Home
last modified time | relevance | path

Searched refs:vk8x0123 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x9-minmax-sse.c165 const __m128 vk8x0123 = _mm_load_ps(w + 72); 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()
236 const __m128 vk8x0123 = _mm_load_ps(w + 72); 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()
284 const __m128 vk8x0123 = _mm_load_ps(w + 72); 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.c163 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
232 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
279 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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.c148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
199 const float32x4_t vk8x0123 = vld1q_f32(w + 68); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
200 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
245 const float32x4_t vk8x0123 = vld1q_f32(w + 72); 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.c165 const __m128 vk8x0123 = _mm_load_ps(w + 72); 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()
239 const __m128 vk8x0123 = _mm_load_ps(w + 72); 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()
289 const __m128 vk8x0123 = _mm_load_ps(w + 72); 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.c148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
199 const float32x4_t vk8x0123 = vld1q_f32(w + 68); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
200 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
245 const float32x4_t vk8x0123 = vld1q_f32(w + 72); 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.c165 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
240 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
290 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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.c165 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
237 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
285 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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.c148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
202 const float32x4_t vk8x0123 = vld1q_f32(w + 68); 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()
250 const float32x4_t vk8x0123 = vld1q_f32(w + 72); 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.c165 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
237 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
285 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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.c148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
202 const float32x4_t vk8x0123 = vld1q_f32(w + 68); 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()
250 const float32x4_t vk8x0123 = vld1q_f32(w + 72); 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.c139 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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()
188 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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.c137 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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()
184 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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.c139 const __m128 vk8x0123 = _mm_load_ps(w + 36); 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()
187 const __m128 vk8x0123 = _mm_load_ps(w + 36); 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.c122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
168 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; 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.c122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
168 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; 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.c165 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
240 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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()
290 const v128_t vk8x0123 = wasm_v128_load(w + 72); 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.c139 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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()
190 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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.c122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
170 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; 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.c139 const __m128 vk8x0123 = _mm_load_ps(w + 36); 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()
189 const __m128 vk8x0123 = _mm_load_ps(w + 36); 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.c139 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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()
188 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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.c122 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
170 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; 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.c139 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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()
190 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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.c219 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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()
366 const v128_t vk8x0123 = wasm_v128_load(w + 36); 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.c219 const __m128 vk8x0123 = _mm_load_ps(w + 36); 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()
363 const __m128 vk8x0123 = _mm_load_ps(w + 36); 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.c202 const float32x4_t vk8x0123 = vld1q_f32(w); w += 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()
314 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4; 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