[CPU][RISC-V] Enable oneDNN W8A8 INT8 to run on RISC-V (#44478)

Signed-off-by: velonica0 <like@mail.nankai.edu.cn>
This commit is contained in:
velonica0
2026-06-11 04:09:05 +00:00
committed by GitHub
parent 248e33c40d
commit f31bc2ea60
4 changed files with 42 additions and 5 deletions
+4
View File
@@ -57,6 +57,10 @@ typedef RVVTYPE(vfloat32, LMUL_512, _t) fixed_fp32x16_t
typedef RVVTYPE(vfloat32, LMUL_1024, _t) fixed_fp32x32_t
__attribute__((riscv_rvv_vector_bits(1024)));
// int8
typedef RVVTYPE(vint8, LMUL_128, _t) fixed_i8x16_t
__attribute__((riscv_rvv_vector_bits(128)));
// int32
typedef RVVTYPE(vint32, LMUL_256, _t) fixed_i32x8_t
__attribute__((riscv_rvv_vector_bits(256)));
+29
View File
@@ -734,10 +734,18 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
return FP32Vec16(
RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
}
FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
return FP32Vec16(
RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, elem_num));
}
FP32Vec16 min(const FP32Vec16& b) const {
return FP32Vec16(
RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
}
FP32Vec16 min(const FP32Vec16& b, const int elem_num) const {
return FP32Vec16(
RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, elem_num));
}
FP32Vec16 abs() const {
return FP32Vec16(RVVI(__riscv_vfabs_v_f32, LMUL_512)(reg, VEC_ELEM_NUM));
}
@@ -867,6 +875,27 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
}
};
struct INT8Vec16 : public Vec<INT8Vec16> {
constexpr static int VEC_ELEM_NUM = 16;
fixed_i8x16_t reg;
explicit INT8Vec16(const FP32Vec16& vec) {
auto i32_vec =
RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_512)(vec.reg, VEC_ELEM_NUM);
auto i16_vec = RVVI(__riscv_vnclip_wx_i16, LMUL_256)(
i32_vec, 0, __RISCV_VXRM_RNU, VEC_ELEM_NUM);
reg = RVVI(__riscv_vnclip_wx_i8, LMUL_128)(i16_vec, 0, __RISCV_VXRM_RNU,
VEC_ELEM_NUM);
}
void save(int8_t* ptr) const {
RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, VEC_ELEM_NUM);
}
void save(int8_t* ptr, int elem_num) const {
RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, elem_num);
}
};
// ============================================================================
// Type Traits & Global Helpers
// ============================================================================
+3 -2
View File
@@ -329,8 +329,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
// Quantization
#if defined(__AVX512F__) || defined(__AVX2__) || \
(defined(__aarch64__) && !defined(__APPLE__)) || defined(__powerpc64__)
#if defined(__AVX512F__) || defined(__AVX2__) || \
(defined(__aarch64__) && !defined(__APPLE__)) || defined(__powerpc64__) || \
defined(__riscv_v)
// Helper function to release oneDNN handlers
ops.def("release_dnnl_matmul_handler(int handler) -> ()",
&release_dnnl_matmul_handler);