/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmsimd.c | 313 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 424 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 315 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 430 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-sse.c | 315 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 427 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-neon-acc2.c | 266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 378 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neonfma.c | 266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 376 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-neon.c | 266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 376 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up4x25-minmax-neonfma-acc2.c | 266 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 378 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmsimd-x86.c | 315 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 428 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 315 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 430 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-sse-acc2.c | 315 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 429 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-wasmsimd-arm.c | 315 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 428 const v128_t vk24x0123 = wasm_v128_load(w + 100); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up8x25-wasmsimd.c | 387 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 552 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 663 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up8x25-minmax-neon-acc2.c | 340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 458 const float32x4_t vk24x0123 = vld1q_f32(w + 196); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 570 const float32x4_t vk24x0123 = vld1q_f32(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-sse.c | 389 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 556 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 668 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
|
D | up8x25-minmax-wasmsimd-arm.c | 389 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 557 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 669 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
|
D | up8x25-minmax-neon.c | 340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 455 const float32x4_t vk24x0123 = vld1q_f32(w + 196); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 565 const float32x4_t vk24x0123 = vld1q_f32(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 389 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 560 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 674 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 389 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 560 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 674 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
|
D | up8x25-minmax-neonfma.c | 340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 455 const float32x4_t vk24x0123 = vld1q_f32(w + 196); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 565 const float32x4_t vk24x0123 = vld1q_f32(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-sse-acc2.c | 389 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 559 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 673 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-wasmsimd-x86.c | 389 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 557 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 669 const v128_t vk24x0123 = wasm_v128_load(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
|
D | up8x25-minmax-neonfma-acc2.c | 340 const float32x4_t vk24x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 458 const float32x4_t vk24x0123 = vld1q_f32(w + 196); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 570 const float32x4_t vk24x0123 = vld1q_f32(w + 200); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|