/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmsimd.c | 271 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 396 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-x86.c | 273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 400 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 402 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-sse.c | 273 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 399 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-sse-acc2.c | 273 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 401 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-neonfma.c | 238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 348 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-wasmsimd-arm.c | 273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 400 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up4x25-minmax-neonfma-acc2.c | 238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 350 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 402 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-neon-acc2.c | 238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 350 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neon.c | 238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 348 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up8x25-wasmsimd.c | 324 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 510 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 635 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up8x25-minmax-sse.c | 326 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 514 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 640 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
|
D | up8x25-minmax-neon.c | 291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 427 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 537 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-neonfma.c | 291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 427 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 537 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-neonfma-acc2.c | 291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 430 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 542 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|
D | up8x25-minmax-wasmsimd-arm.c | 326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 515 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 641 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
|
D | up8x25-minmax-neon-acc2.c | 291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 430 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 542 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-wasmsimd-x86.c | 326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 515 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 641 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
|
D | up8x25-minmax-sse-acc2.c | 326 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 517 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 645 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 518 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 646 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 518 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 646 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|