vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : remove ggml_task_type and GGML_PERF

Browse source 
ggml-ci
This commit is contained in:
slaren 2024-06-19 17:59:18 +02:00
parent 9c77ec1d74
commit d27f26ea0c
4 changed files with 308 additions and 910 deletions
Download patch file Download diff file Expand all files Collapse all files

1060
ggml.c
View file

File diff suppressed because it is too large Load diff

35
ggml.h
View file

@ -585,11 +585,7 @@ extern "C" {
struct ggml_tensor * grad;
struct ggml_tensor * src[GGML_MAX_SRC];
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
// source tensor and offset for views
struct ggml_tensor * view_src;
size_t view_offs;
@ -599,7 +595,7 @@ extern "C" {
void * extra; // extra things e.g. for ggml-cuda.cu
char padding[8];
// char padding[4];
};
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@ -646,11 +642,6 @@ extern "C" {
struct ggml_hash_set visited_hash_table;
enum ggml_cgraph_eval_order order;
// performance
int perf_runs;
int64_t perf_cycles;
int64_t perf_time_us;
};
// scratch buffer
@ -667,28 +658,6 @@ extern "C" {
bool no_alloc; // don't allocate memory for the tensor data
};
// compute types
// NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
// This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
enum ggml_task_type {
GGML_TASK_TYPE_INIT = 0,
GGML_TASK_TYPE_COMPUTE,
GGML_TASK_TYPE_FINALIZE,
};
struct ggml_compute_params {
enum ggml_task_type type;
// ith = thread index, nth = number of threads
int ith, nth;
// work buffer for all threads
size_t wsize;
void * wdata;
};
// numa strategies
enum ggml_numa_strategy {
GGML_NUMA_STRATEGY_DISABLED = 0,

37
sgemm.cpp
View file

@ -249,8 +249,7 @@ class tinyBLAS {
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n, int task) {
if (task == GGML_TASK_TYPE_COMPUTE)
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
@ -458,8 +457,7 @@ class tinyBLAS_Q0_ARM {
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n, int task) {
if (task == GGML_TASK_TYPE_COMPUTE)
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
@ -596,8 +594,7 @@ class tinyBLAS_Q0_AVX {
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n, int task) {
if (task == GGML_TASK_TYPE_COMPUTE)
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
@ -829,7 +826,7 @@ class tinyBLAS_Q0_AVX {
* For example, for single-threaded single-precision GEMM you can say
*
* llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
* 0, 1, GGML_TASK_TYPE_COMPUTE,
* 0, 1,
* GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
*
* @param m is rows in `A` and `C`
@ -843,14 +840,13 @@ class tinyBLAS_Q0_AVX {
* @param ldc is row stride of `C`
* @param ith is thread id (must be less than `nth`)
* @param nth is number of threads (must be greater than zero)
* @param task is GGML task type
* @param Atype is GGML data type of `A`
* @param Btype is GGML data type of `B`
* @param Ctype is GGML data type of `C`
* @return true if this function was able to service the matmul request
*/
bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
int64_t ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype) {
int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
assert(m >= 0);
assert(n >= 0);
@ -877,7 +873,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif defined(__AVX__) || defined(__AVX2__)
if (k % 8)
@ -887,7 +883,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif defined(__ARM_NEON)
if (n < 4)
@ -899,7 +895,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#else
return false;
@ -917,7 +913,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
if (k % 8)
@ -929,7 +925,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
if (n < 8)
@ -943,7 +939,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif defined(__ARM_NEON) && !defined(_MSC_VER)
if (k % 4)
@ -955,7 +951,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#else
return false;
@ -971,7 +967,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_DOTPROD)
tinyBLAS_Q0_ARM<block_q8_0> tb{
@ -979,7 +975,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#else
return false;
@ -995,7 +991,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_DOTPROD)
tinyBLAS_Q0_ARM<block_q4_0> tb{
@ -1003,7 +999,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n, task);
tb.matmul(m, n);
return true;
#else
return false;
@ -1025,7 +1021,6 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
(void)ldc;
(void)ith;
(void)nth;
(void)task;
(void)Atype;
(void)Btype;
(void)Ctype;

2
sgemm.h
View file

@ -7,7 +7,7 @@ extern "C" {
bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
const void *, int64_t, void *, int64_t, int, int,
int, int, int, int);
int, int, int);
#ifdef __cplusplus
}