Home
last modified time | relevance | path

Searched refs:vk12x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-arm-acc2.c243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
244 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
382 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
383 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-sse.c243 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
244 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
379 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
380 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-neon-acc2.c218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
219 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
330 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
331 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neonfma.c218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
219 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
328 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
329 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-neon.c218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
219 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
328 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
329 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup4x25-wasmsimd.c241 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
242 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
376 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
377 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma-acc2.c218 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
219 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
330 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
331 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-x86.c243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
244 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
380 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
381 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
244 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
382 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
383 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse-acc2.c243 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
244 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
381 const __m128 vk12x0123 = _mm_load_ps(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
382 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-minmax-wasmsimd-arm.c243 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
244 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
380 const v128_t vk12x0123 = wasm_v128_load(w + 52); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
381 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
258 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
410 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
411 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
522 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
523 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-sse.c281 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
283 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
484 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
485 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
620 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
621 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-wasmsimd-arm.c281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
283 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
485 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
486 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
621 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
622 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon.c256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
258 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
407 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
408 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
517 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
518 vacc0123p0 = vmlaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-wasmsimd-arm-acc2.c281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
283 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
488 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
489 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
626 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
627 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
283 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
488 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
489 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
626 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
627 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-neonfma.c256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
258 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
407 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
408 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
517 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
518 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-minmax-sse-acc2.c281 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
283 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
487 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
488 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
625 const __m128 vk12x0123 = _mm_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
626 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-wasmsimd.c279 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
281 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
480 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
481 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
615 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
616 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-wasmsimd-x86.c281 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
283 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
485 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
486 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
621 const v128_t vk12x0123 = wasm_v128_load(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
622 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi12x0123, vk12x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-neonfma-acc2.c256 const float32x4_t vk12x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
258 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
410 const float32x4_t vk12x0123 = vld1q_f32(w + 100); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
411 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
522 const float32x4_t vk12x0123 = vld1q_f32(w + 104); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
523 vacc0123p0 = vfmaq_f32(vacc0123p0, vi12x0123, vk12x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()