Home
last modified time | relevance | path

Searched refs:vi16x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c264 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
395 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
397 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c264 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
397 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
399 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c264 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
268 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
394 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
396 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c264 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
268 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
396 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
398 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c262 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
266 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
391 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
393 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c233 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
235 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
343 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
345 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c264 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
395 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
397 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c233 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
235 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
345 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
347 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c264 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
268 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
397 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
399 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c233 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
235 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
345 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
347 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c233 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
235 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
343 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
345 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c313 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
319 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
505 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
509 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
635 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
637 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c282 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
286 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
422 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
424 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
532 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
534 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c282 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
286 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
422 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
424 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
532 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
534 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c311 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
317 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
501 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
505 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
630 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
632 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c282 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
286 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
425 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
427 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
537 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
539 vacc0123p0 = vfmaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c313 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
319 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
506 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
510 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
636 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
638 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c282 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
286 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
425 const float32x4_t vi16x0123 = vld1q_f32(i16); i16 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
427 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
537 const float32x4_t vi16x0123 = vld1q_f32(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
539 vacc0123p0 = vmlaq_f32(vacc0123p0, vi16x0123, vk16x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c313 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
319 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
506 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
510 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
636 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
638 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c313 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
319 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
508 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
512 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
640 const __m128 vi16x0123 = _mm_loadu_ps(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
642 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c313 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
319 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
509 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
513 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
641 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
643 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c313 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
319 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
509 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
513 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
641 const v128_t vi16x0123 = wasm_v128_load(i16); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
643 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi16x0123, vk16x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()