| /external/XNNPACK/src/f32-spmm/gen/ |
| D | 32x1-minmax-neon.c | 76 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
84 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
92 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
134 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
138 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
142 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
172 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
174 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
176 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
201 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
[all …]
|
| D | 32x1-minmax-neonfma.c | 76 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
84 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
92 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
134 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
138 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
142 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
172 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
174 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
176 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
201 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
[all …]
|
| D | 12x1-minmax-neon.c | 60 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_12x1__neon() local
63 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x1__neon()
66 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_12x1__neon()
97 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_12x1__neon() local
99 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x1__neon()
101 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_12x1__neon()
126 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_12x1__neon() local
127 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x1__neon()
128 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_12x1__neon()
|
| D | 12x1-minmax-neonfma.c | 60 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma() local
63 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma()
66 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma()
97 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma() local
99 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma()
101 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma()
126 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma() local
127 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma()
128 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma()
|
| D | 16x1-minmax-wasmsimd-arm.c | 61 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm() local
65 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm()
69 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm()
101 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm() local
103 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm()
105 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm()
131 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm() local
132 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm()
133 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm()
|
| D | 16x1-minmax-wasmsimd-x86.c | 61 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local
65 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86()
69 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86()
101 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local
103 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86()
105 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86()
131 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local
132 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86()
133 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86()
|
| D | 16x1-minmax-neon.c | 63 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local
67 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon()
71 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon()
103 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local
105 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon()
107 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon()
132 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local
133 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon()
134 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon()
|
| D | 16x1-minmax-neonfma.c | 63 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local
67 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma()
71 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma()
103 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local
105 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma()
107 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma()
132 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local
133 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma()
134 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma()
|
| D | 16x1-minmax-sse.c | 61 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__sse() local
65 vout0123 = _mm_max_ps(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__sse()
69 _mm_storeu_ps(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__sse()
101 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__sse() local
103 vout0123 = _mm_max_ps(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__sse()
105 _mm_storeu_ps(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__sse()
130 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__sse() local
131 vout0123 = _mm_max_ps(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__sse()
132 _mm_storeu_ps(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__sse()
|
| D | 32x1-minmax-wasmsimd-arm.c | 73 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
81 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
89 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
131 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
135 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
139 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
170 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
172 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
174 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
200 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
[all …]
|
| D | 32x1-minmax-wasmsimd-x86-pipelined.c | 87 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
95 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
103 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
145 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
149 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
153 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
184 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
186 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
188 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
214 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
[all …]
|
| D | 32x1-minmax-neon-pipelined.c | 87 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
95 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
103 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
148 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
152 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
156 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
189 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
191 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
193 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
221 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
[all …]
|
| D | 32x1-minmax-wasmsimd-arm-pipelined.c | 87 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
95 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
103 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
145 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
149 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
153 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
184 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
186 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
188 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
214 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
[all …]
|
| D | 32x1-minmax-neonfma-pipelined.c | 87 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
95 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
103 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
148 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
152 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
156 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
189 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
191 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
193 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
221 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
[all …]
|
| D | 32x1-minmax-sse.c | 73 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
81 vout0123 = _mm_max_ps(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
89 _mm_storeu_ps(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
131 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
135 vout0123 = _mm_max_ps(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
139 _mm_storeu_ps(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
169 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
171 vout0123 = _mm_max_ps(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
173 _mm_storeu_ps(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
198 __m128 vout0123 = _mm_min_ps(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
[all …]
|
| D | 32x1-minmax-wasmsimd-x86.c | 73 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
81 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
89 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
131 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
135 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
139 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
170 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
172 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
174 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
200 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
[all …]
|
| D | 16x1-minmax-wasmsimd-arm-pipelined.c | 71 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined() local
75 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined()
79 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined()
111 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined() local
113 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined()
115 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined()
141 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined() local
142 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined()
143 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined()
|
| D | 16x1-minmax-neon-pipelined.c | 70 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local
74 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined()
78 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined()
112 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local
114 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined()
116 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined()
143 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local
144 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined()
145 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined()
|
| D | 16x1-minmax-wasmsimd-x86-pipelined.c | 71 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local
75 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined()
79 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined()
111 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local
113 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined()
115 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined()
141 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local
142 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined()
143 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined()
|
| D | 16x1-minmax-neonfma-pipelined.c | 70 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined() local
74 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined()
78 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined()
112 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined() local
114 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined()
116 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined()
143 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined() local
144 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined()
145 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined()
|
| D | 8x1-minmax-wasmsimd-x86.c | 55 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86() local
57 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86()
59 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86()
86 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86() local
87 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86()
88 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_x86()
|
| D | 8x1-minmax-wasmsimd-arm.c | 55 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm() local
57 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm()
59 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm()
86 v128_t vout0123 = wasm_f32x4_min(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm() local
87 vout0123 = wasm_f32x4_max(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm()
88 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_8x1__wasmsimd_arm()
|
| D | 16x1-minmax-neonfma-x2.c | 104 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2() local
108 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2()
112 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2()
144 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2() local
146 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2()
148 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2()
173 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2() local
174 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2()
175 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma_x2()
|
| D | 16x1-minmax-neon-x2.c | 104 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2() local
108 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2()
112 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2()
144 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2() local
146 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2()
148 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2()
173 float32x4_t vout0123 = vminq_f32(vacc0123, vmax); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2() local
174 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2()
175 vst1q_f32(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__neon_x2()
|
| D | 16x1-minmax-wasmsimd-x86-x2.c | 101 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2() local
105 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2()
109 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2()
141 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2() local
143 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2()
145 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2()
171 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2() local
172 vout0123 = wasm_f32x4_pmax(vmin, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2()
173 wasm_v128_store(output, vout0123); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2()
|