Home
last modified time | relevance | path

Searched refs:vk6x0123 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x9-minmax-neon-acc2.c134 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
136 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
194 const float32x4_t vk6x0123 = vld1q_f32(w + 52); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
195 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
242 const float32x4_t vk6x0123 = vld1q_f32(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
243 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
Dup8x9-minmax-neonfma.c134 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
136 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
191 const float32x4_t vk6x0123 = vld1q_f32(w + 52); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
192 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
237 const float32x4_t vk6x0123 = vld1q_f32(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
238 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
Dup8x9-minmax-sse.c147 const __m128 vk6x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
149 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
224 const __m128 vk6x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
225 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
276 const __m128 vk6x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
277 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
Dup8x9-minmax-sse-acc2.c147 const __m128 vk6x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
149 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
227 const __m128 vk6x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
228 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
281 const __m128 vk6x0123 = _mm_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
282 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
Dup8x9-minmax-wasmsimd-arm.c147 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
149 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
225 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
226 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
277 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
278 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
Dup8x9-minmax-neon.c134 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
136 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
191 const float32x4_t vk6x0123 = vld1q_f32(w + 52); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
192 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
237 const float32x4_t vk6x0123 = vld1q_f32(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
238 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
Dup8x9-minmax-wasmsimd-arm-acc2.c147 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
149 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
228 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
229 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
282 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
283 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
Dup8x9-minmax-neonfma-acc2.c134 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
136 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
194 const float32x4_t vk6x0123 = vld1q_f32(w + 52); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
195 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
242 const float32x4_t vk6x0123 = vld1q_f32(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
243 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
Dup8x9-wasmsimd.c145 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
147 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
220 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
221 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
271 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
272 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
Dup8x9-minmax-wasmsimd-x86.c147 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
149 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
225 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
226 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
277 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
278 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
Dup4x9-wasmsimd.c125 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
126 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
176 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
177 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
Dup4x9-minmax-wasmsimd-arm.c127 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
128 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
180 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
181 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
Dup4x9-minmax-sse.c127 const __m128 vk6x0123 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
128 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
179 const __m128 vk6x0123 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
180 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
Dup4x9-minmax-neon.c114 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
115 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
160 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
161 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
Dup4x9-minmax-neonfma.c114 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
115 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
160 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
161 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
Dup8x9-minmax-wasmsimd-x86-acc2.c147 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
149 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
228 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
229 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
282 const v128_t vk6x0123 = wasm_v128_load(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
283 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
Dup4x9-minmax-wasmsimd-x86.c127 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
128 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
180 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
181 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
Dup4x9-minmax-sse-acc2.c127 const __m128 vk6x0123 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
128 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
181 const __m128 vk6x0123 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
182 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
Dup4x9-minmax-neon-acc2.c114 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
115 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
162 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
163 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
Dup4x9-minmax-wasmsimd-arm-acc2.c127 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
128 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
182 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
183 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
Dup4x9-minmax-wasmsimd-x86-acc2.c127 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
128 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
182 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
183 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
Dup4x9-minmax-neonfma-acc2.c114 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
115 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
162 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
163 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c207 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
208 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
356 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
357 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c207 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
208 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
358 const v128_t vk6x0123 = wasm_v128_load(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
359 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c207 const __m128 vk6x0123 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
208 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
355 const __m128 vk6x0123 = _mm_load_ps(w + 28); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
356 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()

12