/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmsimd.c | 223 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 364 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 370 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-sse.c | 225 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 367 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-neon-acc2.c | 206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 318 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neonfma.c | 206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 316 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-neon.c | 206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 316 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up4x25-minmax-neonfma-acc2.c | 206 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 318 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmsimd-x86.c | 225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 368 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 370 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-sse-acc2.c | 225 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 369 const __m128 vk9x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-wasmsimd-arm.c | 225 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 368 const v128_t vk9x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up8x25-wasmsimd.c | 252 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 462 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 603 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up8x25-minmax-neon-acc2.c | 235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 398 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 510 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-sse.c | 254 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 466 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 608 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
|
D | up8x25-minmax-wasmsimd-arm.c | 254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 467 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 609 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
|
D | up8x25-minmax-neon.c | 235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 395 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 505 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 470 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 614 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 470 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 614 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
|
D | up8x25-minmax-neonfma.c | 235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 395 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 505 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-sse-acc2.c | 254 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 469 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 613 const __m128 vk9x0123 = _mm_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-wasmsimd-x86.c | 254 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 467 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 609 const v128_t vk9x0123 = wasm_v128_load(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
|
D | up8x25-minmax-neonfma-acc2.c | 235 const float32x4_t vk9x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 398 const float32x4_t vk9x0123 = vld1q_f32(w + 76); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 510 const float32x4_t vk9x0123 = vld1q_f32(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|