| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 125 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
222 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
303 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 83 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
129 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 125 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
224 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
307 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 137 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
267 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
375 float vo1p1 = vi2x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 149 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
249 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
305 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
|
| D | 3x3p1-minmax-neon-2x4-acc2.c | 85 float32x4_t vo1p1 = vmulq_lane_f32(vi2x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
156 float32x4_t vo1p1 = vmulq_lane_f32(vi2x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c | 88 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
158 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
|
| D | 3x3p1-minmax-neonfma-2x4-acc2.c | 85 float32x4_t vo1p1 = vmulq_lane_f32(vi2x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
156 float32x4_t vo1p1 = vmulq_lane_f32(vi2x4567, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
|
| D | 3x3s2p1-minmax-scalar-2x1-acc2.c | 91 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2() local
150 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c | 88 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
158 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 149 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
251 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
309 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c | 103 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2() local
166 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2() local
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c | 103 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2() local
166 v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2() local
|
| D | 3x3p1-minmax-ssse3-2x4-acc2.c | 96 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2() local
160 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2() local
|
| D | 3x3s2p1-minmax-neonfma-2x4-acc2.c | 90 float32x4_t vo1p1 = vmulq_lane_f32(vi2x8ACE9BDF.val[0], vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
163 float32x4_t vo1p1 = vmulq_lane_f32(vi2x8ACE, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 172 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
309 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
389 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
|
| D | 5x5p2-minmax-neon-2x4-acc2.c | 97 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
243 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
385 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c | 100 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local
247 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local
390 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c | 100 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local
245 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local
386 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local
|
| D | 5x5p2-minmax-neonfma-2x4-acc2.c | 97 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
243 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
385 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c | 100 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local
247 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local
390 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c | 100 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local
245 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local
386 v128_t vo1p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local
|
| D | 3x3p1-minmax-sse-2x4-acc2.c | 113 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
203 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
| D | 5x5p2-minmax-neonfma-2x4-acc3.c | 97 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
245 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
389 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc3() local
|
| D | 5x5p2-minmax-neon-2x4-acc3.c | 97 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
245 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
389 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc3() local
|