Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-igemm/MRx4c8-minmax-wasmsimd.c.in |
| 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 | |
| 16 | |
| 17 | void xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64( |
| 18 | size_t mr, |
| 19 | size_t nc, |
| 20 | size_t kc, |
| 21 | size_t ks, |
| 22 | const int8_t** restrict a, |
| 23 | const void* restrict w, |
| 24 | int8_t* restrict c, |
| 25 | size_t cm_stride, |
| 26 | size_t cn_stride, |
| 27 | size_t a_offset, |
| 28 | const int8_t* zero, |
| 29 | const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| 30 | { |
| 31 | assert(mr != 0); |
| 32 | assert(mr <= 1); |
| 33 | assert(nc != 0); |
| 34 | assert(kc != 0); |
| 35 | assert(ks != 0); |
| 36 | assert(ks % (1 * sizeof(void*)) == 0); |
| 37 | assert(a_offset % sizeof(int8_t) == 0); |
| 38 | assert(a != NULL); |
| 39 | assert(w != NULL); |
| 40 | assert(c != NULL); |
| 41 | |
| 42 | int8_t* c0 = c; |
| 43 | |
| 44 | const v128_t vzero = wasm_f64x2_splat(0.0); |
| 45 | do { |
| 46 | v128_t vacc0x0 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[0]); |
| 47 | v128_t vacc0x1 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[1]); |
| 48 | v128_t vacc0x2 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[2]); |
| 49 | v128_t vacc0x3 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[3]); |
| 50 | w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| 51 | |
| 52 | size_t p = ks; |
| 53 | do { |
| 54 | const int8_t* restrict a0 = a[0]; |
| 55 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 56 | a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 57 | } |
| 58 | a += 1; |
| 59 | |
| 60 | size_t k = 0; |
| 61 | while (k < kc) { |
| 62 | const v128_t vxa0 = wasm_i16x8_load_8x8(a0); |
| 63 | a0 += 8; |
| 64 | |
| 65 | const v128_t vxb0 = wasm_i16x8_load_8x8(w); |
| 66 | |
| 67 | const v128_t vprod0x0 = wasm_i16x8_mul(vxa0, vxb0); |
| 68 | vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_low_i16x8(vprod0x0)); |
| 69 | vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_high_i16x8(vprod0x0)); |
| 70 | const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); |
| 71 | |
| 72 | const v128_t vprod0x1 = wasm_i16x8_mul(vxa0, vxb1); |
| 73 | vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_low_i16x8(vprod0x1)); |
| 74 | vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_high_i16x8(vprod0x1)); |
| 75 | const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); |
| 76 | |
| 77 | const v128_t vprod0x2 = wasm_i16x8_mul(vxa0, vxb2); |
| 78 | vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_low_i16x8(vprod0x2)); |
| 79 | vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_high_i16x8(vprod0x2)); |
| 80 | const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); |
| 81 | |
| 82 | const v128_t vprod0x3 = wasm_i16x8_mul(vxa0, vxb3); |
| 83 | vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_low_i16x8(vprod0x3)); |
| 84 | vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_high_i16x8(vprod0x3)); |
| 85 | |
| 86 | w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t)); |
| 87 | k += 8 * sizeof(int8_t); |
| 88 | } |
| 89 | p -= 1 * sizeof(void*); |
| 90 | } while (p != 0); |
| 91 | |
| 92 | 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)); |
| 93 | 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)); |
| 94 | |
| 95 | v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7)); |
| 96 | |
| 97 | const v128_t vsign0x0123 = wasm_i32x4_lt(vacc0x0123, vzero); |
| 98 | |
| 99 | const v128_t vacc0x01 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 0, 4, 1, 5); |
| 100 | |
| 101 | const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier); |
| 102 | const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding); |
| 103 | const v128_t vprod0x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x01, vmultiplier), vrounding); |
| 104 | const v128_t vacc0x23 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 2, 6, 3, 7); |
| 105 | |
| 106 | const v128_t vprod0x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x23, vmultiplier), vrounding); |
| 107 | |
| 108 | const v128_t vq31prod0x0123 = wasm_v32x4_shuffle(vprod0x01, vprod0x23, 1, 3, 5, 7); |
| 109 | |
| 110 | const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask); |
| 111 | const v128_t vrem0x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0x0123, vremainder_mask), wasm_i32x4_lt(vq31prod0x0123, vzero)); |
| 112 | |
| 113 | const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold); |
| 114 | const int32_t vshift = params->wasmsimd.shift; |
| 115 | vacc0x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0x0123, vshift), wasm_i32x4_gt(vrem0x0123, vthreshold)); |
| 116 | |
| 117 | const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point); |
| 118 | v128_t vacc00x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123), voutput_zero_point); |
| 119 | |
| 120 | v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123); |
| 121 | |
| 122 | const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min); |
| 123 | vout = wasm_i8x16_max(vout, voutput_min); |
| 124 | |
| 125 | const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max); |
| 126 | vout = wasm_i8x16_min(vout, voutput_max); |
| 127 | |
| 128 | if (nc >= 4) { |
| 129 | *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0); |
| 130 | |
| 131 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 132 | |
| 133 | a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 134 | |
| 135 | nc -= 4; |
| 136 | } else { |
| 137 | if (nc & 2) { |
| 138 | *((uint16_t*) c0) = (uint16_t) wasm_i16x8_extract_lane(vout, 0); |
| 139 | c0 += 2; |
| 140 | vout = wasm_u32x4_shr(vout, 16); |
| 141 | } |
| 142 | if (nc & 1) { |
| 143 | *c0 = (int8_t) wasm_i8x16_extract_lane(vout, 0); |
| 144 | } |
| 145 | |
| 146 | nc = 0; |
| 147 | } |
| 148 | } while (nc != 0); |
| 149 | } |