/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-scalar-1x1-acc4.c | 114 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local 180 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local 236 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
|
D | 3x3p1-minmax-scalar-1x1-acc4.c | 79 float vo0p3 = vi0x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local 108 float vo0p3 = vi0x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
|
D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 126 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local 191 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local 225 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
|
D | 5x5p2-minmax-scalar-1x1-acc5.c | 114 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local 181 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local 238 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
|
D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 126 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local 192 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local 227 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
|
D | 3x3s2p1-minmax-scalar-1x1-acc4.c | 84 float vo0p3 = vi0x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local 117 float vo0p3 = vi0x1 * vk01; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local
|
D | 3x3p1-minmax-neon-1x4-acc4.c | 81 float32x4_t vo0p3 = vmulq_lane_f32(vi0x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local 135 float32x4_t vo0p3 = vmulq_lane_f32(vi0x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local
|
D | 3x3p1-minmax-neonfma-1x4-acc4.c | 81 float32x4_t vo0p3 = vmulq_lane_f32(vi0x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local 135 float32x4_t vo0p3 = vmulq_lane_f32(vi0x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c | 84 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4() local 137 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c | 84 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4() local 137 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4() local
|
D | 3x3s2p1-minmax-neon-1x4-acc4.c | 77 float32x4_t vo0p3 = vmulq_lane_f32(vi2x8ACE9BDF.val[0], vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() local 128 float32x4_t vo0p3 = vmulq_lane_f32(vi2x8ACE, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() local
|
D | 3x3s2p1-minmax-neonfma-1x4-acc4.c | 77 float32x4_t vo0p3 = vmulq_lane_f32(vi2x8ACE9BDF.val[0], vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() local 128 float32x4_t vo0p3 = vmulq_lane_f32(vi2x8ACE, vw89, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() local
|
D | 3x3p1-minmax-ssse3-1x4-acc4.c | 90 __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4() local 137 __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 97 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4() local 143 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 97 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4() local 143 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4() local
|
D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c | 93 v128_t vo0p3 = wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4() local 146 v128_t vo0p3 = wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4() local
|
D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c | 93 v128_t vo0p3 = wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local 146 v128_t vo0p3 = wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 93 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 202 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local 308 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 93 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 203 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local 310 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc5() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 93 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 202 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local 308 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4_acc4() local
|
D | 5x5p2-minmax-neonfma-1x4-acc4.c | 90 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 200 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local 307 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc4() local
|
D | 5x5p2-minmax-neon-1x4-acc4.c | 90 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 200 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local 307 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 93 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 203 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local 310 v128_t vo0p3 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc5() local
|
D | 3x3p1-minmax-sse-1x4-acc4.c | 106 __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local 173 __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
|
D | 5x5p2-minmax-neonfma-1x4-acc5.c | 90 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 201 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local 309 float32x4_t vo0p3 = vmulq_lane_f32(vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc5() local
|