[Refactor] Remove dead kernel code (#57621)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-09-20 10:39:23 -07:00
committed by GitHub
parent 49ee12d742
commit 4868312128
12 changed files with 4 additions and 559 deletions
@@ -74,18 +74,9 @@ class MatrixView_half_rw {
const int width)
: data(data), height(height), width(width) {}
__device__ __forceinline__ half item(int row, int column) const {
return data[row * width + column];
}
__device__ __forceinline__ half* item_ptr(int row, int column) {
return &data[row * width + column];
}
__device__ __forceinline__ void set(int row, int column, half value) {
data[row * width + column] = value;
}
__device__ __forceinline__ void set_half2(int row, int column, half2 value) {
((half2*)data)[(row * width + column) / 2] = value;
}
__device__ __forceinline__ void set4(int row, int column, half v0, half v1,
half v2, half v3) {
@@ -113,14 +104,6 @@ class MatrixView_q4_row {
return (data[row * width / 8 + column / 8] >> shift) & 0x0f;
}
__device__ __forceinline__ void item2(int (&items)[2], int row,
int column) const {
int shift = (column & 0x07) * 4;
uint32_t d = data[row * width / 8 + column / 8] >> shift;
items[0] = d & 0x0f;
items[1] = (d >> 4) & 0x0f;
}
__device__ __forceinline__ void item4(int (&items)[4], int row,
int column) const {
int shift = (column & 0x07) * 4;
@@ -148,14 +131,6 @@ class MatrixView_q2_row {
return (data[row * width / 16 + column / 16] >> shift) & 0x03;
}
__device__ __forceinline__ void item2(int (&items)[2], int row,
int column) const {
int shift = (column & 0x0f) * 2;
uint32_t d = data[row * width / 16 + column / 16] >> shift;
items[0] = d & 0x03;
items[1] = (d >> 2) & 0x03;
}
__device__ __forceinline__ void item4(int (&items)[4], int row,
int column) const {
int shift = (column & 0x0f) * 2;
@@ -237,14 +212,6 @@ class MatrixView_q8_row {
return (data[row * width / 4 + column / 4] >> shift) & 0xff;
}
__device__ __forceinline__ void item2(int (&items)[2], int row,
int column) const {
int shift = (column & 0x03) * 8;
uint32_t d = data[row * width / 4 + column / 4] >> shift;
items[0] = d & 0xff;
items[1] = (d >> 8) & 0xff;
}
__device__ __forceinline__ void item4(int (&items)[4], int row,
int column) const {
int shift = (column & 0x03) * 2;
@@ -67,14 +67,6 @@ __device__ inline void cp_async2_ca_pred(void* smem_ptr, const void* glob_ptr,
}
}
__device__ inline void cp_async4_ca_pred(void* smem_ptr, const void* glob_ptr,
bool pred = true) {
if (pred) {
reinterpret_cast<int4*>(smem_ptr)[0] =
reinterpret_cast<const int4*>(glob_ptr)[0];
}
}
__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
bool pred = true) {
if (pred) {
@@ -121,19 +113,6 @@ __device__ inline void cp_async2_ca_pred(void* smem_ptr, const void* glob_ptr,
"r"(smem), "l"(glob_ptr), "n"(BYTES));
}
__device__ inline void cp_async4_ca_pred(void* smem_ptr, const void* glob_ptr,
bool pred = true) {
const int BYTES = 16;
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
asm volatile(
"{\n"
" .reg .pred p;\n"
" setp.ne.b32 p, %0, 0;\n"
" @p cp.async.ca.shared.global [%1], [%2], %3;\n"
"}\n" ::"r"((int)pred),
"r"(smem), "l"(glob_ptr), "n"(BYTES));
}
__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
bool pred = true) {
const int BYTES = 16;
@@ -170,4 +149,4 @@ __device__ inline void cp_async_wait() {
} // namespace MARLIN_NAMESPACE_NAME
#endif
#endif