Home
last modified time | relevance | path

Searched refs:vk10x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c231 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
374 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
375 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c231 const __m128 vk10x0123 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
232 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
371 const __m128 vk10x0123 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
372 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c210 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
211 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
322 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
323 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma.c210 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
211 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
320 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
321 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon.c210 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
211 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
320 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
321 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-wasmsimd.c229 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
230 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
368 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
369 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma-acc2.c210 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
211 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
322 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
323 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c231 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
372 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c231 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
374 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
375 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse-acc2.c231 const __m128 vk10x0123 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
232 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
373 const __m128 vk10x0123 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
374 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-wasmsimd-arm.c231 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
372 const v128_t vk10x0123 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c242 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
244 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
402 const float32x4_t vk10x0123 = vld1q_f32(w + 84); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
403 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
514 const float32x4_t vk10x0123 = vld1q_f32(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
515 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c263 const __m128 vk10x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
265 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
472 const __m128 vk10x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
473 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
612 const __m128 vk10x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
613 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c263 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
265 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
473 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
474 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
613 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
614 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon.c242 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
244 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
399 const float32x4_t vk10x0123 = vld1q_f32(w + 84); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
400 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
509 const float32x4_t vk10x0123 = vld1q_f32(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
510 vacc0123p0 = vmlaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-wasmsimd-arm-acc2.c263 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
265 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
476 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
477 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
618 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
619 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c263 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
265 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
476 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
477 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
618 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
619 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-neonfma.c242 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
244 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
399 const float32x4_t vk10x0123 = vld1q_f32(w + 84); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
400 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
509 const float32x4_t vk10x0123 = vld1q_f32(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
510 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-sse-acc2.c263 const __m128 vk10x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
265 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
475 const __m128 vk10x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
476 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
617 const __m128 vk10x0123 = _mm_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
618 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-wasmsimd.c261 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
263 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
468 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
469 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
607 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
608 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-wasmsimd-x86.c263 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
265 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
473 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
474 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
613 const v128_t vk10x0123 = wasm_v128_load(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
614 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi10x0123, vk10x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-neonfma-acc2.c242 const float32x4_t vk10x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
244 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
402 const float32x4_t vk10x0123 = vld1q_f32(w + 84); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
403 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
514 const float32x4_t vk10x0123 = vld1q_f32(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
515 vacc0123p0 = vfmaq_f32(vacc0123p0, vi10x0123, vk10x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()