/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmsimd.c | 235 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 372 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-x86.c | 237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 376 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 378 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-sse.c | 237 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 375 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-sse-acc2.c | 237 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 377 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-neonfma.c | 214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 324 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-wasmsimd-arm.c | 237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 376 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up4x25-minmax-neonfma-acc2.c | 214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 326 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 378 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-neon-acc2.c | 214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 326 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neon.c | 214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 324 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up8x25-wasmsimd.c | 270 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 474 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 611 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up8x25-minmax-sse.c | 272 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 478 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 616 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
|
D | up8x25-minmax-neon.c | 249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 403 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 513 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-neonfma.c | 249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 403 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 513 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-neonfma-acc2.c | 249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 406 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 518 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|
D | up8x25-minmax-wasmsimd-arm.c | 272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 479 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 617 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
|
D | up8x25-minmax-neon-acc2.c | 249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 406 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 518 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-wasmsimd-x86.c | 272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 479 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 617 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
|
D | up8x25-minmax-sse-acc2.c | 272 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 481 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 621 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 482 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 622 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 482 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 622 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|