Home
last modified time | relevance | path

Searched refs:vk2x4567 (Results 1 – 25 of 33) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x4-wasmsimd.c85 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()
Dup8x4-minmax-sse-acc2.c87 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()
Dup8x4-minmax-neon.c82 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local
84 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon()
Dup8x4-minmax-wasmsimd-arm-acc2.c87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
89 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
Dup8x4-minmax-neonfma.c82 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local
84 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma()
Dup8x4-minmax-neon-acc2.c82 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
84 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
Dup8x4-minmax-neonfma-acc2.c82 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()
Dup8x4-minmax-wasmsimd-arm.c87 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local
89 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm()
Dup8x4-minmax-wasmsimd-x86.c87 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()
Dup8x4-minmax-sse.c87 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local
89 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse()
Dup8x4-minmax-wasmsimd-x86-acc2.c87 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()
Dup8x9-minmax-sse.c112 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
114 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
Dup8x9-wasmsimd.c110 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
112 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
Dup8x9-minmax-neon.c107 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
109 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
Dup8x9-minmax-sse-acc2.c112 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
114 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
Dup8x9-minmax-neonfma.c107 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
109 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
Dup8x9-minmax-wasmsimd-arm-acc2.c112 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
114 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
Dup8x9-minmax-wasmsimd-x86.c112 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
114 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
Dup8x9-minmax-neonfma-acc2.c107 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
109 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
Dup8x9-minmax-wasmsimd-arm.c112 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
114 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
Dup8x9-minmax-neon-acc2.c107 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
109 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
Dup8x9-minmax-wasmsimd-x86-acc2.c112 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
114 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
Dup8x25-minmax-neon-acc2.c187 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
189 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c192 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
194 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c192 const v128_t vk2x4567 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
194 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi2x4567, vk2x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()

12