Searched refs:va0000 (Results 1 – 5 of 5) sorted by relevance
/external/XNNPACK/src/f32-ppmm/gen/ |
D | 4x8-minmax-wasmsimd-arm-splat.c | 70 const v128_t va0000 = wasm_v32x4_shuffle(va0123, va0123, 0, 0, 0, 0); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_arm_splat() local 75 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0000, vb0123)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_arm_splat() 79 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0000, vb4567)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_arm_splat()
|
D | 4x8-minmax-sse.c | 68 const __m128 va0000 = _mm_shuffle_ps(va0123, va0123, _MM_SHUFFLE(0, 0, 0, 0)); in xnn_f32_ppmm_minmax_ukernel_4x8__sse() local 73 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0000, vb0123)); in xnn_f32_ppmm_minmax_ukernel_4x8__sse() 77 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0000, vb4567)); in xnn_f32_ppmm_minmax_ukernel_4x8__sse()
|
D | 4x8-minmax-wasmsimd-x86-splat.c | 68 const v128_t va0000 = wasm_v32x4_shuffle(va0123, va0123, 0, 0, 0, 0); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() local 73 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0000, vb0123)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat() 77 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0000, vb4567)); in xnn_f32_ppmm_minmax_ukernel_4x8__wasmsimd_x86_splat()
|
D | 4x8-minmax-neonfma.c | 76 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0); in xnn_f32_ppmm_minmax_ukernel_4x8__neonfma() local 81 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123); in xnn_f32_ppmm_minmax_ukernel_4x8__neonfma() 85 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567); in xnn_f32_ppmm_minmax_ukernel_4x8__neonfma()
|
D | 8x8-minmax-neonfma.c | 109 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() local 118 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 126 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma()
|