/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmsimd.c | 241 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 376 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 382 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-sse.c | 243 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 379 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-neon-acc2.c | 218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 330 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neonfma.c | 218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 328 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-neon.c | 218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 328 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up4x25-minmax-neonfma-acc2.c | 218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 330 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmsimd-x86.c | 243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 380 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 382 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-sse-acc2.c | 243 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 381 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-wasmsimd-arm.c | 243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 380 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up8x25-wasmsimd.c | 279 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 480 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 615 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up8x25-minmax-neon-acc2.c | 256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 410 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 522 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-sse.c | 281 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 484 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 620 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
|
D | up8x25-minmax-wasmsimd-arm.c | 281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 485 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 621 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
|
D | up8x25-minmax-neon.c | 256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 407 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 517 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 488 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 626 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 488 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 626 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
|
D | up8x25-minmax-neonfma.c | 256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 407 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 517 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-sse-acc2.c | 281 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 487 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 625 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-wasmsimd-x86.c | 281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 485 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 621 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
|
D | up8x25-minmax-neonfma-acc2.c | 256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 410 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 522 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|