #if XNN_ARCH_ARM64
89 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
90 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
91 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
92 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
93 #else
94 int32x4_t vacc0x0123 = vcombine_s32(vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01)), vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23)));
95 int32x4_t vacc0x4567 = vcombine_s32(vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45)), vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67)));
96 int32x4_t vacc0x89AB = vcombine_s32(vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89)), vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB)));
97 int32x4_t vacc0xCDEF = vcombine_s32(vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)), vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF)));
98 #endif
99
100 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
101 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
102 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
103 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
104
105 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
106 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
107 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
108 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
109 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
110
111 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
112 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
113 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
114 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
115
116 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
117 #if XNN_ARCH_ARM64
118 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
119 #else
120 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
121 #endif
122 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
123 #if XNN_ARCH_ARM64
124 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
125 #else
126 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
127 #endif
128 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
129 #if XNN_ARCH_ARM64
130 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
131 #else
132 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
133 #endif
134 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
135 #if XNN_ARCH_ARM64
136 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
137 #else
138 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
139 #endif
140
141 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
142 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
143 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
144 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
145 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
146
147 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
148 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
149 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
150 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
151 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
152
153 if XNN_LIKELY(nc >= 16) {
154 vst1q_f32(c0, vout0x0123);
155 vst1q_f32(c0 + 4, vout0x4567);
156 vst1q_f32(c0 + 8, vout0x89AB);
157 vst1q_f32(c0 + 12, vout0xCDEF);
158
159 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
160
161 c0 = (float*) ((uintptr_t) c0 + cn_stride);
162
163 nc -= 16;
164 } else {
165 if (nc & 8) {
166 vst1q_f32(c0, vout0x0123); c0 += 4;
167 vout0x0123 = vout0x89AB;
168 vst1q_f32(c0, vout0x4567); c0 += 4;
169 vout0x4567 = vout0xCDEF;
170 }
171 if (nc & 4) {
172 vst1q_f32(c0, vout0x0123); c0 += 4;
173 vout0x0123 = vout0x4567;
174 }
175 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
176 if (nc & 2) {
177 vst1_f32(c0, vout0x01); c0 += 2;
178 vout0x01 = vget_high_f32(vout0x0123);
179 }
180 if (nc & 1) {
181 vst1_lane_f32(c0, vout0x01, 0);
182 }
183 nc = 0;
184 }
185 } while (nc != 0);
186 }