| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x9-wasmrelaxedsimd-fma.c | 128 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local
130 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma()
|
| D | up8x9-minmax-wasmrelaxedsimd-fma.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local
132 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma()
|
| D | up8x9-minmax-sse.c | 130 const __m128 vk4x4567 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
132 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
|
| D | up8x9-minmax-wasmsimd-arm.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
|
| D | up8x9-minmax-wasmrelaxedsimd-acc2.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2()
|
| D | up8x9-wasmsimd.c | 128 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
130 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
|
| D | up8x9-minmax-neonfma.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
123 vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
|
| D | up8x9-minmax-neon-acc2.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
123 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
|
| D | up8x9-minmax-neon.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
123 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
|
| D | up8x9-minmax-wasmsimd-arm-acc2.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
|
| D | up8x9-minmax-wasmsimd-x86.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
|
| D | up8x9-minmax-wasmrelaxedsimd-fma-acc2.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local
132 vacc4567p0 = __builtin_wasm_fma_f32x4(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2()
|
| D | up8x9-minmax-wasmrelaxedsimd.c | 130 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local
132 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd()
|
| D | up8x9-minmax-neonfma-acc2.c | 121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
123 vacc4567p0 = vfmaq_f32(vacc4567p0, vi4x4567, vk4x4567); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-minmax-sse-acc2.c | 130 const __m128 vk4x4567 = _mm_load_ps(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
132 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
|
| D | up8x9-wasmsimd-acc2.c | 128 const v128_t vk4x4567 = wasm_v128_load(w + 44); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local
130 vacc4567p0 = wasm_f32x4_add(vacc4567p0, wasm_f32x4_mul(vi4x4567, vk4x4567)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2()
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-fp32-avx-mul32.c | 128 …const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * s… in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul32() local
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul32()
|
| D | up8x9-minmax-fp32-sse41-mul32.c | 128 …const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * s… in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32() local
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32()
|
| D | up8x9-minmax-fp32-xop-mul32.c | 133 …const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * s… in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32() local
137 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32()
|
| /external/XNNPACK/src/qc8-dwconv/gen/ |
| D | up8x9-minmax-fp32-sse41-mul32.c | 128 …const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * s… in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32() local
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32()
|
| D | up8x9-minmax-fp32-xop-mul32.c | 133 …const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * s… in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32() local
137 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32()
|
| D | up8x9-minmax-fp32-avx-mul32.c | 128 …const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * s… in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul32() local
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul32()
|
| /external/XNNPACK/src/qu8-dwconv/gen/ |
| D | up8x9-minmax-fp32-xop-mul32.c | 134 …const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintpt… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32() local
138 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32()
|
| D | up8x9-minmax-fp32-sse41-mul32.c | 129 …const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintpt… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32() local
133 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32()
|