Home
last modified time | relevance | path

Searched refs:vi8x0123 (Results 1 – 25 of 50) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x9-minmax-sse.c161 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
167 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
233 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
237 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
283 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
285 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
Dup8x9-wasmsimd.c159 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
165 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
229 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
233 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
278 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
280 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
Dup8x9-minmax-neon.c146 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
150 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
198 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
200 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
244 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
246 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
Dup8x9-minmax-sse-acc2.c161 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
167 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
236 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
240 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
288 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
290 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
Dup8x9-minmax-neonfma.c146 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
150 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
198 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
200 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
244 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
246 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
Dup8x9-minmax-wasmsimd-arm-acc2.c161 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
167 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
237 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
241 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
289 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
291 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
Dup8x9-minmax-wasmsimd-x86.c161 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
167 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
234 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
238 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
284 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
286 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
Dup8x9-minmax-neonfma-acc2.c146 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
150 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
201 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
203 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
249 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
251 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
Dup8x9-minmax-wasmsimd-arm.c161 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
167 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
234 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
238 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
284 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
286 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
Dup8x9-minmax-neon-acc2.c146 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
150 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
201 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
203 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
249 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
251 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
Dup4x9-minmax-wasmsimd-arm.c136 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
140 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
187 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
189 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
Dup4x9-wasmsimd.c134 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
138 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
183 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
185 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
Dup4x9-minmax-sse.c136 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
140 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
186 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
188 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
Dup4x9-minmax-neonfma.c121 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
123 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
167 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
169 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
Dup4x9-minmax-neon.c121 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
123 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
167 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
169 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
Dup8x9-minmax-wasmsimd-x86-acc2.c161 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
167 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
237 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
241 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
289 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
291 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
Dup4x9-minmax-wasmsimd-x86-acc2.c136 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
140 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
189 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
191 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
Dup4x9-minmax-neon-acc2.c121 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
123 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
169 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
171 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
Dup4x9-minmax-sse-acc2.c136 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
140 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
188 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
190 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
Dup4x9-minmax-wasmsimd-x86.c136 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
140 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
187 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
189 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
Dup4x9-minmax-neonfma-acc2.c121 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
123 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
169 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
171 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
Dup4x9-minmax-wasmsimd-arm-acc2.c136 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
140 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
189 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
191 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c216 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
220 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
365 const v128_t vi8x0123 = wasm_v128_load(i8); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
367 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c216 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
220 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
362 const __m128 vi8x0123 = _mm_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
364 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c201 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
203 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
313 const float32x4_t vi8x0123 = vld1q_f32(i8); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
315 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()

12