| /external/llvm-project/llvm/test/MC/ARM/ |
| D | neont2-minmax-encoding.s | 67 vmin.s8 d1, d2, d3
68 vmin.s16 d4, d5, d6
69 vmin.s32 d7, d8, d9
70 vmin.u8 d10, d11, d12
71 vmin.u16 d13, d14, d15
72 vmin.u32 d16, d17, d18
73 vmin.f32 d19, d20, d21
75 vmin.s8 d2, d3
76 vmin.s16 d5, d6
77 vmin.s32 d8, d9
[all …]
|
| D | neon-minmax-encoding.s | 65 vmin.s8 d1, d2, d3
66 vmin.s16 d4, d5, d6
67 vmin.s32 d7, d8, d9
68 vmin.u8 d10, d11, d12
69 vmin.u16 d13, d14, d15
70 vmin.u32 d16, d17, d18
71 vmin.f32 d19, d20, d21
73 vmin.s8 d2, d3
74 vmin.s16 d5, d6
75 vmin.s32 d8, d9
[all …]
|
| D | mve-minmax.s | 17 # CHECK: vmin.s8 q3, q0, q7 @ encoding: [0x00,0xef,0x5e,0x66]
18 # CHECK-NOFP: vmin.s8 q3, q0, q7 @ encoding: [0x00,0xef,0x5e,0x66]
19 vmin.s8 q3, q0, q7
21 # CHECK: vmin.s16 q0, q1, q2 @ encoding: [0x12,0xef,0x54,0x06]
22 # CHECK-NOFP: vmin.s16 q0, q1, q2 @ encoding: [0x12,0xef,0x54,0x06]
23 vmin.s16 q0, q1, q2
25 # CHECK: vmin.s32 q0, q1, q2 @ encoding: [0x22,0xef,0x54,0x06]
26 # CHECK-NOFP: vmin.s32 q0, q1, q2 @ encoding: [0x22,0xef,0x54,0x06]
27 vmin.s32 q0, q1, q2
29 # CHECK: vmin.u8 q0, q1, q2 @ encoding: [0x02,0xff,0x54,0x06]
[all …]
|
| /external/llvm/test/MC/ARM/ |
| D | neon-minmax-encoding.s | 65 vmin.s8 d1, d2, d3
66 vmin.s16 d4, d5, d6
67 vmin.s32 d7, d8, d9
68 vmin.u8 d10, d11, d12
69 vmin.u16 d13, d14, d15
70 vmin.u32 d16, d17, d18
71 vmin.f32 d19, d20, d21
73 vmin.s8 d2, d3
74 vmin.s16 d5, d6
75 vmin.s32 d8, d9
[all …]
|
| D | neont2-minmax-encoding.s | 67 vmin.s8 d1, d2, d3
68 vmin.s16 d4, d5, d6
69 vmin.s32 d7, d8, d9
70 vmin.u8 d10, d11, d12
71 vmin.u16 d13, d14, d15
72 vmin.u32 d16, d17, d18
73 vmin.f32 d19, d20, d21
75 vmin.s8 d2, d3
76 vmin.s16 d5, d6
77 vmin.s32 d8, d9
[all …]
|
| /external/capstone/suite/MC/ARM/ |
| D | neon-minmax-encoding.s.cs | 30 0x13,0x16,0x02,0xf2 = vmin.s8 d1, d2, d3
31 0x16,0x46,0x15,0xf2 = vmin.s16 d4, d5, d6
32 0x19,0x76,0x28,0xf2 = vmin.s32 d7, d8, d9
33 0x1c,0xa6,0x0b,0xf3 = vmin.u8 d10, d11, d12
34 0x1f,0xd6,0x1e,0xf3 = vmin.u16 d13, d14, d15
35 0xb2,0x06,0x61,0xf3 = vmin.u32 d16, d17, d18
36 0xa5,0x3f,0x64,0xf2 = vmin.f32 d19, d20, d21
37 0x13,0x26,0x02,0xf2 = vmin.s8 d2, d2, d3
38 0x16,0x56,0x15,0xf2 = vmin.s16 d5, d5, d6
39 0x19,0x86,0x28,0xf2 = vmin.s32 d8, d8, d9
[all …]
|
| D | neont2-minmax-encoding.s.cs | 30 0x02,0xef,0x13,0x16 = vmin.s8 d1, d2, d3
31 0x15,0xef,0x16,0x46 = vmin.s16 d4, d5, d6
32 0x28,0xef,0x19,0x76 = vmin.s32 d7, d8, d9
33 0x0b,0xff,0x1c,0xa6 = vmin.u8 d10, d11, d12
34 0x1e,0xff,0x1f,0xd6 = vmin.u16 d13, d14, d15
35 0x61,0xff,0xb2,0x06 = vmin.u32 d16, d17, d18
36 0x64,0xef,0xa5,0x3f = vmin.f32 d19, d20, d21
37 0x02,0xef,0x13,0x26 = vmin.s8 d2, d2, d3
38 0x15,0xef,0x16,0x56 = vmin.s16 d5, d5, d6
39 0x28,0xef,0x19,0x86 = vmin.s32 d8, d8, d9
[all …]
|
| /external/XNNPACK/src/f32-spmm/gen/ |
| D | 8x4-minmax-scalar.c | 31 const float vmin = params->scalar.min; in xnn_f32_spmm_minmax_ukernel_8x4__scalar() local
155 vout0x0 = math_max_f32(vout0x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
156 vout1x0 = math_max_f32(vout1x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
157 vout2x0 = math_max_f32(vout2x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
158 vout3x0 = math_max_f32(vout3x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
159 vout4x0 = math_max_f32(vout4x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
160 vout5x0 = math_max_f32(vout5x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
161 vout6x0 = math_max_f32(vout6x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
162 vout7x0 = math_max_f32(vout7x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
163 vout0x1 = math_max_f32(vout0x1, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__scalar()
[all …]
|
| D | 8x2-minmax-scalar.c | 31 const float vmin = params->scalar.min; in xnn_f32_spmm_minmax_ukernel_8x2__scalar() local
105 vout0x0 = math_max_f32(vout0x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
106 vout1x0 = math_max_f32(vout1x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
107 vout2x0 = math_max_f32(vout2x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
108 vout3x0 = math_max_f32(vout3x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
109 vout4x0 = math_max_f32(vout4x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
110 vout5x0 = math_max_f32(vout5x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
111 vout6x0 = math_max_f32(vout6x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
112 vout7x0 = math_max_f32(vout7x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
113 vout0x1 = math_max_f32(vout0x1, vmin); in xnn_f32_spmm_minmax_ukernel_8x2__scalar()
[all …]
|
| D | 32x1-minmax-wasmsimd-x86.c | 32 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
81 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
82 vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
83 vout89AB = wasm_v128_bitselect(vmin, vout89AB, wasm_f32x4_lt(vout89AB, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
84 voutCDEF = wasm_v128_bitselect(vmin, voutCDEF, wasm_f32x4_lt(voutCDEF, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
85 voutGHIJ = wasm_v128_bitselect(vmin, voutGHIJ, wasm_f32x4_lt(voutGHIJ, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
86 voutKLMN = wasm_v128_bitselect(vmin, voutKLMN, wasm_f32x4_lt(voutKLMN, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
87 voutOPQR = wasm_v128_bitselect(vmin, voutOPQR, wasm_f32x4_lt(voutOPQR, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
88 voutSTUV = wasm_v128_bitselect(vmin, voutSTUV, wasm_f32x4_lt(voutSTUV, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
135 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
[all …]
|
| D | 32x4-minmax-neonfma.c | 32 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
157 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
158 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
159 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
160 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
161 voutGHIJn0 = vmaxq_f32(voutGHIJn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
162 voutKLMNn0 = vmaxq_f32(voutKLMNn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
163 voutOPQRn0 = vmaxq_f32(voutOPQRn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
164 voutSTUVn0 = vmaxq_f32(voutSTUVn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
165 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
[all …]
|
| D | 32x1-minmax-wasmsimd-x86-pipelined.c | 32 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
95 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
96 vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
97 vout89AB = wasm_v128_bitselect(vmin, vout89AB, wasm_f32x4_lt(vout89AB, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
98 voutCDEF = wasm_v128_bitselect(vmin, voutCDEF, wasm_f32x4_lt(voutCDEF, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
99 voutGHIJ = wasm_v128_bitselect(vmin, voutGHIJ, wasm_f32x4_lt(voutGHIJ, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
100 voutKLMN = wasm_v128_bitselect(vmin, voutKLMN, wasm_f32x4_lt(voutKLMN, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
101 voutOPQR = wasm_v128_bitselect(vmin, voutOPQR, wasm_f32x4_lt(voutOPQR, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
102 voutSTUV = wasm_v128_bitselect(vmin, voutSTUV, wasm_f32x4_lt(voutSTUV, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
149 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
[all …]
|
| D | 8x1-minmax-scalar.c | 31 const float vmin = params->scalar.min; in xnn_f32_spmm_minmax_ukernel_8x1__scalar() local
80 vout0x0 = math_max_f32(vout0x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
81 vout1x0 = math_max_f32(vout1x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
82 vout2x0 = math_max_f32(vout2x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
83 vout3x0 = math_max_f32(vout3x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
84 vout4x0 = math_max_f32(vout4x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
85 vout5x0 = math_max_f32(vout5x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
86 vout6x0 = math_max_f32(vout6x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
87 vout7x0 = math_max_f32(vout7x0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
149 vout0 = math_max_f32(vout0, vmin); in xnn_f32_spmm_minmax_ukernel_8x1__scalar()
[all …]
|
| D | 32x1-minmax-wasmsimd-x86-pipelined-x2.c | 32 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local
135 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
136 vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
137 vout89AB = wasm_v128_bitselect(vmin, vout89AB, wasm_f32x4_lt(vout89AB, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
138 voutCDEF = wasm_v128_bitselect(vmin, voutCDEF, wasm_f32x4_lt(voutCDEF, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
139 voutGHIJ = wasm_v128_bitselect(vmin, voutGHIJ, wasm_f32x4_lt(voutGHIJ, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
140 voutKLMN = wasm_v128_bitselect(vmin, voutKLMN, wasm_f32x4_lt(voutKLMN, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
141 voutOPQR = wasm_v128_bitselect(vmin, voutOPQR, wasm_f32x4_lt(voutOPQR, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
142 voutSTUV = wasm_v128_bitselect(vmin, voutSTUV, wasm_f32x4_lt(voutSTUV, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
189 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
[all …]
|
| D | 16x4-minmax-neonfma.c | 32 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local
104 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
105 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
106 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
107 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
108 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
109 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
110 vout89ABn1 = vmaxq_f32(vout89ABn1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
111 voutCDEFn1 = vmaxq_f32(voutCDEFn1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
112 vout0123n2 = vmaxq_f32(vout0123n2, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
[all …]
|
| D | 32x1-minmax-wasmsimd-x86-x2.c | 32 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() local
150 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
151 vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
152 vout89AB = wasm_v128_bitselect(vmin, vout89AB, wasm_f32x4_lt(vout89AB, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
153 voutCDEF = wasm_v128_bitselect(vmin, voutCDEF, wasm_f32x4_lt(voutCDEF, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
154 voutGHIJ = wasm_v128_bitselect(vmin, voutGHIJ, wasm_f32x4_lt(voutGHIJ, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
155 voutKLMN = wasm_v128_bitselect(vmin, voutKLMN, wasm_f32x4_lt(voutKLMN, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
156 voutOPQR = wasm_v128_bitselect(vmin, voutOPQR, wasm_f32x4_lt(voutOPQR, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
157 voutSTUV = wasm_v128_bitselect(vmin, voutSTUV, wasm_f32x4_lt(voutSTUV, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
204 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
[all …]
|
| D | 32x2-minmax-neonfma.c | 32 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
109 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
110 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
111 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
112 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
113 voutGHIJn0 = vmaxq_f32(voutGHIJn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
114 voutKLMNn0 = vmaxq_f32(voutKLMNn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
115 voutOPQRn0 = vmaxq_f32(voutOPQRn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
116 voutSTUVn0 = vmaxq_f32(voutSTUVn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
117 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
[all …]
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x8-minmax-wasmsimd-x86-loadsplat.c | 121 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() local
122 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
123 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
124 vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
125 vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
126 vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
127 vacc5x0123 = wasm_v128_bitselect(vmin, vacc5x0123, wasm_f32x4_lt(vacc5x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
128 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
129 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
130 vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
[all …]
|
| D | 5x8-minmax-wasmsimd-x86-loadsplat.c | 109 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() local
110 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
111 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
112 vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
113 vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
114 vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
115 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
116 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
117 vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
118 vacc3x4567 = wasm_v128_bitselect(vmin, vacc3x4567, wasm_f32x4_lt(vacc3x4567, vmin)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
[all …]
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 6x8inc-minmax-wasmsimd-x86-loadsplat.c | 123 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() local
124 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
125 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
126 vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
127 vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
128 vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
129 vacc5x0123 = wasm_v128_bitselect(vmin, vacc5x0123, wasm_f32x4_lt(vacc5x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
130 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
131 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
132 vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
[all …]
|
| D | 5x8inc-minmax-wasmsimd-x86-loadsplat.c | 111 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() local
112 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
113 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
114 vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
115 vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
116 vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
117 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
118 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
119 vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
120 vacc3x4567 = wasm_v128_bitselect(vmin, vacc3x4567, wasm_f32x4_lt(vacc3x4567, vmin)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
[all …]
|
| /external/XNNPACK/src/f32-vmulcaddc/gen/ |
| D | c8-minmax-wasmsimd-x86-2x.c | 44 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x() local
68 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
69 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
70 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
71 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
100 vacc0 = wasm_v128_bitselect(vmin, vacc0, wasm_f32x4_lt(vacc0, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
101 vacc1 = wasm_v128_bitselect(vmin, vacc1, wasm_f32x4_lt(vacc1, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
126 vacc0 = wasm_v128_bitselect(vmin, vacc0, wasm_f32x4_lt(vacc0, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
127 vacc1 = wasm_v128_bitselect(vmin, vacc1, wasm_f32x4_lt(vacc1, vmin)); in xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmsimd_x86_2x()
|
| /external/XNNPACK/src/f32-ppmm/gen/ |
| D | 4x4-minmax-scalar.c | 118 const float vmin = params->scalar.min; in xnn_f32_ppmm_minmax_ukernel_4x4__scalar() local
119 vacc0x0 = math_max_f32(vacc0x0, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
120 vacc1x0 = math_max_f32(vacc1x0, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
121 vacc2x0 = math_max_f32(vacc2x0, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
122 vacc3x0 = math_max_f32(vacc3x0, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
123 vacc0x1 = math_max_f32(vacc0x1, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
124 vacc1x1 = math_max_f32(vacc1x1, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
125 vacc2x1 = math_max_f32(vacc2x1, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
126 vacc3x1 = math_max_f32(vacc3x1, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
127 vacc0x2 = math_max_f32(vacc0x2, vmin); in xnn_f32_ppmm_minmax_ukernel_4x4__scalar()
[all …]
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 6x8-minmax-wasmsimd-x86-loadsplat.c | 150 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat() local
151 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
152 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
153 vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
154 vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
155 vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
156 vacc5x0123 = wasm_v128_bitselect(vmin, vacc5x0123, wasm_f32x4_lt(vacc5x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
157 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
158 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
159 vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat()
[all …]
|
| D | 5x8-minmax-wasmsimd-x86-loadsplat.c | 135 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat() local
136 vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
137 vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
138 vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
139 vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
140 vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
141 vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
142 vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
143 vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
144 vacc3x4567 = wasm_v128_bitselect(vmin, vacc3x4567, wasm_f32x4_lt(vacc3x4567, vmin)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat()
[all …]
|