Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 2 | // Template: src/qs8-igemm/MRx4c8-wasmsimd-mul16.c.in |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 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 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 18 | void xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld128( |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 19 | size_t mr, |
| 20 | size_t nc, |
| 21 | size_t kc, |
| 22 | size_t ks, |
| 23 | const int8_t** restrict a, |
| 24 | const void* restrict w, |
| 25 | int8_t* restrict c, |
| 26 | size_t cm_stride, |
| 27 | size_t cn_stride, |
| 28 | size_t a_offset, |
| 29 | const int8_t* zero, |
| 30 | const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN |
| 31 | { |
| 32 | assert(mr != 0); |
| 33 | assert(mr <= 1); |
| 34 | assert(nc != 0); |
| 35 | assert(kc != 0); |
| 36 | assert(ks != 0); |
| 37 | assert(ks % (1 * sizeof(void*)) == 0); |
| 38 | assert(a_offset % sizeof(int8_t) == 0); |
| 39 | assert(a != NULL); |
| 40 | assert(w != NULL); |
| 41 | assert(c != NULL); |
| 42 | |
| 43 | kc = round_up_po2(kc, 8); |
| 44 | int8_t* c0 = c; |
| 45 | |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 46 | do { |
Marat Dukhan | 4810905 | 2021-08-31 17:31:57 -0700 | [diff] [blame] | 47 | v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]); |
| 48 | v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]); |
| 49 | v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]); |
| 50 | v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]); |
Marat Dukhan | 43bee05 | 2021-07-14 20:57:18 -0700 | [diff] [blame] | 51 | w = (const void*) ((const int32_t*) w + 4); |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 52 | |
| 53 | size_t p = ks; |
| 54 | do { |
| 55 | const int8_t* restrict a0 = a[0]; |
| 56 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 57 | a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 58 | } |
| 59 | a += 1; |
| 60 | |
| 61 | size_t k = 0; |
| 62 | while (k < kc) { |
| 63 | const v128_t vxa0 = wasm_i16x8_load8x8(a0); |
| 64 | a0 += 8; |
| 65 | |
| 66 | const v128_t vb01 = wasm_v128_load(w); |
| 67 | const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01); |
| 68 | const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01); |
| 69 | |
| 70 | const v128_t vprod0x0 = wasm_i16x8_mul(vxb0, vxa0); |
| 71 | vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_low_i16x8(vprod0x0)); |
| 72 | |
| 73 | const v128_t vprod0x1 = wasm_i16x8_mul(vxb1, vxa0); |
| 74 | vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_low_i16x8(vprod0x1)); |
| 75 | vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_high_i16x8(vprod0x0)); |
| 76 | |
| 77 | vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_high_i16x8(vprod0x1)); |
Marat Dukhan | 43bee05 | 2021-07-14 20:57:18 -0700 | [diff] [blame] | 78 | const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16); |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 79 | const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23); |
| 80 | const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23); |
| 81 | |
| 82 | const v128_t vprod0x2 = wasm_i16x8_mul(vxb2, vxa0); |
| 83 | vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_low_i16x8(vprod0x2)); |
| 84 | |
| 85 | const v128_t vprod0x3 = wasm_i16x8_mul(vxb3, vxa0); |
| 86 | vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_low_i16x8(vprod0x3)); |
| 87 | vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_high_i16x8(vprod0x2)); |
| 88 | |
| 89 | vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_high_i16x8(vprod0x3)); |
| 90 | |
Marat Dukhan | 43bee05 | 2021-07-14 20:57:18 -0700 | [diff] [blame] | 91 | w = (const void*) ((const int8_t*) w + 32); |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 92 | k += 8 * sizeof(int8_t); |
| 93 | } |
| 94 | p -= 1 * sizeof(void*); |
| 95 | } while (p != 0); |
| 96 | |
| 97 | 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)); |
| 98 | 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)); |
| 99 | |
| 100 | v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7)); |
| 101 | |
| 102 | vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123); |
| 103 | |
| 104 | const v128_t vscale0123 = wasm_v128_load(w); |
| 105 | w = (const void*) ((const float*) w + 4); |
| 106 | vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123); |
| 107 | |
| 108 | const v128_t voutput_min_less_zero_point = wasm_v128_load(params->wasmsimd.output_min_less_zero_point); |
| 109 | vacc0x0123 = wasm_f32x4_max(vacc0x0123, voutput_min_less_zero_point); |
| 110 | |
| 111 | const v128_t voutput_max_less_zero_point = wasm_v128_load(params->wasmsimd.output_max_less_zero_point); |
| 112 | vacc0x0123 = wasm_f32x4_min(vacc0x0123, voutput_max_less_zero_point); |
| 113 | |
| 114 | const v128_t vmagic_bias = wasm_v128_load(params->wasmsimd.magic_bias); |
| 115 | vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias); |
| 116 | |
| 117 | const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load(params->wasmsimd.magic_bias_less_output_zero_point); |
| 118 | vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point); |
| 119 | |
Marat Dukhan | 07706f6 | 2021-08-08 23:48:40 -0700 | [diff] [blame] | 120 | v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123); |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 121 | |
Marat Dukhan | 07706f6 | 2021-08-08 23:48:40 -0700 | [diff] [blame] | 122 | v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123); |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 123 | |
| 124 | if (nc >= 4) { |
| 125 | *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0); |
| 126 | |
| 127 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 128 | |
| 129 | a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 130 | |
| 131 | nc -= 4; |
| 132 | } else { |
Marat Dukhan | 7a8dd87 | 2021-08-19 02:39:13 -0700 | [diff] [blame] | 133 | uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0); |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 134 | if (nc & 2) { |
Marat Dukhan | 7a8dd87 | 2021-08-19 02:39:13 -0700 | [diff] [blame] | 135 | *((uint16_t*) c0) = (uint16_t) vout0; |
| 136 | vout0 >>= 16; |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 137 | c0 += 2; |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 138 | } |
| 139 | if (nc & 1) { |
Marat Dukhan | 7a8dd87 | 2021-08-19 02:39:13 -0700 | [diff] [blame] | 140 | *c0 = (int8_t) vout0; |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 141 | } |
| 142 | |
| 143 | nc = 0; |
| 144 | } |
| 145 | } while (nc != 0); |
| 146 | } |