/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmsimd.c | 307 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 420 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-x86.c | 309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 424 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 426 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-sse.c | 309 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 423 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-sse-acc2.c | 309 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 425 const __m128 vk23x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-neonfma.c | 262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 372 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-wasmsimd-arm.c | 309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 424 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up4x25-minmax-neonfma-acc2.c | 262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 374 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 309 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 426 const v128_t vk23x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-neon-acc2.c | 262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 374 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neon.c | 262 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 372 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up8x25-wasmsimd.c | 378 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 546 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 659 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up8x25-minmax-sse.c | 380 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 550 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local 664 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
|
D | up8x25-minmax-neon.c | 333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 451 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 561 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-neonfma.c | 333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 451 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 561 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-neonfma-acc2.c | 333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 454 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 566 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|
D | up8x25-minmax-wasmsimd-arm.c | 380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 551 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local 665 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
|
D | up8x25-minmax-neon-acc2.c | 333 const float32x4_t vk23x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 454 const float32x4_t vk23x0123 = vld1q_f32(w + 188); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 566 const float32x4_t vk23x0123 = vld1q_f32(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-wasmsimd-x86.c | 380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 551 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local 665 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
|
D | up8x25-minmax-sse-acc2.c | 380 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 553 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 669 const __m128 vk23x0123 = _mm_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-wasmsimd-x86-acc2.c | 380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 554 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local 670 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 380 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 554 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 670 const v128_t vk23x0123 = wasm_v128_load(w + 192); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|