Searched refs:va1111 (Results 1 – 5 of 5) sorted by relevance
/external/XNNPACK/src/f32-ppmm/gen/ |
D | 4x8-minmax-wasmsimd-arm-splat.c | 71 const v128_t va1111 = wasm_v32x4_shuffle(va0123, va0123, 1, 1, 1, 1); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_arm_splat() local 76 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1111, vb0123)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_arm_splat() 80 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1111, vb4567)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_arm_splat()
|
D | 4x8-minmax-sse.c | 69 const __m128 va1111 = _mm_shuffle_ps(va0123, va0123, _MM_SHUFFLE(1, 1, 1, 1)); in xnn_f32_ppmm_minmax_ukernel_4x8__sse() local 74 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1111, vb0123)); in xnn_f32_ppmm_minmax_ukernel_4x8__sse() 78 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1111, vb4567)); in xnn_f32_ppmm_minmax_ukernel_4x8__sse()
|
D | 4x8-minmax-wasmsimd-x86-splat.c | 69 const v128_t va1111 = wasm_v32x4_shuffle(va0123, va0123, 1, 1, 1, 1); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() local 74 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1111, vb0123)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 78 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1111, vb4567)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat()
|
D | 4x8-minmax-neonfma.c | 77 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1); in xnn_f32_ppmm_minmax_ukernel_4x8__neonfma() local 82 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123); in xnn_f32_ppmm_minmax_ukernel_4x8__neonfma() 86 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567); in xnn_f32_ppmm_minmax_ukernel_4x8__neonfma()
|
D | 8x8-minmax-neonfma.c | 110 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() local 119 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 127 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma()
|