/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-wasmrelaxedsimd-fma.c | 289 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local 408 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_ukernel_up4x25__wasmrelaxedsimd_fma() local
|
D | up4x25-wasmsimd.c | 289 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 408 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up8x25-wasmrelaxedsimd-fma.c | 351 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local 528 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local 647 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma() local
|
D | up8x25-wasmsimd.c | 351 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 528 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local 647 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
|
D | up4x25-minmax-neonfma.c | 250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local 360 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
|
D | up4x25-minmax-neon-acc2.c | 250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local 362 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
|
D | up4x25-minmax-neonfma-acc2.c | 250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local 362 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
|
D | up4x25-minmax-wasmrelaxedsimd-fma.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local 412 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma() local
|
D | up4x25-minmax-wasmrelaxedsimd.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local 412 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd() local
|
D | up4x25-minmax-neon.c | 250 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local 360 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
|
D | up4x25-minmax-wasmrelaxedsimd-fma-acc2.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local 414 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_fma_acc2() local
|
D | up4x25-minmax-sse-acc2.c | 291 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local 413 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
|
D | up4x25-minmax-sse.c | 291 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 411 const __m128 vk20x0123 = _mm_load_ps(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 414 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 414 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|
D | up4x25-minmax-wasmsimd-arm.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local 412 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
|
D | up4x25-minmax-wasmrelaxedsimd-acc2.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local 414 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmrelaxedsimd_acc2() local
|
D | up4x25-minmax-wasmsimd-x86.c | 291 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 412 const v128_t vk20x0123 = wasm_v128_load(w + 84); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up8x25-minmax-sse-acc2.c | 353 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 535 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local 657 const __m128 vk20x0123 = _mm_load_ps(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
|
D | up8x25-minmax-neonfma-acc2.c | 312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 442 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local 554 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
|
D | up8x25-minmax-wasmrelaxedsimd-fma.c | 353 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local 533 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local 653 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma() local
|
D | up8x25-minmax-neonfma.c | 312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 439 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local 549 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
|
D | up8x25-minmax-neon.c | 312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 439 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local 549 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
|
D | up8x25-minmax-neon-acc2.c | 312 const float32x4_t vk20x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 442 const float32x4_t vk20x0123 = vld1q_f32(w + 164); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local 554 const float32x4_t vk20x0123 = vld1q_f32(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
|
D | up8x25-minmax-wasmsimd-arm-acc2.c | 353 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 536 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local 658 const v128_t vk20x0123 = wasm_v128_load(w + 168); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
|