Home
last modified time | relevance | path

Searched refs:vk17x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
274 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
400 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
401 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
402 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
403 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c273 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
274 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
399 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
400 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c273 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
274 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
401 const __m128 vk17x0123 = _mm_load_ps(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
402 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c271 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
272 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
396 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
397 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
239 vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
348 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
349 vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
274 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
400 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
401 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
239 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
350 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
351 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c273 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
402 const v128_t vk17x0123 = wasm_v128_load(w + 72); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
403 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
239 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
350 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
351 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c238 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
239 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
348 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
349 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c326 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
328 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
514 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
515 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
640 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
641 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
293 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
427 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
428 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
537 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
538 vacc0123p0 = vmlaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
293 vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
427 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
428 vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
537 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
538 vacc0123p0 = vfmaq_f32(vacc0123p0, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c324 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
326 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
510 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
511 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
635 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
636 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
293 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
430 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
431 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
542 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
543 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
328 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
515 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
516 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
641 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
642 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c291 const float32x4_t vk17x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
293 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
430 const float32x4_t vk17x0123 = vld1q_f32(w + 140); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
431 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
542 const float32x4_t vk17x0123 = vld1q_f32(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
543 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
328 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
515 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
516 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
641 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
642 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c326 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
328 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
517 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
518 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
645 const __m128 vk17x0123 = _mm_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
646 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
328 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
518 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
519 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
646 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
647 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c326 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
328 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
518 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
519 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
646 const v128_t vk17x0123 = wasm_v128_load(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
647 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()