| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-2x1.c | 81 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
125 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
220 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
301 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
|
| D | 5x5p2-minmax-scalar-2x1.c | 123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
218 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
297 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 81 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
127 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
222 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
305 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
|
| D | 5x5p2-minmax-scalar-3x1.c | 134 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
261 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
366 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
|
| D | 3x3p1-minmax-scalar-3x1.c | 90 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
148 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 134 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
264 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
372 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 147 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
247 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
303 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 147 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
245 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
299 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
|
| D | 3x3p1-minmax-neonfma-2x4.c | 74 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
143 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
|
| D | 3x3p1-minmax-neon-2x4.c | 74 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
143 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-2x4.c | 77 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4() local
145 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4() local
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-2x4.c | 77 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4() local
145 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4() local
|
| D | 3x3s2p1-minmax-scalar-2x1.c | 89 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1() local
146 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1() local
|
| D | 3x3s2p1-minmax-scalar-2x1-acc2.c | 89 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2() local
148 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2() local
|
| D | 3x3p1-minmax-neonfma-2x4-acc2.c | 74 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
145 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
|
| D | 3x3p1-minmax-neon-2x4-acc2.c | 74 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
145 float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c | 77 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
147 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c | 77 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
147 v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
|
| D | 3x3p1-minmax-scalar-4x1.c | 99 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
171 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 147 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
249 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
307 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c | 101 v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4() local
162 v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4() local
|
| D | 3x3p1-minmax-ssse3-2x4.c | 94 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4() local
156 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4() local
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c | 101 v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4() local
162 v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4() local
|