Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-gemm/avx512-broadcast.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <immintrin.h> |
| 13 | |
| 14 | #include <xnnpack/gemm.h> |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 15 | #include <xnnpack/intrinsics-polyfill.h> |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 16 | |
| 17 | |
| 18 | void xnn_f32_gemm_ukernel_8x16__avx512f_broadcast( |
| 19 | size_t mr, |
| 20 | size_t nc, |
| 21 | size_t kc, |
| 22 | const float*restrict a, |
| 23 | size_t a_stride, |
| 24 | const float*restrict w, |
| 25 | float*restrict c, |
| 26 | size_t cm_stride, |
| 27 | size_t cn_stride, |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 28 | const union xnn_f32_minmax_params params[restrict static 1]) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 29 | { |
| 30 | assert(mr != 0); |
| 31 | assert(mr <= 8); |
| 32 | assert(nc != 0); |
| 33 | assert(kc != 0); |
| 34 | assert(kc % sizeof(float) == 0); |
| 35 | assert(a != NULL); |
| 36 | assert(w != NULL); |
| 37 | assert(c != NULL); |
| 38 | |
| 39 | const float* a0 = a; |
| 40 | float* c0 = c; |
| 41 | const float* a1 = (const float*) ((uintptr_t) a0 + a_stride); |
| 42 | float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 43 | if XNN_UNPREDICTABLE(mr < 2) { |
| 44 | a1 = a0; |
| 45 | c1 = c0; |
| 46 | } |
| 47 | const float* a2 = (const float*) ((uintptr_t) a1 + a_stride); |
| 48 | float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 49 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 50 | a2 = a1; |
| 51 | c2 = c1; |
| 52 | } |
| 53 | const float* a3 = (const float*) ((uintptr_t) a2 + a_stride); |
| 54 | float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 55 | if XNN_UNPREDICTABLE(mr < 4) { |
| 56 | a3 = a2; |
| 57 | c3 = c2; |
| 58 | } |
| 59 | const float* a4 = (const float*) ((uintptr_t) a3 + a_stride); |
| 60 | float* c4 = (float*) ((uintptr_t) c3 + cm_stride); |
| 61 | if XNN_UNPREDICTABLE(mr <= 4) { |
| 62 | a4 = a3; |
| 63 | c4 = c3; |
| 64 | } |
| 65 | const float* a5 = (const float*) ((uintptr_t) a4 + a_stride); |
| 66 | float* c5 = (float*) ((uintptr_t) c4 + cm_stride); |
| 67 | if XNN_UNPREDICTABLE(mr < 6) { |
| 68 | a5 = a4; |
| 69 | c5 = c4; |
| 70 | } |
| 71 | const float* a6 = (const float*) ((uintptr_t) a5 + a_stride); |
| 72 | float* c6 = (float*) ((uintptr_t) c5 + cm_stride); |
| 73 | if XNN_UNPREDICTABLE(mr <= 6) { |
| 74 | a6 = a5; |
| 75 | c6 = c5; |
| 76 | } |
| 77 | const float* a7 = (const float*) ((uintptr_t) a6 + a_stride); |
| 78 | float* c7 = (float*) ((uintptr_t) c6 + cm_stride); |
| 79 | if XNN_UNPREDICTABLE(mr != 8) { |
| 80 | a7 = a6; |
| 81 | c7 = c6; |
| 82 | } |
| 83 | |
| 84 | do { |
| 85 | __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); |
| 86 | __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 87 | __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 88 | __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 89 | __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 90 | __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 91 | __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 92 | __m512 vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF; |
| 93 | w += 16; |
| 94 | |
| 95 | size_t k = kc; |
| 96 | do { |
| 97 | const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); |
| 98 | w += 16; |
| 99 | |
| 100 | vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789ABCDEF); |
| 101 | vacc1x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a1), vb0123456789ABCDEF, vacc1x0123456789ABCDEF); |
| 102 | vacc2x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a2), vb0123456789ABCDEF, vacc2x0123456789ABCDEF); |
| 103 | vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789ABCDEF); |
| 104 | vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789ABCDEF); |
| 105 | vacc5x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a5), vb0123456789ABCDEF, vacc5x0123456789ABCDEF); |
| 106 | vacc6x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a6), vb0123456789ABCDEF, vacc6x0123456789ABCDEF); |
| 107 | vacc7x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a7), vb0123456789ABCDEF, vacc7x0123456789ABCDEF); |
| 108 | |
| 109 | a0 += 1; |
| 110 | a1 += 1; |
| 111 | a2 += 1; |
| 112 | a3 += 1; |
| 113 | a4 += 1; |
| 114 | a5 += 1; |
| 115 | a6 += 1; |
| 116 | a7 += 1; |
| 117 | |
| 118 | k -= sizeof(float); |
| 119 | } while (k != 0); |
| 120 | |
| 121 | const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max)); |
| 122 | vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); |
| 123 | vacc1x0123456789ABCDEF = _mm512_min_ps(vacc1x0123456789ABCDEF, vmax); |
| 124 | vacc2x0123456789ABCDEF = _mm512_min_ps(vacc2x0123456789ABCDEF, vmax); |
| 125 | vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); |
| 126 | vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); |
| 127 | vacc5x0123456789ABCDEF = _mm512_min_ps(vacc5x0123456789ABCDEF, vmax); |
| 128 | vacc6x0123456789ABCDEF = _mm512_min_ps(vacc6x0123456789ABCDEF, vmax); |
| 129 | vacc7x0123456789ABCDEF = _mm512_min_ps(vacc7x0123456789ABCDEF, vmax); |
| 130 | |
| 131 | const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min)); |
| 132 | vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); |
| 133 | vacc1x0123456789ABCDEF = _mm512_max_ps(vacc1x0123456789ABCDEF, vmin); |
| 134 | vacc2x0123456789ABCDEF = _mm512_max_ps(vacc2x0123456789ABCDEF, vmin); |
| 135 | vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); |
| 136 | vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); |
| 137 | vacc5x0123456789ABCDEF = _mm512_max_ps(vacc5x0123456789ABCDEF, vmin); |
| 138 | vacc6x0123456789ABCDEF = _mm512_max_ps(vacc6x0123456789ABCDEF, vmin); |
| 139 | vacc7x0123456789ABCDEF = _mm512_max_ps(vacc7x0123456789ABCDEF, vmin); |
| 140 | |
| 141 | if XNN_LIKELY(nc >= 16) { |
| 142 | _mm512_storeu_ps(c7, vacc7x0123456789ABCDEF); |
| 143 | c7 = (float*) ((uintptr_t) c7 + cn_stride); |
| 144 | _mm512_storeu_ps(c6, vacc6x0123456789ABCDEF); |
| 145 | c6 = (float*) ((uintptr_t) c6 + cn_stride); |
| 146 | _mm512_storeu_ps(c5, vacc5x0123456789ABCDEF); |
| 147 | c5 = (float*) ((uintptr_t) c5 + cn_stride); |
| 148 | _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); |
| 149 | c4 = (float*) ((uintptr_t) c4 + cn_stride); |
| 150 | _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); |
| 151 | c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 152 | _mm512_storeu_ps(c2, vacc2x0123456789ABCDEF); |
| 153 | c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 154 | _mm512_storeu_ps(c1, vacc1x0123456789ABCDEF); |
| 155 | c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 156 | _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); |
| 157 | c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 158 | |
| 159 | a7 = (const float*) ((uintptr_t) a7 - kc); |
| 160 | a6 = (const float*) ((uintptr_t) a6 - kc); |
| 161 | a5 = (const float*) ((uintptr_t) a5 - kc); |
| 162 | a4 = (const float*) ((uintptr_t) a4 - kc); |
| 163 | a3 = (const float*) ((uintptr_t) a3 - kc); |
| 164 | a2 = (const float*) ((uintptr_t) a2 - kc); |
| 165 | a1 = (const float*) ((uintptr_t) a1 - kc); |
| 166 | a0 = (const float*) ((uintptr_t) a0 - kc); |
| 167 | |
| 168 | nc -= 16; |
| 169 | } else { |
| 170 | if (nc & 15) { |
| 171 | // Prepare mask for valid 32-bit elements (depends on nc). |
| 172 | const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << nc) - UINT32_C(1))); |
| 173 | |
| 174 | _mm512_mask_storeu_ps(c7, vmask, vacc7x0123456789ABCDEF); |
| 175 | _mm512_mask_storeu_ps(c6, vmask, vacc6x0123456789ABCDEF); |
| 176 | _mm512_mask_storeu_ps(c5, vmask, vacc5x0123456789ABCDEF); |
| 177 | _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); |
| 178 | _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); |
| 179 | _mm512_mask_storeu_ps(c2, vmask, vacc2x0123456789ABCDEF); |
| 180 | _mm512_mask_storeu_ps(c1, vmask, vacc1x0123456789ABCDEF); |
| 181 | _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); |
| 182 | } |
| 183 | |
| 184 | nc = 0; |
| 185 | } |
| 186 | } while (nc != 0); |
| 187 | } |