Home
last modified time | relevance | path

Searched refs:viSTUV (Results 1 – 19 of 19) sorted by relevance

/external/XNNPACK/src/f32-spmm/gen/
D32x1-minmax-wasmsimd-arm-pipelined-x2.c48 v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local
70 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2()
81 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2()
89 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2()
100 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2()
112 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2()
124 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2()
D32x1-minmax-wasmsimd-x86-pipelined-x2.c48 v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local
70 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
81 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
89 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
100 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
112 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
124 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2()
D32x1-minmax-wasmsimd-arm-pipelined.c48 v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
72 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
84 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined()
D32x1-minmax-neon-pipelined.c48 float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
70 vaccSTUV = vmlaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
84 viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined()
D32x1-minmax-neonfma-pipelined.c48 float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
70 vaccSTUV = vfmaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
84 viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined()
D32x1-minmax-wasmsimd-x86-pipelined.c48 v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
72 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
84 viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined()
D32x4-minmax-neonfma.c84 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
97 vaccSTUVn0 = vfmaq_laneq_f32(vaccSTUVn0, viSTUV, vw, 0); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
105 vaccSTUVn1 = vfmaq_laneq_f32(vaccSTUVn1, viSTUV, vw, 1); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
113 vaccSTUVn2 = vfmaq_laneq_f32(vaccSTUVn2, viSTUV, vw, 2); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
121 vaccSTUVn3 = vfmaq_laneq_f32(vaccSTUVn3, viSTUV, vw, 3); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
251 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
264 vaccSTUV = vfmaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
D32x2-minmax-neonfma.c68 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
81 vaccSTUVn0 = vfmaq_lane_f32(vaccSTUVn0, viSTUV, vw, 0); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
89 vaccSTUVn1 = vfmaq_lane_f32(vaccSTUVn1, viSTUV, vw, 1); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
169 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
182 vaccSTUV = vfmaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
D32x1-minmax-neon.c60 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
73 vaccSTUV = vmlaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x1__neon()
D32x1-minmax-neonfma.c60 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
73 vaccSTUV = vfmaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma()
D32x1-minmax-wasmsimd-arm.c60 const v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
70 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm()
D32x1-minmax-sse.c60 const __m128 viSTUV = _mm_loadu_ps(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
70 vaccSTUV = _mm_add_ps(vaccSTUV, _mm_mul_ps(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__sse()
D32x1-minmax-wasmsimd-x86.c60 const v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
70 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86()
D32x1-minmax-neon-x2.c131 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() local
144 vaccSTUV = vmlaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2()
D32x1-minmax-neonfma-x2.c131 const float32x4_t viSTUV = vld1q_f32(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2() local
144 vaccSTUV = vfmaq_f32(vaccSTUV, viSTUV, vw); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2()
D32x1-minmax-wasmsimd-arm-x2.c129 const v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2() local
139 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2()
D32x1-minmax-wasmsimd-x86-x2.c129 const v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() local
139 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2()
D32x1-minmax-wasmsimd-arm-x4.c201 const v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x4() local
211 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x4()
D32x1-minmax-wasmsimd-x86-x4.c201 const v128_t viSTUV = wasm_v128_load(input + 28); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x4() local
211 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw)); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x4()