/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmrelaxedsimd-fma.c | 265 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local 392 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
|
D | up4x25-wasmsimd.c | 265 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 392 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up8x25-wasmrelaxedsimd-fma.c | 315 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local 504 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local 631 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
|
D | up8x25-wasmsimd.c | 315 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 504 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 631 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up4x25-minmax-neonfma.c | 234 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 344 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-neon-acc2.c | 234 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 346 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neonfma-acc2.c | 234 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 346 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmrelaxedsimd-fma.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local 396 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
|
D | up4x25-minmax-wasmrelaxedsimd.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local 396 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
|
D | up4x25-minmax-neon.c | 234 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 344 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up4x25-minmax-wasmrelaxedsimd-fma-acc2.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local 398 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
|
D | up4x25-minmax-sse-acc2.c | 267 const __m128 vk16x0123 = _mm_load_ps(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 397 const __m128 vk16x0123 = _mm_load_ps(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-sse.c | 267 const __m128 vk16x0123 = _mm_load_ps(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 395 const __m128 vk16x0123 = _mm_load_ps(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 398 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 398 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-wasmsimd-arm.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 396 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up4x25-minmax-wasmrelaxedsimd-acc2.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local 398 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
|
D | up4x25-minmax-wasmsimd-x86.c | 267 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 396 const v128_t vk16x0123 = wasm_v128_load(w + 68); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up8x25-minmax-sse-acc2.c | 317 const __m128 vk16x0123 = _mm_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 511 const __m128 vk16x0123 = _mm_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 641 const __m128 vk16x0123 = _mm_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-neonfma-acc2.c | 284 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 426 const float32x4_t vk16x0123 = vld1q_f32(w + 132); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 538 const float32x4_t vk16x0123 = vld1q_f32(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|
D | up8x25-minmax-wasmrelaxedsimd-fma.c | 317 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local 509 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local 637 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
|
D | up8x25-minmax-neonfma.c | 284 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 423 const float32x4_t vk16x0123 = vld1q_f32(w + 132); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 533 const float32x4_t vk16x0123 = vld1q_f32(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-neon.c | 284 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 423 const float32x4_t vk16x0123 = vld1q_f32(w + 132); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 533 const float32x4_t vk16x0123 = vld1q_f32(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-neon-acc2.c | 284 const float32x4_t vk16x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 426 const float32x4_t vk16x0123 = vld1q_f32(w + 132); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 538 const float32x4_t vk16x0123 = vld1q_f32(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 317 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 512 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 642 const v128_t vk16x0123 = wasm_v128_load(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|