| /external/XNNPACK/src/f32-conv-hwc2chw/ |
| D | 3x3s2p1c3x4-wasmsimd-2x2.c | 115 const v128_t vi2x1 = wasm_v128_load(i2); i2 += 4; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
122 …vo1x1 = wasm_f32x4_add(vo1x1, wasm_f32x4_mul(vk00c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
135 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk20c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
190 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk01c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
203 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk21c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
211 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk01c1, wasm_v32x4_shuffle(vi2x1, vi2x1, 1, 1, 1, 1))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
224 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk21c1, wasm_v32x4_shuffle(vi2x1, vi2x1, 1, 1, 1, 1))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
239 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk01c2, wasm_v32x4_shuffle(vi2x1, vi2x1, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
252 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk21c2, wasm_v32x4_shuffle(vi2x1, vi2x1, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
260 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk02c0, wasm_v32x4_shuffle(vi2x1, vi2x1, 3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
[all …]
|
| D | 3x3s2p1c3x4-sse-2x2.c | 114 const __m128 vi2x1 = _mm_loadu_ps(i2); i2 += 4; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
121 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk00c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
134 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk20c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
189 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
202 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
210 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c1, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(1, 1, 1, 1))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
223 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c1, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(1, 1, 1, 1))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
238 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk01c2, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
251 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk21c2, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
259 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk02c0, _mm_shuffle_ps(vi2x1, vi2x1, _MM_SHUFFLE(3, 3, 3, 3))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
[all …]
|
| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-3x1.c | 77 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
101 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
107 vo2p0 += vi2x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
109 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
111 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
117 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
159 vo2p0 += vi2x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
161 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
163 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
|
| D | 3x3p1-minmax-scalar-2x1.c | 70 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
89 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
95 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
96 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
101 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
134 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
135 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 70 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
89 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
95 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
96 vo0p1 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
101 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
136 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
137 vo0p1 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
|
| D | 3x3p1-minmax-scalar-4x1.c | 84 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
113 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
120 vo2p0 += vi2x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
123 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
126 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
133 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
185 vo2p0 += vi2x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
188 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
191 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 3x3p1-minmax-scalar-1x1-acc3.c | 63 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local
77 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
81 vo0p2 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
85 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
109 vo0p2 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
|
| D | 3x3p1-minmax-scalar-1x1.c | 63 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1() local
77 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
81 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
85 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
107 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
|
| D | 3x3p1-minmax-scalar-1x1-acc2.c | 63 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2() local
77 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
81 vo0p1 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
85 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
108 vo0p1 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-2x1.c | 93 float vi2x1 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
135 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
143 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
144 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
153 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
230 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
238 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
239 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
248 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
310 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
[all …]
|
| D | 3x3p1-minmax-scalar-1x1-acc4.c | 63 float vi2x1 = *i2++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
77 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
81 vo0p1 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
85 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
110 vo0p1 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-3x1.c | 100 float vi2x1 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
151 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
159 vo2p0 += vi2x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
161 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
163 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
175 vi2x1 = vi2x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
278 vi2x0 = vi2x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
286 vo2p0 += vi2x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
288 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
290 vo0p0 += vi2x1 * vk21; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
[all …]
|
| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p1c3x8-neon-2x1.c | 171 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
179 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
181 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
194 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
196 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
203 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
205 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
218 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
220 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
227 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
[all …]
|
| D | 3x3s2p1c3x8-neonfma-2x1.c | 173 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
181 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
183 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
196 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
198 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
205 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
207 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
220 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
222 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
229 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x2.c | 106 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
118 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
141 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
143 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
155 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
157 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
180 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
182 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
274 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neonfma-2x2.c | 108 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
118 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
120 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
143 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
145 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
237 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
239 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
262 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
264 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
276 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x2.c | 108 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
118 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
120 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
143 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
145 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
157 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
159 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
182 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
184 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
276 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neon-2x2.c | 106 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
118 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
141 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
143 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
235 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
237 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk01c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
260 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
262 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
274 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 200 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
205 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
207 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
220 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
222 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
229 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
231 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
244 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
246 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
253 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 198 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
203 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
205 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
218 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
220 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
227 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
229 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk01c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
242 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
244 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
251 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
[all …]
|
| D | 3x3s2p0p1c3x4-neonfma-2x2.c | 103 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
111 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
126 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
135 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
150 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
211 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
226 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
235 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
250 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
266 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neon-2x2.c | 101 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
109 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
124 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
185 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
200 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
209 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
224 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
240 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
255 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
264 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x2.c | 101 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
109 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
124 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
133 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
148 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
209 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
224 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
233 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
248 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
264 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 103 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
111 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
126 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
187 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
202 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
211 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
226 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
242 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
257 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
266 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neonfma-2x1.c | 144 const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
151 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
160 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
166 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
175 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
181 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
190 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
196 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
205 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
247 vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
[all …]
|