Home
last modified time | relevance | path

Searched refs:float16x8_t (Results 1 – 25 of 172) sorted by relevance

1234567

/external/llvm-project/clang/test/CodeGen/
Daarch64-v8.2a-neon-intrinsics.c20 float16x8_t test_vabsq_f16(float16x8_t a) { in test_vabsq_f16()
36 uint16x8_t test_vceqzq_f16(float16x8_t a) { in test_vceqzq_f16()
52 uint16x8_t test_vcgezq_f16(float16x8_t a) { in test_vcgezq_f16()
68 uint16x8_t test_vcgtzq_f16(float16x8_t a) { in test_vcgtzq_f16()
84 uint16x8_t test_vclezq_f16(float16x8_t a) { in test_vclezq_f16()
100 uint16x8_t test_vcltzq_f16(float16x8_t a) { in test_vcltzq_f16()
114 float16x8_t test_vcvtq_f16_s16 (int16x8_t a) { in test_vcvtq_f16_s16()
128 float16x8_t test_vcvtq_f16_u16 (uint16x8_t a) { in test_vcvtq_f16_u16()
142 int16x8_t test_vcvtq_s16_f16 (float16x8_t a) { in test_vcvtq_s16_f16()
156 uint16x8_t test_vcvtq_u16_f16 (float16x8_t a) { in test_vcvtq_u16_f16()
[all …]
Darm-v8.2a-neon-intrinsics.c20 float16x8_t test_vabsq_f16(float16x8_t a) { in test_vabsq_f16()
36 uint16x8_t test_vceqzq_f16(float16x8_t a) { in test_vceqzq_f16()
52 uint16x8_t test_vcgezq_f16(float16x8_t a) { in test_vcgezq_f16()
68 uint16x8_t test_vcgtzq_f16(float16x8_t a) { in test_vcgtzq_f16()
84 uint16x8_t test_vclezq_f16(float16x8_t a) { in test_vclezq_f16()
100 uint16x8_t test_vcltzq_f16(float16x8_t a) { in test_vcltzq_f16()
114 float16x8_t test_vcvtq_f16_s16 (int16x8_t a) { in test_vcvtq_f16_s16()
128 float16x8_t test_vcvtq_f16_u16 (uint16x8_t a) { in test_vcvtq_f16_u16()
142 int16x8_t test_vcvtq_s16_f16 (float16x8_t a) { in test_vcvtq_s16_f16()
156 int16x8_t test_vcvtq_u16_f16 (float16x8_t a) { in test_vcvtq_u16_f16()
[all …]
Daarch64-v8.2a-neon-intrinsics-constrained.c48 float16x8_t test_vsqrtq_f16(float16x8_t a) { in test_vsqrtq_f16()
66 float16x8_t test_vfmaq_f16(float16x8_t a, float16x8_t b, float16x8_t c) { in test_vfmaq_f16()
86 float16x8_t test_vfmsq_f16(float16x8_t a, float16x8_t b, float16x8_t c) { in test_vfmsq_f16()
118 float16x8_t test_vfmaq_lane_f16(float16x8_t a, float16x8_t b, float16x4_t c) { in test_vfmaq_lane_f16()
134 float16x4_t test_vfma_laneq_f16(float16x4_t a, float16x4_t b, float16x8_t c) { in test_vfma_laneq_f16()
150 float16x8_t test_vfmaq_laneq_f16(float16x8_t a, float16x8_t b, float16x8_t c) { in test_vfmaq_laneq_f16()
180 float16x8_t test_vfmaq_n_f16(float16x8_t a, float16x8_t b, float16_t c) { in test_vfmaq_n_f16()
200 float16_t test_vfmah_laneq_f16(float16_t a, float16_t b, float16x8_t c) { in test_vfmah_laneq_f16()
234 float16x8_t test_vfmsq_lane_f16(float16x8_t a, float16x8_t b, float16x4_t c) { in test_vfmsq_lane_f16()
252 float16x4_t test_vfms_laneq_f16(float16x4_t a, float16x4_t b, float16x8_t c) { in test_vfms_laneq_f16()
[all …]
Daarch64-neon-fp16fml.c69 float32x4_t test_vfmlalq_low_f16(float32x4_t a, float16x8_t b, float16x8_t c) { in test_vfmlalq_low_f16()
81 float32x4_t test_vfmlslq_low_f16(float32x4_t a, float16x8_t b, float16x8_t c) { in test_vfmlslq_low_f16()
93 float32x4_t test_vfmlalq_high_f16(float32x4_t a, float16x8_t b, float16x8_t c) { in test_vfmlalq_high_f16()
105 float32x4_t test_vfmlslq_high_f16(float32x4_t a, float16x8_t b, float16x8_t c) { in test_vfmlslq_high_f16()
303 float32x4_t test_vfmlalq_lane_low_f16(float32x4_t a, float16x8_t b, float16x4_t c) { in test_vfmlalq_lane_low_f16()
395 float32x4_t test_vfmlalq_lane_high_f16(float32x4_t a, float16x8_t b, float16x4_t c) { in test_vfmlalq_lane_high_f16()
447 float32x2_t test_vfmlal_laneq_low_f16(float32x2_t a, float16x4_t b, float16x8_t c) { in test_vfmlal_laneq_low_f16()
499 float32x2_t test_vfmlal_laneq_high_f16(float32x2_t a, float16x4_t b, float16x8_t c) { in test_vfmlal_laneq_high_f16()
591 float32x4_t test_vfmlalq_laneq_low_f16(float32x4_t a, float16x8_t b, float16x8_t c) { in test_vfmlalq_laneq_low_f16()
683 float32x4_t test_vfmlalq_laneq_high_f16(float32x4_t a, float16x8_t b, float16x8_t c) { in test_vfmlalq_laneq_high_f16()
[all …]
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x25-minmax-neonfp16arith-acc2.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
167 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
168 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
171 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
172 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
173 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
174 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
178 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
179 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
[all …]
Dup16x25-minmax-neonfp16arith.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
167 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
168 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
171 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
172 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
173 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
174 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
178 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
179 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
[all …]
Dup16x9-minmax-neonfp16arith-acc2.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
88 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
91 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
92 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
93 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
94 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
98 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
99 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
[all …]
Dup16x9-minmax-neonfp16arith.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
88 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
91 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
92 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
93 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
94 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
98 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
99 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
[all …]
Dup8x25-minmax-neonfp16arith-acc2.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
167 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
170 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
171 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
174 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
175 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
176 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
178 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
179 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
[all …]
Dup8x25-minmax-neonfp16arith.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
167 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
170 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
171 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
174 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
175 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
178 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
179 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
182 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
[all …]
Dup16x4-minmax-neonfp16arith-acc2.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
63 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
66 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
67 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
68 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
69 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
73 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
74 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
[all …]
Dup16x4-minmax-neonfp16arith.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
63 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
66 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
67 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
68 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
69 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
73 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
74 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
[all …]
Dup8x9-minmax-neonfp16arith-acc2.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
90 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
91 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
94 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
95 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
96 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
98 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
99 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2()
[all …]
Dup8x9-minmax-neonfp16arith.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
87 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
90 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
91 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
94 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
95 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
98 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
99 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
102 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith()
[all …]
Dup8x4-minmax-neonfp16arith-acc2.c33 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
34 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
62 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
65 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
66 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
69 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
70 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
71 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
73 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
74 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
[all …]
/external/XNNPACK/src/f16-gavgpool/
D7p7x-minmax-neonfp16arith-c8.c40 const float16x8_t vi0 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
41 const float16x8_t vi1 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
42 const float16x8_t vi2 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
43 const float16x8_t vi3 = vld1q_f16(i3); i3 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
44 const float16x8_t vi4 = vld1q_f16(i4); i4 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
45 const float16x8_t vi5 = vld1q_f16(i5); i5 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
46 const float16x8_t vi6 = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
48 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
49 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
50 const float16x8_t vsum45 = vaddq_f16(vi4, vi5); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
[all …]
D7x-minmax-neonfp16arith-c8.c52 const float16x8_t vscale = vld1q_dup_f16(&params->scale); in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
53 const float16x8_t vmin = vld1q_dup_f16(&params->min); in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
54 const float16x8_t vmax = vld1q_dup_f16(&params->max); in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
57 const float16x8_t vi0 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
58 const float16x8_t vi1 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
59 const float16x8_t vi2 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
60 const float16x8_t vi3 = vld1q_f16(i3); i3 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
61 const float16x8_t vi4 = vld1q_f16(i4); i4 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
62 const float16x8_t vi5 = vld1q_f16(i5); i5 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
63 const float16x8_t vi6 = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8()
[all …]
/external/llvm-project/clang/test/CodeGen/arm-mve-intrinsics/
Dvcmlaq.c12 float16x8_t test_vcmlaq_f16(float16x8_t a, float16x8_t b, float16x8_t c) in test_vcmlaq_f16()
40 float16x8_t test_vcmlaq_rot90_f16(float16x8_t a, float16x8_t b, float16x8_t c) in test_vcmlaq_rot90_f16()
68 float16x8_t test_vcmlaq_rot180_f16(float16x8_t a, float16x8_t b, float16x8_t c) in test_vcmlaq_rot180_f16()
96 float16x8_t test_vcmlaq_rot270_f16(float16x8_t a, float16x8_t b, float16x8_t c) in test_vcmlaq_rot270_f16()
126 float16x8_t test_vcmlaq_m_f16(float16x8_t a, float16x8_t b, float16x8_t c, mve_pred16_t p) in test_vcmlaq_m_f16()
158 float16x8_t test_vcmlaq_rot90_m_f16(float16x8_t a, float16x8_t b, float16x8_t c, mve_pred16_t p) in test_vcmlaq_rot90_m_f16()
190 float16x8_t test_vcmlaq_rot180_m_f16(float16x8_t a, float16x8_t b, float16x8_t c, mve_pred16_t p) in test_vcmlaq_rot180_m_f16()
222 float16x8_t test_vcmlaq_rot270_m_f16(float16x8_t a, float16x8_t b, float16x8_t c, mve_pred16_t p) in test_vcmlaq_rot270_m_f16()
Dvcmulq.c12 float16x8_t test_vcmulq_f16(float16x8_t a, float16x8_t b) in test_vcmulq_f16()
40 float16x8_t test_vcmulq_rot90_f16(float16x8_t a, float16x8_t b) in test_vcmulq_rot90_f16()
68 float16x8_t test_vcmulq_rot180_f16(float16x8_t a, float16x8_t b) in test_vcmulq_rot180_f16()
96 float16x8_t test_vcmulq_rot270_f16(float16x8_t a, float16x8_t b) in test_vcmulq_rot270_f16()
126 float16x8_t test_vcmulq_m_f16(float16x8_t inactive, float16x8_t a, float16x8_t b, mve_pred16_t p) in test_vcmulq_m_f16()
158 float16x8_t test_vcmulq_rot90_m_f16(float16x8_t inactive, float16x8_t a, float16x8_t b, mve_pred16_… in test_vcmulq_rot90_m_f16()
190 float16x8_t test_vcmulq_rot180_m_f16(float16x8_t inactive, float16x8_t a, float16x8_t b, mve_pred16… in test_vcmulq_rot180_m_f16()
222 float16x8_t test_vcmulq_rot270_m_f16(float16x8_t inactive, float16x8_t a, float16x8_t b, mve_pred16… in test_vcmulq_rot270_m_f16()
254 float16x8_t test_vcmulq_x_f16(float16x8_t a, float16x8_t b, mve_pred16_t p) in test_vcmulq_x_f16()
286 float16x8_t test_vcmulq_rot90_x_f16(float16x8_t a, float16x8_t b, mve_pred16_t p) in test_vcmulq_rot90_x_f16()
[all …]
Dvrnd.c12 float16x8_t test_vrndaq_f16(float16x8_t a) in test_vrndaq_f16()
40 float16x8_t test_vrndmq_f16(float16x8_t a) in test_vrndmq_f16()
68 float16x8_t test_vrndpq_f16(float16x8_t a) in test_vrndpq_f16()
96 float16x8_t test_vrndq_f16(float16x8_t a) in test_vrndq_f16()
124 float16x8_t test_vrndxq_f16(float16x8_t a) in test_vrndxq_f16()
152 float16x8_t test_vrndnq_f16(float16x8_t a) in test_vrndnq_f16()
182 float16x8_t test_vrndaq_m_f16(float16x8_t inactive, float16x8_t a, mve_pred16_t p) in test_vrndaq_m_f16()
214 float16x8_t test_vrndmq_m_f16(float16x8_t inactive, float16x8_t a, mve_pred16_t p) in test_vrndmq_m_f16()
246 float16x8_t test_vrndnq_m_f16(float16x8_t inactive, float16x8_t a, mve_pred16_t p) in test_vrndnq_m_f16()
278 float16x8_t test_vrndpq_m_f16(float16x8_t inactive, float16x8_t a, mve_pred16_t p) in test_vrndpq_m_f16()
[all …]
/external/XNNPACK/src/f16-gemm/gen/
D1x16-minmax-neonfp16arith-ld64.c45 …float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t)); in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
46 …float16x8_t vacc0x89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t)); in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
52 …const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
53 …const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
59 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0); in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
64 …const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
65 …const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
71 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1); in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
76 …const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
77 …const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
[all …]
/external/XNNPACK/src/f16-gemm/gen-inc/
D1x16inc-minmax-neonfp16arith-ld64.c47 …float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
48 …float16x8_t vacc0x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
54 …const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
55 …const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
61 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0); in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
66 …const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
67 …const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
73 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1); in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
78 …const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
79 …const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64()
[all …]
D8x8inc-minmax-neonfp16arith-ld64.c89 …float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
90 …float16x8_t vacc1x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
91 …float16x8_t vacc2x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
92 …float16x8_t vacc3x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
93 …float16x8_t vacc4x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
94 …float16x8_t vacc5x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
95 …float16x8_t vacc6x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
96 …float16x8_t vacc7x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
109 …const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
121 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0); in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
[all …]
D4x16inc-minmax-neonfp16arith-ld64.c65 …float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
66 …float16x8_t vacc0x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
67 …float16x8_t vacc1x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
68 …float16x8_t vacc1x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
69 …float16x8_t vacc2x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
70 …float16x8_t vacc2x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
71 …float16x8_t vacc3x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
72 …float16x8_t vacc3x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
81 …const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
82 …const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8… in xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64()
[all …]
/external/XNNPACK/src/f16-igemm/gen/
D1x16-minmax-neonfp16arith-ld64.c47 …float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t)); in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
48 …float16x8_t vacc0x89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t)); in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
63 …const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x… in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
64 …const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x… in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
70 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0); in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
75 …const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x… in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
76 …const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x… in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
82 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1); in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
87 …const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x… in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
88 …const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x… in xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64()
[all …]

1234567