/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-scalar-1x1-acc3.c | 113 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local 178 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local 233 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
|
D | 3x3p1-minmax-scalar-1x1-acc3.c | 73 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local 105 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local
|
D | 5x5p2-minmax-scalar-1x1-acc4.c | 113 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local 179 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local 235 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
|
D | 3x3p1-minmax-scalar-1x1-acc4.c | 73 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local 106 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
|
D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 125 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local 190 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local 224 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
|
D | 5x5p2-minmax-scalar-1x1-acc5.c | 113 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local 180 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local 237 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
|
D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 125 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local 191 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local 226 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
|
D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 125 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local 189 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local 222 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
|
D | 3x3s2p1-minmax-scalar-1x1-acc3.c | 75 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3() local 114 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3() local
|
D | 3x3s2p1-minmax-scalar-1x1-acc4.c | 75 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local 115 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c | 76 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3() local 128 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3() local
|
D | 3x3p1-minmax-neon-1x4-acc3.c | 75 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local 128 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c | 76 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3() local 128 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3() local
|
D | 3x3p1-minmax-neonfma-1x4-acc3.c | 75 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local 128 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local
|
D | 5x5p2-minmax-scalar-2x1-acc3.c | 126 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local 225 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local 308 float vo0p2 = vi2x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c | 76 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4() local 129 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4() local
|
D | 3x3s2p1-minmax-neon-1x4-acc3.c | 75 float32x4_t vo0p2 = vmulq_lane_f32(vi1x8ACE9BDF.val[0], vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() local 125 float32x4_t vo0p2 = vmulq_lane_f32(vi1x8ACE, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c | 76 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4() local 129 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4() local
|
D | 3x3s2p1-minmax-neonfma-1x4-acc3.c | 75 float32x4_t vo0p2 = vmulq_lane_f32(vi1x8ACE9BDF.val[0], vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() local 125 float32x4_t vo0p2 = vmulq_lane_f32(vi1x8ACE, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() local
|
D | 3x3p1-minmax-neonfma-1x4-acc4.c | 75 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local 129 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local
|
D | 3x3p1-minmax-neon-1x4-acc4.c | 75 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local 129 float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local
|
/external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-neonfp16arith-1x4-acc3.c | 75 float16x4_t vo0p2 = vmul_laneq_f16(vi1x8ACE9BDF.val[0], vw01234567, 5); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc3() local 125 float16x4_t vo0p2 = vmul_laneq_f16(vi1x8ACE, vw01234567, 5); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc3() local
|
D | 3x3p1-minmax-neonfp16arith-1x8-acc3.c | 76 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3() local 132 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3() local
|
D | 3x3p1-minmax-neonfp16arith-1x8-acc4.c | 76 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4() local 133 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4() local
|
D | 3x3s2p1-minmax-neonfp16arith-1x4-acc4.c | 75 float16x4_t vo0p2 = vmul_laneq_f16(vi1x8ACE9BDF.val[0], vw01234567, 5); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc4() local 126 float16x4_t vo0p2 = vmul_laneq_f16(vi1x8ACE, vw01234567, 5); in xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc4() local
|