| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x3-wasmrelaxedsimd-fma.c | 80 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma() local
82 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma()
|
| D | up8x3-wasmsimd.c | 80 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local
82 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd()
|
| D | up8x3-minmax-wasmrelaxedsimd-fma.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma() local
84 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma()
|
| D | up8x3-minmax-wasmsimd-x86-acc2.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local
84 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2()
|
| D | up8x3-minmax-wasmrelaxedsimd-fma-acc2.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2() local
84 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2()
|
| D | up8x3-minmax-sse.c | 82 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local
84 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse()
|
| D | up8x3-minmax-neon.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local
79 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon()
|
| D | up8x3-minmax-sse-acc2.c | 82 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local
84 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2()
|
| D | up8x3-minmax-neonfma.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local
79 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma()
|
| D | up8x3-minmax-wasmsimd-x86.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local
84 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86()
|
| D | up8x3-minmax-neonfma-acc2.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local
79 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2()
|
| D | up8x3-minmax-wasmsimd-arm-acc2.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local
84 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2()
|
| D | up8x3-minmax-wasmrelaxedsimd.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd() local
84 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd()
|
| D | up8x3-minmax-neon-acc2.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local
79 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2()
|
| D | up8x3-minmax-wasmrelaxedsimd-acc2.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2() local
84 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2()
|
| D | up8x4-wasmrelaxedsimd-fma.c | 85 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma() local
87 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma()
|
| D | up8x3-minmax-wasmsimd-arm.c | 82 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local
84 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm()
|
| D | up8x4-minmax-wasmrelaxedsimd-fma.c | 87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma() local
89 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma()
|
| D | up8x4-wasmsimd.c | 85 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
87 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd()
|
| D | up8x4-minmax-wasmrelaxedsimd-acc2.c | 87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2() local
89 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2()
|
| D | up8x4-minmax-wasmsimd-x86-acc2.c | 87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
89 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
|
| D | up8x4-minmax-wasmsimd-x86.c | 87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
89 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86()
|
| D | up8x4-minmax-neonfma-acc2.c | 82 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
84 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
|
| D | up8x4-minmax-wasmrelaxedsimd.c | 87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd() local
89 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd()
|
| D | up8x4-minmax-sse-acc2.c | 87 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
89 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
|