Home
last modified time | relevance | path

Searched refs:vk5x0123 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x9-minmax-neon-acc2.c127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
129 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
190 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
191 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
238 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
239 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
Dup8x9-minmax-neonfma.c127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
129 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
187 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
188 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
233 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
234 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
Dup8x9-minmax-sse.c138 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
140 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
218 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
219 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
272 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
273 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
Dup8x9-minmax-sse-acc2.c138 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
140 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
221 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
222 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
277 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
278 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
Dup8x9-minmax-wasmsimd-arm.c138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
140 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
219 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
220 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
273 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
274 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
Dup8x9-minmax-neon.c127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
129 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
187 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
188 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
233 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
234 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
Dup8x9-minmax-wasmsimd-arm-acc2.c138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
140 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
222 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
223 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
278 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
279 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
Dup8x9-minmax-neonfma-acc2.c127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
129 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
190 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
191 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
238 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
239 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
Dup8x9-wasmsimd.c136 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
138 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
214 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
215 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
267 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
Dup8x9-minmax-wasmsimd-x86.c138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
140 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
219 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
220 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
273 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
274 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
Dup4x9-wasmsimd.c119 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
120 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
172 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
173 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
Dup4x9-minmax-wasmsimd-arm.c121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
122 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
176 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
177 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
Dup4x9-minmax-sse.c121 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
122 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
175 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
176 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
Dup4x9-minmax-neon.c110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
111 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
156 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
157 vacc0123p0 = vmlaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
Dup4x9-minmax-neonfma.c110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
111 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
156 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
157 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
Dup8x9-minmax-wasmsimd-x86-acc2.c138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
140 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
222 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
223 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
278 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
279 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
Dup4x9-minmax-wasmsimd-x86.c121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
122 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
176 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
177 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
Dup4x9-minmax-sse-acc2.c121 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
122 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
177 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
178 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
Dup4x9-minmax-neon-acc2.c110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
111 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
158 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
159 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
Dup4x9-minmax-wasmsimd-arm-acc2.c121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
122 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
178 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
179 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
Dup4x9-minmax-wasmsimd-x86-acc2.c121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
122 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
178 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
179 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
Dup4x9-minmax-neonfma-acc2.c110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
111 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
158 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
159 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c201 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
202 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
352 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
353 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c201 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
202 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
354 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
355 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c201 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
202 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
351 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
352 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()

12