Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-gemm/MRx4c8-wasmsimd-dot16x2.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2021 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 <wasm_simd128.h> |
| 13 | |
| 14 | #include <xnnpack/gemm.h> |
| 15 | #include <xnnpack/math.h> |
| 16 | |
| 17 | |
| 18 | void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64( |
| 19 | size_t mr, |
| 20 | size_t nc, |
| 21 | size_t kc, |
| 22 | const int8_t* restrict a, |
| 23 | size_t a_stride, |
| 24 | const void* restrict w, |
| 25 | int8_t* restrict c, |
| 26 | size_t cm_stride, |
| 27 | size_t cn_stride, |
| 28 | const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN |
| 29 | { |
| 30 | assert(mr != 0); |
| 31 | assert(mr <= 3); |
| 32 | assert(nc != 0); |
| 33 | assert(kc != 0); |
| 34 | assert(kc % sizeof(int8_t) == 0); |
| 35 | assert(a != NULL); |
| 36 | assert(w != NULL); |
| 37 | assert(c != NULL); |
| 38 | |
| 39 | kc = round_up_po2(kc, 8); |
| 40 | const int8_t* a0 = a; |
| 41 | int8_t* c0 = c; |
| 42 | const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 43 | int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 44 | if XNN_UNPREDICTABLE(mr < 2) { |
| 45 | a1 = a0; |
| 46 | c1 = c0; |
| 47 | } |
| 48 | const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 49 | int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 50 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 51 | a2 = a1; |
| 52 | c2 = c1; |
| 53 | } |
| 54 | |
| 55 | do { |
| 56 | v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]); |
| 57 | v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]); |
| 58 | v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]); |
| 59 | v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]); |
| 60 | v128_t vacc1x0 = vacc0x0; |
| 61 | v128_t vacc1x1 = vacc0x1; |
| 62 | v128_t vacc1x2 = vacc0x2; |
| 63 | v128_t vacc1x3 = vacc0x3; |
| 64 | v128_t vacc2x0 = vacc0x0; |
| 65 | v128_t vacc2x1 = vacc0x1; |
| 66 | v128_t vacc2x2 = vacc0x2; |
| 67 | v128_t vacc2x3 = vacc0x3; |
| 68 | w = (const void*) ((const int32_t*) w + 4); |
| 69 | |
| 70 | size_t k = 0; |
| 71 | while (k < kc) { |
| 72 | const v128_t vxa0 = wasm_i16x8_load8x8(a0); |
| 73 | a0 += 8; |
| 74 | const v128_t vxa1 = wasm_i16x8_load8x8(a1); |
| 75 | a1 += 8; |
| 76 | const v128_t vxa2 = wasm_i16x8_load8x8(a2); |
| 77 | a2 += 8; |
| 78 | |
| 79 | const v128_t vxb0 = wasm_i16x8_load8x8(w); |
| 80 | |
| 81 | vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0)); |
| 82 | vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0)); |
| 83 | vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0)); |
| 84 | const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8); |
| 85 | |
| 86 | vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1)); |
| 87 | vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1)); |
| 88 | vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1)); |
| 89 | const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16); |
| 90 | |
| 91 | vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2)); |
| 92 | vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2)); |
| 93 | vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2)); |
| 94 | const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24); |
| 95 | |
| 96 | vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3)); |
| 97 | vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3)); |
| 98 | vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3)); |
| 99 | |
| 100 | w = (const void*) ((const int8_t*) w + 32); |
| 101 | k += 8 * sizeof(int8_t); |
| 102 | } |
| 103 | |
| 104 | const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7)); |
| 105 | const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7)); |
| 106 | const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7)); |
| 107 | const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7)); |
| 108 | const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7)); |
| 109 | const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7)); |
| 110 | |
| 111 | v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7)); |
| 112 | v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7)); |
| 113 | v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7)); |
| 114 | |
| 115 | vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123); |
| 116 | vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123); |
| 117 | vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123); |
| 118 | |
| 119 | const v128_t vscale0123 = wasm_v128_load(w); |
| 120 | w = (const void*) ((const float*) w + 4); |
| 121 | vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123); |
| 122 | vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123); |
| 123 | vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123); |
| 124 | |
| 125 | const v128_t voutput_min_less_zero_point = wasm_v128_load(params->wasmsimd.output_min_less_zero_point); |
| 126 | vacc0x0123 = wasm_f32x4_max(voutput_min_less_zero_point, vacc0x0123); |
| 127 | vacc1x0123 = wasm_f32x4_max(voutput_min_less_zero_point, vacc1x0123); |
| 128 | vacc2x0123 = wasm_f32x4_max(voutput_min_less_zero_point, vacc2x0123); |
| 129 | |
| 130 | const v128_t voutput_max_less_zero_point = wasm_v128_load(params->wasmsimd.output_max_less_zero_point); |
| 131 | vacc0x0123 = wasm_f32x4_min(voutput_max_less_zero_point, vacc0x0123); |
| 132 | vacc1x0123 = wasm_f32x4_min(voutput_max_less_zero_point, vacc1x0123); |
| 133 | vacc2x0123 = wasm_f32x4_min(voutput_max_less_zero_point, vacc2x0123); |
| 134 | |
| 135 | const v128_t vmagic_bias = wasm_v128_load(params->wasmsimd.magic_bias); |
| 136 | vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias); |
| 137 | vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias); |
| 138 | vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias); |
| 139 | |
| 140 | const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load(params->wasmsimd.magic_bias_less_output_zero_point); |
| 141 | vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point); |
| 142 | vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point); |
| 143 | vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point); |
| 144 | |
| 145 | v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123); |
| 146 | v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123); |
| 147 | |
| 148 | v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123); |
| 149 | |
| 150 | if (nc >= 4) { |
| 151 | *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0); |
| 152 | *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1); |
| 153 | *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2); |
| 154 | |
| 155 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 156 | c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 157 | c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 158 | |
| 159 | a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 160 | a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 161 | a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 162 | |
| 163 | nc -= 4; |
| 164 | } else { |
| 165 | uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0); |
| 166 | uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1); |
| 167 | uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2); |
| 168 | if (nc & 2) { |
| 169 | *((uint16_t*) c0) = (uint16_t) vout0; |
| 170 | vout0 >>= 16; |
| 171 | c0 += 2; |
| 172 | *((uint16_t*) c1) = (uint16_t) vout1; |
| 173 | vout1 >>= 16; |
| 174 | c1 += 2; |
| 175 | *((uint16_t*) c2) = (uint16_t) vout2; |
| 176 | vout2 >>= 16; |
| 177 | c2 += 2; |
| 178 | } |
| 179 | if (nc & 1) { |
| 180 | *c0 = (int8_t) vout0; |
| 181 | *c1 = (int8_t) vout1; |
| 182 | *c2 = (int8_t) vout2; |
| 183 | } |
| 184 | |
| 185 | nc = 0; |
| 186 | } |
| 187 | } while (nc != 0); |
| 188 | } |