GCC Code Coverage Report

Directory: brotensor/
Coverage: low: ≥ 0% medium: ≥ 75.0% high: ≥ 90.0%
Coverage Exec / Excl / Total
Lines: 94.8% 220 / 0 / 232
Functions: 100.0% 24 / 0 / 24
Branches: 38.3% 262 / 0 / 684
src/api/native_tensor_int8.cpp
Line Branch Exec Source
1 // native_tensor_int8.cpp — bodies for the INT8 / GGUF-quant group, restored
2 // from the QuickJS binding (src/js/tensor_bindings_int8.cpp).
3 //
4 // Every op except quantizeInt8PerRowHost is GPU-only (see the header): on a
5 // CPU-only build the dispatcher throws "not implemented on CPU", the
6 // BROTENSOR_API_CATCH below records it, and the JS wrapper rethrows it as an
7 // Error. Nothing here has a CPU fallback.
8
9 #include "native_tensor_int8_decl.h"
10 #include "api_internal.h"
11 #include "native_register.h"
12
13 #include <cstdint>
14 #include <string>
15 #include <vector>
16
17 using namespace brotensor::api;
18
19 namespace {
20
21 // The last quantizeInt8PerRowHost result on this thread. The weights cross as
22 // the native's `i8[]` return, the scales through the companion native — one
23 // quantise pass, two typed arrays, as the old `{ weights, scales }` object.
24 // Both are copied out by the runtime (release == nullptr), so the vectors may
25 // be reused by the next call.
26 thread_local std::vector<int8_t> tl_quant_weights;
27 thread_local std::vector<float> tl_quant_scales;
28
29 40 void clearBuffer(bronze_native_buffer* out) {
30 40 out->data = nullptr;
31 40 out->length = 0;
32 40 out->release = nullptr;
33 40 out->ctx = nullptr;
34 40 }
35
36 using brotensor::Tensor;
37
38 44 Tensor* T(void* p) { return toTensor(p); }
39
40 // The device ops check weight dtypes and (out,in) shapes themselves; what
41 // they take on trust is the activation extent the caller's dims imply, the
42 // optional bias lengths and the raw `const float*` masks. Those are checked
43 // here.
44
45 // An optional per-output bias of `out` elements carried in a kDyn slot.
46 22 bool needBias(const char* L, const char* name, uint64_t bits, int64_t out) {
47 22 const Tensor* b = tensorFromValue(bits);
48
1/2
✓ Branch 0 taken 22 times.
✗ Branch 1 not taken.
 22 return !b || needElems(L, name, b, out);
49 }
50
51 } // namespace
52
53 extern "C" {
54
55 // ---- GGUF k-quant dequant --------------------------------------------------
56
57 2 void bro_tensor_dequantQ4kToFp16(void* W_q, void* W_fp16) {
58
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("dequantQ4kToFp16", {W_q, W_fp16})) return;
59 BROTENSOR_API_TRY
60
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
 2 brotensor::dequant_q4k_to_fp16(*toTensor(W_q), *toTensor(W_fp16));
61
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("dequantQ4kToFp16")
62 4 }
63
64 2 void bro_tensor_dequantQ6kToFp16(void* W_q, void* W_fp16) {
65
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("dequantQ6kToFp16", {W_q, W_fp16})) return;
66 BROTENSOR_API_TRY
67
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
 2 brotensor::dequant_q6k_to_fp16(*toTensor(W_q), *toTensor(W_fp16));
68
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("dequantQ6kToFp16")
69 4 }
70
71 2 void bro_tensor_dequantQ8_0ToFp16(void* W_q, void* W_fp16) {
72
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("dequantQ8_0ToFp16", {W_q, W_fp16})) return;
73 BROTENSOR_API_TRY
74
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
 2 brotensor::dequant_q8_0_to_fp16(*toTensor(W_q), *toTensor(W_fp16));
75
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("dequantQ8_0ToFp16")
76 4 }
77
78 // ---- GGUF k-quant weight-only linear ---------------------------------------
79
80 2 void bro_tensor_linearForwardQ4kFp16(void* W_q, uint64_t bias_bits, void* x, void* y) {
81
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("linearForwardQ4kFp16", {W_q, x, y})) return;
82 BROTENSOR_API_TRY
83
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
 2 brotensor::linear_forward_q4k_fp16(*toTensor(W_q), tensorFromValue(bias_bits), *toTensor(x), *toTensor(y));
84
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardQ4kFp16")
85 4 }
86
87 2 void bro_tensor_linearForwardQ6kFp16(void* W_q, uint64_t bias_bits, void* x, void* y) {
88
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("linearForwardQ6kFp16", {W_q, x, y})) return;
89 BROTENSOR_API_TRY
90
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
 2 brotensor::linear_forward_q6k_fp16(*toTensor(W_q), tensorFromValue(bias_bits), *toTensor(x), *toTensor(y));
91
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardQ6kFp16")
92 4 }
93
94 2 void bro_tensor_linearForwardQ8_0Fp16(void* W_q, uint64_t bias_bits, void* x, void* y) {
95
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("linearForwardQ8_0Fp16", {W_q, x, y})) return;
96 BROTENSOR_API_TRY
97
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
 2 brotensor::linear_forward_q8_0_fp16(*toTensor(W_q), tensorFromValue(bias_bits), *toTensor(x), *toTensor(y));
98
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardQ8_0Fp16")
99 4 }
100
101 2 void bro_tensor_linearForwardBatchedQ4kFp16(void* W_q, uint64_t bias_bits, void* X_BD, void* Y_BD) {
102
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("linearForwardBatchedQ4kFp16", {W_q, X_BD, Y_BD})) return;
103 BROTENSOR_API_TRY
104
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
 2 brotensor::linear_forward_batched_q4k_fp16(*toTensor(W_q), tensorFromValue(bias_bits), *toTensor(X_BD), *toTensor(Y_BD));
105
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardBatchedQ4kFp16")
106 4 }
107
108 2 void bro_tensor_linearForwardBatchedQ6kFp16(void* W_q, uint64_t bias_bits, void* X_BD, void* Y_BD) {
109
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("linearForwardBatchedQ6kFp16", {W_q, X_BD, Y_BD})) return;
110 BROTENSOR_API_TRY
111
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
 2 brotensor::linear_forward_batched_q6k_fp16(*toTensor(W_q), tensorFromValue(bias_bits), *toTensor(X_BD), *toTensor(Y_BD));
112
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardBatchedQ6kFp16")
113 4 }
114
115 2 void bro_tensor_linearForwardBatchedQ8_0Fp16(void* W_q, uint64_t bias_bits, void* X_BD, void* Y_BD) {
116
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("linearForwardBatchedQ8_0Fp16", {W_q, X_BD, Y_BD})) return;
117 BROTENSOR_API_TRY
118
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
 2 brotensor::linear_forward_batched_q8_0_fp16(*toTensor(W_q), tensorFromValue(bias_bits), *toTensor(X_BD), *toTensor(Y_BD));
119
5/12
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardBatchedQ8_0Fp16")
120 4 }
121
122 // ---- W8A16 host quantiser --------------------------------------------------
123
124 20 void bro_tensor_quantizeInt8PerRowHost(const uint16_t* W_fp16, uint32_t W_fp16_len,
125 int32_t out, int32_t in, bronze_native_buffer* weights_out) {
126
1/2
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
 20 if (!weights_out) return;
127 20 clearBuffer(weights_out);
128 20 tl_quant_weights.clear();
129 20 tl_quant_scales.clear();
130
1/2
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
 20 if (!W_fp16) {
131 ✗ setError("quantizeInt8PerRowHost: W_fp16 must be a Uint16Array (binary16 bit pattern)");
132 ✗ return;
133 }
134
2/4
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 20 times.
 20 if (out <= 0 || in <= 0) {
135 ✗ setError("quantizeInt8PerRowHost: out, in must be > 0");
136 ✗ return;
137 }
138 20 const size_t count = static_cast<size_t>(out) * static_cast<size_t>(in);
139
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
 20 if (static_cast<size_t>(W_fp16_len) < count) {
140 ✗ setError("quantizeInt8PerRowHost: W_fp16 view too small for (out, in)");
141 ✗ return;
142 }
143 BROTENSOR_API_TRY
144
1/2
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
 20 tl_quant_weights.resize(count);
145
1/2
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
 20 tl_quant_scales.resize(static_cast<size_t>(out));
146
2/4
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 20 times.
✗ Branch 3 not taken.
 40 brotensor::quantize_int8_per_row_host(W_fp16, out, in,
147 20 tl_quant_weights.data(), tl_quant_scales.data());
148 20 weights_out->data = tl_quant_weights.data();
149 20 weights_out->length = static_cast<uint32_t>(tl_quant_weights.size());
150
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 20 BROTENSOR_API_CATCH("quantizeInt8PerRowHost")
151 20 }
152
153 20 void bro_tensor_quantizeInt8PerRowHostScales(bronze_native_buffer* scales_out) {
154
1/2
✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
 20 if (!scales_out) return;
155 20 clearBuffer(scales_out);
156
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
 20 if (tl_quant_scales.empty()) return;
157 20 scales_out->data = tl_quant_scales.data();
158 20 scales_out->length = static_cast<uint32_t>(tl_quant_scales.size());
159 20 }
160
161 // ---- W8A16 dense -----------------------------------------------------------
162
163 2 void bro_tensor_matmulInt8wFp16(void* W_int8, void* scales, void* X, void* Y) {
164
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need("matmulInt8wFp16", {W_int8, scales, X, Y})) return;
165 BROTENSOR_API_TRY
166
5/10
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
 2 brotensor::matmul_int8w_fp16(*toTensor(W_int8), *toTensor(scales), *toTensor(X), *toTensor(Y));
167
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("matmulInt8wFp16")
168 2 }
169
170 2 void bro_tensor_linearForwardBatchedInt8wFp16(void* W_int8, void* scales, uint64_t bias_bits,
171 void* X_BD, void* Y_BD) {
172
2/4
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
 2 if (!need("linearForwardBatchedInt8wFp16", {W_int8, scales, X_BD, Y_BD}) ||
173 2 !needBias("linearForwardBatchedInt8wFp16", "bias", bias_bits, T(W_int8)->rows)) return;
174 BROTENSOR_API_TRY
175
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
 4 brotensor::linear_forward_batched_int8w_fp16(*toTensor(W_int8), *toTensor(scales),
176
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bias_bits),
177
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(X_BD), *toTensor(Y_BD));
178
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("linearForwardBatchedInt8wFp16")
179 2 }
180
181 // ---- W8A16 convolution -----------------------------------------------------
182
183 2 void bro_tensor_conv2dInt8wFp16Forward(void* X, void* W_int8, void* scales, uint64_t bias_bits,
184 int32_t N, int32_t C_in, int32_t H, int32_t Win,
185 int32_t C_out, int32_t kH, int32_t kW,
186 int32_t sH, int32_t sW, int32_t pH, int32_t pW,
187 int32_t dH, int32_t dW, int32_t groups, void* Y) {
188 2 const char* L = "conv2dInt8wFp16Forward";
189
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need(L, {X, W_int8, scales, Y})) return;
190 2 int64_t H_out = 0, W_out = 0;
191
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 if (!convGeom2d(L, C_in, C_out, H, Win, kH, kW, sH, sW, pH, pW, dH, dW, groups, H_out, W_out)) return;
192
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
 2 if (!needNonNegative(L, {N}) || !needElems(L, "X", T(X), elems({N, C_in, H, Win})) ||
193 2 !needBias(L, "bias", bias_bits, C_out)) return;
194 BROTENSOR_API_TRY
195
4/8
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
 4 brotensor::conv2d_int8w_fp16_forward(*toTensor(X), *toTensor(W_int8), *toTensor(scales),
196
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bias_bits),
197 2 N, C_in, H, Win, C_out, kH, kW,
198
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 sH, sW, pH, pW, dH, dW, groups, *toTensor(Y));
199
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("conv2dInt8wFp16Forward")
200 2 }
201
202 2 void bro_tensor_conv3dInt8wFp16Forward(void* X, void* W_int8, void* scales, uint64_t bias_bits,
203 int32_t N, int32_t C_in, int32_t T, int32_t H, int32_t Win,
204 int32_t C_out, int32_t kT, int32_t kH, int32_t kW,
205 int32_t sT, int32_t sH, int32_t sW,
206 int32_t pT, int32_t pH, int32_t pW,
207 int32_t dT, int32_t dH, int32_t dW,
208 int32_t groups, void* Y) {
209 2 const char* L = "conv3dInt8wFp16Forward";
210
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need(L, {X, W_int8, scales, Y})) return;
211 2 int64_t H_out = 0, W_out = 0;
212
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 if (!convGeom2d(L, C_in, C_out, H, Win, kH, kW, sH, sW, pH, pW, dH, dW, groups, H_out, W_out)) return;
213
2/4
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
 2 if (!needPositive(L, {kT, sT, dT}) || !needNonNegative(L, {N, T, pT})) return;
214
2/4
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
 2 if (!needElems(L, "X", toTensor(X), elems({N, C_in, T, H, Win})) || !needBias(L, "bias", bias_bits, C_out)) return;
215 BROTENSOR_API_TRY
216
4/8
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
 4 brotensor::conv3d_int8w_fp16_forward(*toTensor(X), *toTensor(W_int8), *toTensor(scales),
217
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bias_bits),
218 2 N, C_in, T, H, Win, C_out, kT, kH, kW,
219 2 sT, sH, sW, pT, pH, pW, dT, dH, dW,
220
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 groups, *toTensor(Y));
221
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("conv3dInt8wFp16Forward")
222 2 }
223
224 // ---- W8A16 diffusion resblock ----------------------------------------------
225
226 2 void bro_tensor_resblockForwardInt8wFp16(void* X, void* gamma1, void* beta1,
227 void* W1_int8, void* s1, uint64_t b1_bits,
228 uint64_t t_emb_shift_bits,
229 void* gamma2, void* beta2,
230 void* W2_int8, void* s2, uint64_t b2_bits,
231 uint64_t Wskip_bits, uint64_t sskip_bits, uint64_t bskip_bits,
232 int32_t N, int32_t C_in, int32_t C_out, int32_t H, int32_t Win,
233 int32_t numGroups, double eps, void* Y) {
234 2 const char* L = "resblockForwardInt8wFp16";
235
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need(L, {X, gamma1, beta1, W1_int8, s1, gamma2, beta2, W2_int8, s2, Y})) return;
236
2/4
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
 2 if (!needPositive(L, {numGroups}) || !needNonNegative(L, {N, C_in, C_out, H, Win})) return;
237
1/4
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
 2 if (C_in != C_out && !tensorFromValue(Wskip_bits)) {
238 ✗ setError(std::string(L) + ": C_in != C_out needs a Wskip projection");
239 ✗ return;
240 }
241 // The op checks dtypes, numGroups and W1/W2 shapes; the FP16 operands'
242 // extents and the optional tensors are checked here.
243
9/18
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✗ Branch 8 not taken.
✓ Branch 9 taken 2 times.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✗ Branch 12 not taken.
✓ Branch 13 taken 2 times.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
 18 if (!needOperands(L, T(X), {{"X", T(X), elems({N, C_in, H, Win})},
244 4 {"gamma1", T(gamma1), C_in}, {"beta1", T(beta1), C_in},
245 2 {"b1", tensorFromValue(b1_bits), C_out},
246 2 {"t_emb_shift", tensorFromValue(t_emb_shift_bits), C_out},
247 4 {"gamma2", T(gamma2), C_out}, {"beta2", T(beta2), C_out},
248 2 {"b2", tensorFromValue(b2_bits), C_out},
249 2 {"bskip", tensorFromValue(bskip_bits), C_out}})) return;
250
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
 2 if (!needMask(L, T(s1), C_out, "s1") || !needMask(L, T(s2), C_out, "s2")) return;
251
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 if (const Tensor* ws = tensorFromValue(Wskip_bits)) {
252 ✗ const Tensor* ss = tensorFromValue(sskip_bits);
253 ✗ if (!ss) { setError(std::string(L) + ": Wskip_int8 needs sskip scales"); return; }
254 ✗ if (!needElems(L, "Wskip_int8", ws, elems({C_out, C_in})) || !needMask(L, ss, C_out, "sskip")) return;
255 ✗ }
256 BROTENSOR_API_TRY
257
4/8
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
 4 brotensor::resblock_forward_int8w_fp16(*toTensor(X), *toTensor(gamma1), *toTensor(beta1),
258
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(W1_int8), *toTensor(s1),
259
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 tensorFromValue(b1_bits), tensorFromValue(t_emb_shift_bits),
260
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(gamma2), *toTensor(beta2),
261
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(W2_int8), *toTensor(s2),
262
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(b2_bits),
263
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 tensorFromValue(Wskip_bits), tensorFromValue(sskip_bits),
264
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bskip_bits),
265 2 N, C_in, C_out, H, Win, numGroups,
266
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 static_cast<float>(eps), *toTensor(Y));
267
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("resblockForwardInt8wFp16")
268 2 }
269
270 // ---- W8A16 flash-attention triplet -----------------------------------------
271
272 2 void bro_tensor_flashAttentionProjectKvInt8wFp16(void* ctx, void* Wk_int8, void* sk, uint64_t bk_bits,
273 void* Wv_int8, void* sv, uint64_t bv_bits,
274 void* K_out, void* V_out) {
275 2 const char* L = "flashAttentionProjectKvInt8wFp16";
276
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need(L, {ctx, Wk_int8, sk, Wv_int8, sv, K_out, V_out})) return;
277
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
 2 if (!needBias(L, "bk", bk_bits, T(Wk_int8)->rows) || !needBias(L, "bv", bv_bits, T(Wv_int8)->rows)) return;
278 BROTENSOR_API_TRY
279
4/8
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
 4 brotensor::flash_attention_project_kv_int8w_fp16(*toTensor(ctx), *toTensor(Wk_int8), *toTensor(sk),
280
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bk_bits),
281
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(Wv_int8), *toTensor(sv),
282
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bv_bits),
283
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(K_out), *toTensor(V_out));
284
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("flashAttentionProjectKvInt8wFp16")
285 2 }
286
287 2 void bro_tensor_flashAttentionQWithKvCachedInt8wFp16(void* X, void* K, void* V,
288 void* Wq_int8, void* sq, uint64_t bq_bits,
289 void* Wo_int8, void* so, uint64_t bo_bits,
290 uint64_t mask_bits, int32_t numHeads,
291 bool causal, void* O) {
292 2 const char* L = "flashAttentionQWithKvCachedInt8wFp16";
293
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need(L, {X, K, V, Wq_int8, sq, Wo_int8, so, O})) return;
294
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
 2 if (!needBias(L, "bq", bq_bits, T(Wq_int8)->rows) || !needBias(L, "bo", bo_bits, T(Wo_int8)->rows) ||
295 2 !needMask(L, tensorFromValue(mask_bits), T(K)->rows)) return;
296 BROTENSOR_API_TRY
297
4/8
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
 4 brotensor::flash_attention_q_with_kv_cached_int8w_fp16(*toTensor(X), *toTensor(K), *toTensor(V),
298
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(Wq_int8), *toTensor(sq),
299
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bq_bits),
300
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 *toTensor(Wo_int8), *toTensor(so),
301
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 tensorFromValue(bo_bits),
302
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 maskPtr(mask_bits), numHeads, causal,
303
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 *toTensor(O));
304
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("flashAttentionQWithKvCachedInt8wFp16")
305 2 }
306
307 2 void bro_tensor_flashAttentionQkvoInt8wFp16(void* X, uint64_t Ctx_bits,
308 void* Wq_int8, void* sq, uint64_t bq_bits,
309 void* Wk_int8, void* sk, uint64_t bk_bits,
310 void* Wv_int8, void* sv, uint64_t bv_bits,
311 void* Wo_int8, void* so, uint64_t bo_bits,
312 uint64_t mask_bits, int32_t numHeads,
313 bool causal, void* O) {
314 2 const char* L = "flashAttentionQkvoInt8wFp16";
315
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 if (!need(L, {X, Wq_int8, sq, Wk_int8, sk, Wv_int8, sv, Wo_int8, so, O})) return;
316
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
 2 const Tensor* kvSrc = tensorFromValue(Ctx_bits) ? tensorFromValue(Ctx_bits) : T(X);
317
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
 4 if (!needBias(L, "bq", bq_bits, T(Wq_int8)->rows) || !needBias(L, "bk", bk_bits, T(Wk_int8)->rows) ||
318
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 !needBias(L, "bv", bv_bits, T(Wv_int8)->rows) || !needBias(L, "bo", bo_bits, T(Wo_int8)->rows) ||
319 2 !needMask(L, tensorFromValue(mask_bits), kvSrc->rows)) return;
320 BROTENSOR_API_TRY
321
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
 4 brotensor::flash_attention_qkvo_int8w_fp16(*toTensor(X), tensorFromValue(Ctx_bits),
322
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
 2 *toTensor(Wq_int8), *toTensor(sq), tensorFromValue(bq_bits),
323
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
 2 *toTensor(Wk_int8), *toTensor(sk), tensorFromValue(bk_bits),
324
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
 2 *toTensor(Wv_int8), *toTensor(sv), tensorFromValue(bv_bits),
325
3/6
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
 2 *toTensor(Wo_int8), *toTensor(so), tensorFromValue(bo_bits),
326
2/4
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
 2 maskPtr(mask_bits), numHeads, causal, *toTensor(O));
327
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 2 BROTENSOR_API_CATCH("flashAttentionQkvoInt8wFp16")
328 2 }
329
330 // ---- W8A16 T5-style bias self-attention ------------------------------------
331
332 4 void bro_tensor_selfAttentionBiasInt8wFp16(void* X,
333 void* Wq_int8, void* sq,
334 void* Wk_int8, void* sk,
335 void* Wv_int8, void* sv,
336 void* Wo_int8, void* so,
337 uint64_t mask_bits, uint64_t attnBias_bits,
338 int32_t numHeads, double scale, void* O) {
339
2/4
✗ Branch 0 not taken.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 4 times.
 4 if (!need("selfAttentionBiasInt8wFp16", {X, Wq_int8, sq, Wk_int8, sk, Wv_int8, sv, Wo_int8, so, O}) ||
340 4 !needMask("selfAttentionBiasInt8wFp16", tensorFromValue(mask_bits), T(X)->rows)) return;
341 BROTENSOR_API_TRY
342
2/4
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 4 times.
✗ Branch 3 not taken.
 8 brotensor::self_attention_bias_int8w_fp16(*toTensor(X),
343
2/4
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 4 times.
✗ Branch 3 not taken.
 4 *toTensor(Wq_int8), *toTensor(sq),
344
2/4
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 4 times.
✗ Branch 3 not taken.
 4 *toTensor(Wk_int8), *toTensor(sk),
345
2/4
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 4 times.
✗ Branch 3 not taken.
 4 *toTensor(Wv_int8), *toTensor(sv),
346
2/4
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 4 times.
✗ Branch 3 not taken.
 4 *toTensor(Wo_int8), *toTensor(so),
347
2/4
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 4 times.
✗ Branch 3 not taken.
 4 maskPtr(mask_bits), tensorFromValue(attnBias_bits),
348
1/2
✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
 4 numHeads, static_cast<float>(scale), *toTensor(O));
349
0/12
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
 4 BROTENSOR_API_CATCH("selfAttentionBiasInt8wFp16")
350 4 }
351
352 } // extern "C"
353
354 namespace brotensor::api {
355
356 3 bool registerTensorNatives_int8(std::string* error) {
357 using namespace brotensor::api::reg;
358 3 bool ok =
359
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.dequantQ4kToFp16", p(&bro_tensor_dequantQ4kToFp16), "void", {kTensorCls, kTensorCls}, error) &&
360
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.dequantQ6kToFp16", p(&bro_tensor_dequantQ6kToFp16), "void", {kTensorCls, kTensorCls}, error) &&
361
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.dequantQ8_0ToFp16", p(&bro_tensor_dequantQ8_0ToFp16), "void", {kTensorCls, kTensorCls}, error) &&
362
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardQ4kFp16", p(&bro_tensor_linearForwardQ4kFp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
363
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardQ6kFp16", p(&bro_tensor_linearForwardQ6kFp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
364
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardQ8_0Fp16", p(&bro_tensor_linearForwardQ8_0Fp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
365
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardBatchedQ4kFp16", p(&bro_tensor_linearForwardBatchedQ4kFp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
366
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardBatchedQ6kFp16", p(&bro_tensor_linearForwardBatchedQ6kFp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
367
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardBatchedQ8_0Fp16", p(&bro_tensor_linearForwardBatchedQ8_0Fp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
368
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.quantizeInt8PerRowHost", p(&bro_tensor_quantizeInt8PerRowHost), "i8[]", {"u16[]", "i32", "i32"}, error) &&
369
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.quantizeInt8PerRowHostScales", p(&bro_tensor_quantizeInt8PerRowHostScales), "f32[]", {}, error) &&
370
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.matmulInt8wFp16", p(&bro_tensor_matmulInt8wFp16), "void", {kTensorCls, kTensorCls, kTensorCls, kTensorCls}, error) &&
371
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.linearForwardBatchedInt8wFp16", p(&bro_tensor_linearForwardBatchedInt8wFp16), "void", {kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
372
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.conv2dInt8wFp16Forward", p(&bro_tensor_conv2dInt8wFp16Forward), "void", {kTensorCls, kTensorCls, kTensorCls, kDyn, "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", kTensorCls}, error) &&
373
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.conv3dInt8wFp16Forward", p(&bro_tensor_conv3dInt8wFp16Forward), "void", {kTensorCls, kTensorCls, kTensorCls, kDyn, "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", "i32", kTensorCls}, error) &&
374
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.resblockForwardInt8wFp16", p(&bro_tensor_resblockForwardInt8wFp16), "void", {kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kDyn, kDyn, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kDyn, kDyn, kDyn, kDyn, "i32", "i32", "i32", "i32", "i32", "i32", "f64", kTensorCls}, error) &&
375
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.flashAttentionProjectKvInt8wFp16", p(&bro_tensor_flashAttentionProjectKvInt8wFp16), "void", {kTensorCls, kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls}, error) &&
376
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
 3 fn("__bro_native.tensor.flashAttentionQWithKvCachedInt8wFp16", p(&bro_tensor_flashAttentionQWithKvCachedInt8wFp16), "void", {kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls, kDyn, kDyn, "i32", "bool", kTensorCls}, error) &&
377
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
 3 fn("__bro_native.tensor.flashAttentionQkvoInt8wFp16", p(&bro_tensor_flashAttentionQkvoInt8wFp16), "void", {kTensorCls, kDyn, kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls, kDyn, kTensorCls, kTensorCls, kDyn, kDyn, "i32", "bool", kTensorCls}, error) &&
378 3 fn("__bro_native.tensor.selfAttentionBiasInt8wFp16", p(&bro_tensor_selfAttentionBiasInt8wFp16), "void", {kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kTensorCls, kDyn, kDyn, "i32", "f64", kTensorCls}, error);
379 3 return ok;
380 }
381
382 } // namespace brotensor::api
383