提交 09820843 编写于 作者: C chenjiaoAngel

fix gemm_s8

上级 6373783d
......@@ -1367,7 +1367,7 @@ inline void gemm_sdot_int8_kernel(const int8_t* a_ptr,
"fcmge v4.4s, v30.4s, v2.4s \n" /* vcgeq_f32 */ \
"fmul v5.4s, v30.4s, v3.4s \n" /* vmulq_f32 */ \
"fcmge v6.4s, v31.4s, v2.4s \n" /* vcgeq_f32 */ \
"fmul v7.4s, v32.4s, v3.4s \n" /* vmulq_f32 */ \
"fmul v7.4s, v31.4s, v3.4s \n" /* vmulq_f32 */ \
"bif v30.16b, v5.16b, v4.16b \n" /* choose*/ \
"bif v31.16b, v7.16b, v6.16b \n" /* choose*/ \
"12: \n"
......
......@@ -30,8 +30,8 @@ void gemm_s8(bool is_transA,
Dtype* C,
const float* bias,
bool is_bias,
bool is_relu,
const float* scale,
const operators::ActivationParam act_param,
ARMContext* ctx) {
int hblock = get_hblock_int8(ctx);
int m_roundup = hblock * ((M + hblock - 1) / hblock);
......@@ -42,7 +42,7 @@ void gemm_s8(bool is_transA,
prepackA_int8(packed_A, A, lda, 0, M, 0, K, is_transA, ctx);
gemm_prepack_int8(
packed_A, B, bias, C, M, N, K, is_bias, is_relu, is_transB, scale, ctx);
packed_A, B, bias, C, M, N, K, is_bias, is_transB, scale, act_param, ctx);
TargetFree(TargetType::kARM, packed_A);
}
......@@ -56,8 +56,8 @@ template void gemm_s8<float>(bool is_transA,
float* C,
const float* bias,
bool is_bias,
bool is_relu,
const float* scale,
const operators::ActivationParam act_param,
ARMContext* ctx);
template void gemm_s8<int8_t>(bool is_transA,
......@@ -70,8 +70,8 @@ template void gemm_s8<int8_t>(bool is_transA,
int8_t* C,
const float* bias,
bool is_bias,
bool is_relu,
const float* scale,
const operators::ActivationParam act_param,
ARMContext* ctx);
} // namespace math
......
......@@ -34,8 +34,8 @@ void gemm_s8(bool is_transA,
Dtype* C,
const float* bias,
bool is_bias,
bool is_relu,
const float* scale,
const operators::ActivationParam act_param,
ARMContext* ctx);
} // namespace math
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册