| // Auto-generated file. Do not edit! |
| // Template: src/qs8-dwconv/unipass-sse-mul16.c.in |
| // Generator: tools/xngen |
| // |
| // Copyright 2020 Google LLC |
| // |
| // This source code is licensed under the BSD-style license found in the |
| // LICENSE file in the root directory of this source tree. |
| |
| #include <assert.h> |
| |
| #include <smmintrin.h> |
| |
| #include <xnnpack/dwconv.h> |
| |
| |
| void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul16( |
| size_t channels, |
| size_t output_width, |
| const int8_t** input, |
| const void* weights, |
| int8_t* output, |
| size_t input_stride, |
| size_t output_increment, |
| size_t input_offset, |
| const int8_t* zero, |
| const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| { |
| assert(channels != 0); |
| assert(output_width != 0); |
| |
| do { |
| const int8_t* i0 = input[0]; |
| assert(i0 != NULL); |
| if XNN_UNPREDICTABLE(i0 != zero) { |
| i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| } |
| const int8_t* i1 = input[1]; |
| assert(i1 != NULL); |
| if XNN_UNPREDICTABLE(i1 != zero) { |
| i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| } |
| const int8_t* i2 = input[2]; |
| assert(i2 != NULL); |
| if XNN_UNPREDICTABLE(i2 != zero) { |
| i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| } |
| const int8_t* i3 = input[3]; |
| assert(i3 != NULL); |
| if XNN_UNPREDICTABLE(i3 != zero) { |
| i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| } |
| const int8_t* i4 = input[4]; |
| assert(i4 != NULL); |
| if XNN_UNPREDICTABLE(i4 != zero) { |
| i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| } |
| const int8_t* i5 = input[5]; |
| assert(i5 != NULL); |
| if XNN_UNPREDICTABLE(i5 != zero) { |
| i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| } |
| const int8_t* i6 = input[6]; |
| assert(i6 != NULL); |
| if XNN_UNPREDICTABLE(i6 != zero) { |
| i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| } |
| const int8_t* i7 = input[7]; |
| assert(i7 != NULL); |
| if XNN_UNPREDICTABLE(i7 != zero) { |
| i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| } |
| const int8_t* i8 = input[8]; |
| assert(i8 != NULL); |
| if XNN_UNPREDICTABLE(i8 != zero) { |
| i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| } |
| input = (const int8_t**) ((uintptr_t) input + input_stride); |
| |
| size_t c = channels; |
| const void* w = weights; |
| for (; c >= 24; c -= 24) { |
| __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); |
| __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); |
| __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8)); |
| __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12)); |
| __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16)); |
| __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20)); |
| |
| |
| const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); |
| const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567); |
| const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t))); |
| const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567); |
| const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); |
| const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF); |
| const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t))); |
| const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF); |
| const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16)); |
| const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(vi0xGHIJKLMN); |
| const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t))); |
| const __m128i vxk0xGHIJKLMN = _mm_cvtepi8_epi16(vk0xGHIJKLMN); |
| i0 += 24; |
| |
| |
| const __m128i vp0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); |
| const __m128i vp0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); |
| const __m128i vp0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF); |
| const __m128i vp0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF); |
| const __m128i vp0xGHIJKLMNlo = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN); |
| const __m128i vp0xGHIJKLMNhi = _mm_mulhi_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp0x01234567lo, vp0x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp0x89ABCDEFlo, vp0x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp0x89ABCDEFlo, vp0x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp0xGHIJKLMNlo, vp0xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp0xGHIJKLMNlo, vp0xGHIJKLMNhi)); |
| |
| const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); |
| const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567); |
| const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t))); |
| const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567); |
| const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); |
| const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF); |
| const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t))); |
| const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF); |
| const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16)); |
| const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(vi1xGHIJKLMN); |
| const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t))); |
| const __m128i vxk1xGHIJKLMN = _mm_cvtepi8_epi16(vk1xGHIJKLMN); |
| i1 += 24; |
| |
| |
| const __m128i vp1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); |
| const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); |
| const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); |
| const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); |
| const __m128i vp1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); |
| const __m128i vp1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp1x89ABCDEFlo, vp1x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp1x89ABCDEFlo, vp1x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp1xGHIJKLMNlo, vp1xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp1xGHIJKLMNlo, vp1xGHIJKLMNhi)); |
| |
| const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); |
| const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567); |
| const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t))); |
| const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567); |
| const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); |
| const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF); |
| const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t))); |
| const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF); |
| const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16)); |
| const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(vi2xGHIJKLMN); |
| const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t))); |
| const __m128i vxk2xGHIJKLMN = _mm_cvtepi8_epi16(vk2xGHIJKLMN); |
| i2 += 24; |
| |
| |
| const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); |
| const __m128i vp2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); |
| const __m128i vp2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF); |
| const __m128i vp2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF); |
| const __m128i vp2xGHIJKLMNlo = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN); |
| const __m128i vp2xGHIJKLMNhi = _mm_mulhi_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp2x89ABCDEFlo, vp2x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp2x89ABCDEFlo, vp2x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp2xGHIJKLMNlo, vp2xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp2xGHIJKLMNlo, vp2xGHIJKLMNhi)); |
| |
| const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); |
| const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567); |
| const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t))); |
| const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567); |
| const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8)); |
| const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF); |
| const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t))); |
| const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF); |
| const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16)); |
| const __m128i vxi3xGHIJKLMN = _mm_cvtepi8_epi16(vi3xGHIJKLMN); |
| const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t))); |
| const __m128i vxk3xGHIJKLMN = _mm_cvtepi8_epi16(vk3xGHIJKLMN); |
| i3 += 24; |
| |
| |
| const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); |
| const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); |
| const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); |
| const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); |
| const __m128i vp3xGHIJKLMNlo = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN); |
| const __m128i vp3xGHIJKLMNhi = _mm_mulhi_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp3x89ABCDEFlo, vp3x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp3x89ABCDEFlo, vp3x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp3xGHIJKLMNlo, vp3xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp3xGHIJKLMNlo, vp3xGHIJKLMNhi)); |
| |
| const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); |
| const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567); |
| const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t))); |
| const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); |
| const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); |
| const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF); |
| const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t))); |
| const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF); |
| const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); |
| const __m128i vxi4xGHIJKLMN = _mm_cvtepi8_epi16(vi4xGHIJKLMN); |
| const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t))); |
| const __m128i vxk4xGHIJKLMN = _mm_cvtepi8_epi16(vk4xGHIJKLMN); |
| i4 += 24; |
| |
| |
| const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); |
| const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); |
| const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); |
| const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); |
| const __m128i vp4xGHIJKLMNlo = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN); |
| const __m128i vp4xGHIJKLMNhi = _mm_mulhi_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp4x89ABCDEFlo, vp4x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp4x89ABCDEFlo, vp4x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp4xGHIJKLMNlo, vp4xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp4xGHIJKLMNlo, vp4xGHIJKLMNhi)); |
| |
| const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); |
| const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567); |
| const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t))); |
| const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); |
| const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8)); |
| const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF); |
| const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t))); |
| const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF); |
| const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16)); |
| const __m128i vxi5xGHIJKLMN = _mm_cvtepi8_epi16(vi5xGHIJKLMN); |
| const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t))); |
| const __m128i vxk5xGHIJKLMN = _mm_cvtepi8_epi16(vk5xGHIJKLMN); |
| i5 += 24; |
| |
| |
| const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); |
| const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); |
| const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); |
| const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); |
| const __m128i vp5xGHIJKLMNlo = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN); |
| const __m128i vp5xGHIJKLMNhi = _mm_mulhi_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp5x01234567lo, vp5x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp5x89ABCDEFlo, vp5x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp5x89ABCDEFlo, vp5x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp5xGHIJKLMNlo, vp5xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp5xGHIJKLMNlo, vp5xGHIJKLMNhi)); |
| |
| const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); |
| const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567); |
| const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t))); |
| const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567); |
| const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8)); |
| const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF); |
| const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t))); |
| const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF); |
| const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); |
| const __m128i vxi6xGHIJKLMN = _mm_cvtepi8_epi16(vi6xGHIJKLMN); |
| const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t))); |
| const __m128i vxk6xGHIJKLMN = _mm_cvtepi8_epi16(vk6xGHIJKLMN); |
| i6 += 24; |
| |
| |
| const __m128i vp6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); |
| const __m128i vp6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); |
| const __m128i vp6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF); |
| const __m128i vp6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF); |
| const __m128i vp6xGHIJKLMNlo = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN); |
| const __m128i vp6xGHIJKLMNhi = _mm_mulhi_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp6x01234567lo, vp6x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp6x89ABCDEFlo, vp6x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp6x89ABCDEFlo, vp6x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp6xGHIJKLMNlo, vp6xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp6xGHIJKLMNlo, vp6xGHIJKLMNhi)); |
| |
| const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); |
| const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567); |
| const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t))); |
| const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567); |
| const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8)); |
| const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF); |
| const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t))); |
| const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF); |
| const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16)); |
| const __m128i vxi7xGHIJKLMN = _mm_cvtepi8_epi16(vi7xGHIJKLMN); |
| const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t))); |
| const __m128i vxk7xGHIJKLMN = _mm_cvtepi8_epi16(vk7xGHIJKLMN); |
| i7 += 24; |
| |
| |
| const __m128i vp7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); |
| const __m128i vp7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); |
| const __m128i vp7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF); |
| const __m128i vp7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF); |
| const __m128i vp7xGHIJKLMNlo = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN); |
| const __m128i vp7xGHIJKLMNhi = _mm_mulhi_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp7x01234567lo, vp7x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp7x89ABCDEFlo, vp7x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp7x89ABCDEFlo, vp7x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp7xGHIJKLMNlo, vp7xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp7xGHIJKLMNlo, vp7xGHIJKLMNhi)); |
| |
| const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); |
| const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567); |
| const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t))); |
| const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567); |
| const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8)); |
| const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF); |
| const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t))); |
| const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF); |
| const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16)); |
| const __m128i vxi8xGHIJKLMN = _mm_cvtepi8_epi16(vi8xGHIJKLMN); |
| const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t))); |
| const __m128i vxk8xGHIJKLMN = _mm_cvtepi8_epi16(vk8xGHIJKLMN); |
| i8 += 24; |
| |
| |
| const __m128i vp8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); |
| const __m128i vp8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); |
| const __m128i vp8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF); |
| const __m128i vp8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF); |
| const __m128i vp8xGHIJKLMNlo = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN); |
| const __m128i vp8xGHIJKLMNhi = _mm_mulhi_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp8x01234567lo, vp8x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi)); |
| vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp8x89ABCDEFlo, vp8x89ABCDEFhi)); |
| vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp8x89ABCDEFlo, vp8x89ABCDEFhi)); |
| vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vp8xGHIJKLMNlo, vp8xGHIJKLMNhi)); |
| vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vp8xGHIJKLMNlo, vp8xGHIJKLMNhi)); |
| |
| w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)); |
| |
| __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); |
| __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); |
| __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB); |
| __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF); |
| __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ); |
| __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN); |
| |
| const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale); |
| vscaled0123 = _mm_mul_ps(vscaled0123, vscale); |
| vscaled4567 = _mm_mul_ps(vscaled4567, vscale); |
| vscaled89AB = _mm_mul_ps(vscaled89AB, vscale); |
| vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale); |
| vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale); |
| vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale); |
| |
| vacc0123 = _mm_cvtps_epi32(vscaled0123); |
| vacc4567 = _mm_cvtps_epi32(vscaled4567); |
| vacc89AB = _mm_cvtps_epi32(vscaled89AB); |
| vaccCDEF = _mm_cvtps_epi32(vscaledCDEF); |
| vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ); |
| vaccKLMN = _mm_cvtps_epi32(vscaledKLMN); |
| |
| const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point); |
| __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point); |
| __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point); |
| |
| |
| __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF); |
| __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN); |
| |
| const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min); |
| vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min); |
| voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min); |
| |
| const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse4.output_max); |
| vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max); |
| voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max); |
| |
| _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN); |
| output += 24; |
| } |
| if XNN_UNLIKELY(c != 0) { |
| const int8_t* k = (const int8_t*) ((const int32_t*) w + 24); |
| do { |
| __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); |
| __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); |
| |
| |
| const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); |
| const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567); |
| const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k); |
| const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567); |
| i0 += 8; |
| |
| |
| const __m128i vp0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); |
| const __m128i vp0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp0x01234567lo, vp0x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi)); |
| |
| const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); |
| const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567); |
| const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24)); |
| const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567); |
| i1 += 8; |
| |
| |
| const __m128i vp1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); |
| const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); |
| |
| const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); |
| const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567); |
| const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48)); |
| const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567); |
| i2 += 8; |
| |
| |
| const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); |
| const __m128i vp2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); |
| |
| const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); |
| const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567); |
| const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72)); |
| const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567); |
| i3 += 8; |
| |
| |
| const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); |
| const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); |
| |
| const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); |
| const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567); |
| const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96)); |
| const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); |
| i4 += 8; |
| |
| |
| const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); |
| const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); |
| |
| const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); |
| const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567); |
| const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120)); |
| const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567); |
| i5 += 8; |
| |
| |
| const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); |
| const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp5x01234567lo, vp5x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi)); |
| |
| const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); |
| const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567); |
| const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144)); |
| const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567); |
| i6 += 8; |
| |
| |
| const __m128i vp6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); |
| const __m128i vp6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp6x01234567lo, vp6x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi)); |
| |
| const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); |
| const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567); |
| const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168)); |
| const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567); |
| i7 += 8; |
| |
| |
| const __m128i vp7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); |
| const __m128i vp7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp7x01234567lo, vp7x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi)); |
| |
| const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); |
| const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567); |
| const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192)); |
| const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567); |
| i8 += 8; |
| |
| |
| const __m128i vp8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); |
| const __m128i vp8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); |
| |
| vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp8x01234567lo, vp8x01234567hi)); |
| vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi)); |
| |
| k += 8; |
| |
| __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); |
| __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); |
| |
| const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale); |
| vscaled0123 = _mm_mul_ps(vscaled0123, vscale); |
| vscaled4567 = _mm_mul_ps(vscaled4567, vscale); |
| |
| vacc0123 = _mm_cvtps_epi32(vscaled0123); |
| vacc4567 = _mm_cvtps_epi32(vscaled4567); |
| |
| w = (const void*) ((const int32_t*) w + 8); |
| |
| const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point); |
| __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| |
| |
| __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); |
| |
| vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min)); |
| vout0123456701234567 = _mm_min_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_max)); |
| |
| if XNN_LIKELY(c >= 8) { |
| _mm_storel_epi64((__m128i*) output, vout0123456701234567); |
| output += 8; |
| c -= 8; |
| } else { |
| if (c & 4) { |
| *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); |
| vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); |
| output += 4; |
| } |
| if (c & 2) { |
| *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0); |
| vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); |
| output += 2; |
| } |
| if (c & 1) { |
| *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0); |
| output += 1; |
| } |
| c = 0; |
| } |
| } while (c != 0); |
| } |
| |
| output = (int8_t*) ((uintptr_t) output + output_increment); |
| } while (--output_width != 0); |
| } |