Home
last modified time | relevance | path

Searched refs:vk11x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
378 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
379 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c237 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
238 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
375 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
376 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
215 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
326 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
327 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma.c214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
215 vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
324 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
325 vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon.c214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
215 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
324 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
325 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-wasmsimd.c235 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
236 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
372 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma-acc2.c214 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
215 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
326 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
327 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
238 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
376 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
377 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
378 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
379 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse-acc2.c237 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
238 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
377 const __m128 vk11x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
378 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-wasmsimd-arm.c237 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
238 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
376 const v128_t vk11x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
377 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
251 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
406 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
407 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
518 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
519 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c272 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
274 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
478 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
479 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
616 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
617 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
274 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
479 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
480 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
617 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
618 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon.c249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
251 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
403 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
404 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
513 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
514 vacc0123p0 = vmlaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-wasmsimd-arm-acc2.c272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
482 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
483 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
622 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
623 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
482 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
483 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
622 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
623 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-neonfma.c249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
251 vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
403 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
404 vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
513 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
514 vacc0123p0 = vfmaq_f32(vacc0123p0, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-sse-acc2.c272 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
274 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
481 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
482 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
621 const __m128 vk11x0123 = _mm_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
622 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-wasmsimd.c270 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
272 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
474 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
475 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
611 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
612 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-wasmsimd-x86.c272 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
274 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
479 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
480 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
617 const v128_t vk11x0123 = wasm_v128_load(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
618 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-neonfma-acc2.c249 const float32x4_t vk11x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
251 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
406 const float32x4_t vk11x0123 = vld1q_f32(w + 92); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
407 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
518 const float32x4_t vk11x0123 = vld1q_f32(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
519 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()