| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
176 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
230 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
|
| D | 3x3p1-minmax-scalar-1x1-acc2.c | 72 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2() local
103 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2() local
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
177 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
232 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
|
| D | 3x3p1-minmax-scalar-1x1-acc3.c | 72 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local
104 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
178 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
234 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
187 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
219 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
|
| D | 3x3s2p1-minmax-scalar-1x1-acc2.c | 74 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2() local
112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2() local
|
| D | 3x3p1-minmax-scalar-1x1-acc4.c | 72 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
105 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
|
| D | 3x3s2p1-minmax-scalar-1x1-acc3.c | 74 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3() local
113 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3() local
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
189 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
223 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
221 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
302 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
179 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
236 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
|
| D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
188 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
221 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
|
| D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
190 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
225 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
|
| D | 3x3p1-minmax-neonfma-1x4-acc2.c | 73 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2() local
125 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2() local
|
| D | 3x3p1-minmax-neon-1x4-acc2.c | 73 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2() local
125 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c | 76 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2() local
127 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c | 76 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2() local
127 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2() local
|
| D | 3x3s2p1-minmax-scalar-1x1-acc4.c | 74 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local
114 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local
|
| D | 3x3p1-minmax-neonfma-1x4-acc3.c | 73 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local
126 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local
|
| D | 3x3s2p1-minmax-neon-1x4-acc2.c | 73 float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() local
122 float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() local
|
| D | 3x3s2p1-minmax-neonfma-1x4-acc2.c | 73 float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() local
122 float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c | 76 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3() local
128 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3() local
|
| D | 3x3p1-minmax-neon-1x4-acc3.c | 73 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local
126 float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 82 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
128 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
|