Home
last modified time | relevance | path

Searched refs:vk4x0123 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x9-minmax-neon-acc2.c120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
122 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
186 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
187 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
234 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
235 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
Dup8x9-minmax-neonfma.c120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
122 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
183 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
184 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
229 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
230 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
Dup8x9-minmax-sse.c129 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
131 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
212 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
213 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
268 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
269 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
Dup8x9-minmax-sse-acc2.c129 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
131 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
215 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
216 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
273 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
274 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
Dup8x9-minmax-wasmsimd-arm.c129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
131 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
213 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
214 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
269 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
270 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
Dup8x9-minmax-neon.c120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
122 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
183 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
184 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
229 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
230 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
Dup8x9-minmax-wasmsimd-arm-acc2.c129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
131 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
216 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
217 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
274 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
275 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
Dup8x9-minmax-neonfma-acc2.c120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
122 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
186 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
187 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
234 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
235 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
Dup8x9-wasmsimd.c127 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
129 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
208 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
209 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
263 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
264 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
Dup8x9-minmax-wasmsimd-x86.c129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
131 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
213 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
214 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
269 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
270 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
Dup4x9-wasmsimd.c113 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
114 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
168 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
169 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
Dup4x9-minmax-wasmsimd-arm.c115 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
116 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
172 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
173 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
Dup4x9-minmax-sse.c115 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
116 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
171 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
172 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
Dup4x9-minmax-neon.c106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
107 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
152 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
153 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
Dup4x9-minmax-neonfma.c106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
107 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
152 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
153 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
Dup8x9-minmax-wasmsimd-x86-acc2.c129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
131 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
216 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
217 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
274 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
275 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
Dup4x9-minmax-wasmsimd-x86.c115 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
116 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
172 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
173 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
Dup4x9-minmax-sse-acc2.c115 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
116 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
173 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
174 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
Dup4x9-minmax-neon-acc2.c106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
107 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
154 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
155 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
Dup4x9-minmax-wasmsimd-arm-acc2.c115 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
116 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
174 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
175 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
Dup4x9-minmax-wasmsimd-x86-acc2.c115 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
116 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
174 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
175 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
Dup4x9-minmax-neonfma-acc2.c106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
107 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
154 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
155 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c195 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
196 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
348 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
349 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c195 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
196 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
350 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
351 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c195 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
196 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
347 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
348 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()

12