diff options
author | aditya <bluenerd@protonmail.com> | 2023-08-10 12:32:35 +0530 |
---|---|---|
committer | aditya <bluenerd@protonmail.com> | 2023-08-10 12:32:35 +0530 |
commit | a9ff78b3f48dc9f81943c41531c4959ce7e2ae9d (patch) | |
tree | 49ee8c3c9148038f04112802265d928ef1aba428 /ggml.c | |
parent | 2516af4cd61f509c995b4f78fdf123cba33f3509 (diff) | |
parent | 916a9acdd0a411426690400ebe2bb7ce840a6bba (diff) |
resolve merge conflict
Diffstat (limited to 'ggml.c')
-rw-r--r-- | ggml.c | 2577 |
1 files changed, 1446 insertions, 1131 deletions
@@ -31,11 +31,17 @@ #include <unistd.h> #endif +// static_assert should be a #define, but if it's not, +// fall back to the _Static_assert C11 keyword. // if C99 - static_assert is noop // ref: https://stackoverflow.com/a/53923785/4039976 #ifndef static_assert +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L) +#define static_assert(cond, msg) _Static_assert(cond, msg) +#else #define static_assert(cond, msg) struct global_scope_noop_trick #endif +#endif #if defined(_MSC_VER) // disable "possible loss of data" to avoid hundreds of casts @@ -112,10 +118,6 @@ typedef void * thread_ret_t; #endif #endif -#ifdef __HAIKU__ -#define static_assert(cond, msg) _Static_assert(cond, msg) -#endif - /*#define GGML_PERF*/ #define GGML_DEBUG 0 #define GGML_GELU_FP16 @@ -193,8 +195,8 @@ typedef void * thread_ret_t; #define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN) #define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr) #else -inline static void* ggml_aligned_malloc(size_t size) { - void* aligned_memory = NULL; +inline static void * ggml_aligned_malloc(size_t size) { + void * aligned_memory = NULL; #ifdef GGML_USE_METAL int result = posix_memalign(&aligned_memory, getpagesize(), size); #else @@ -3438,7 +3440,9 @@ inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float //inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] *= v; } inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { -#if defined(GGML_SIMD) +#if defined(GGML_USE_ACCELERATE) + vDSP_vsmul(y, 1, &v, y, 1, n); +#elif defined(GGML_SIMD) const int np = (n & ~(GGML_F32_STEP - 1)); GGML_F32_VEC vx = GGML_F32_VEC_SET1(v); @@ -3601,7 +3605,7 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) { #endif } -inline static void ggml_vec_sum_ggf(const int n, ggml_float * s, const float * x) { +inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) { ggml_float sum = 0.0; for (int i = 0; i < n; ++i) { sum += (ggml_float)x[i]; @@ -3609,6 +3613,14 @@ inline static void ggml_vec_sum_ggf(const int n, ggml_float * s, const float * x *s = sum; } +inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) { + float sum = 0.0f; + for (int i = 0; i < n; ++i) { + sum += GGML_FP16_TO_FP32(x[i]); + } + *s = sum; +} + inline static void ggml_vec_max_f32(const int n, float * s, const float * x) { #ifndef GGML_USE_ACCELERATE float max = -INFINITY; @@ -3748,16 +3760,6 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "ARGMAX", "REPEAT", "REPEAT_BACK", - "ABS", - "SGN", - "NEG", - "STEP", - "TANH", - "ELU", - "RELU", - "GELU", - "GELU_QUICK", - "SILU", "SILU_BACK", "NORM", "RMS_NORM", @@ -3787,6 +3789,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CLAMP", "CONV_1D", "CONV_2D", + "POOL_1D", + "POOL_2D", "FLASH_ATTN", "FLASH_FF", @@ -3794,6 +3798,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "WIN_PART", "WIN_UNPART", + "UNARY", + "MAP_UNARY", "MAP_BINARY", @@ -3805,7 +3811,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 66, "GGML_OP_COUNT != 66"); +static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -3826,16 +3832,6 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "argmax(x)", "repeat(x)", "repeat_back(x)", - "abs(x)", - "sgn(x)", - "-x", - "step(x)", - "tanh(x)", - "elu(x)", - "relu(x)", - "gelu(x)", - "gelu_quick(x)", - "silu(x)", "silu_back(x)", "norm(x)", "rms_norm(x)", @@ -3865,6 +3861,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "clamp(x)", "conv_1d(x)", "conv_2d(x)", + "pool_1d(x)", + "pool_2d(x)", "flash_attn(x)", "flash_ff(x)", @@ -3872,6 +3870,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "win_part(x)", "win_unpart(x)", + "unary(x)", + "f(x)", "f(x,y)", @@ -3883,7 +3883,9 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 66, "GGML_OP_COUNT != 66"); +static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62"); + +static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); @@ -4069,8 +4071,8 @@ bool ggml_is_numa(void) { //////////////////////////////////////////////////////////////////////////////// void ggml_print_object(const struct ggml_object * obj) { - GGML_PRINT(" - ggml_object: offset = %zu, size = %zu, next = %p\n", - obj->offs, obj->size, (const void *) obj->next); + GGML_PRINT(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n", + obj->type, obj->offs, obj->size, (const void *) obj->next); } void ggml_print_objects(const struct ggml_context * ctx) { @@ -4108,7 +4110,7 @@ size_t ggml_nbytes(const struct ggml_tensor * tensor) { // // is enough, but just in case, adding the second part - return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]); + return GGML_PAD(MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]), GGML_MEM_ALIGN); } size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) { @@ -4137,6 +4139,10 @@ const char * ggml_op_name(enum ggml_op op) { return GGML_OP_NAME[op]; } +const char * ggml_op_symbol(enum ggml_op op) { + return GGML_OP_SYMBOL[op]; +} + size_t ggml_element_size(const struct ggml_tensor * tensor) { return GGML_TYPE_SIZE[tensor->type]; } @@ -4162,10 +4168,9 @@ static inline bool ggml_is_matrix(const struct ggml_tensor * tensor) { static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); - return - (t0->ne[0] == t1->ne[0]) && - (t0->ne[2] == t1->ne[2]) && - (t0->ne[3] == t1->ne[3]); + return (t0->ne[0] == t1->ne[0]) && + (t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable + (t1->ne[3]%t0->ne[3] == 0); } static inline bool ggml_can_out_prod(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { @@ -4207,7 +4212,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { } size_t ggml_tensor_overhead(void) { - return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE + 16; + return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE; } bool ggml_is_transposed(const struct ggml_tensor * tensor) { @@ -4224,6 +4229,15 @@ bool ggml_is_contiguous(const struct ggml_tensor * tensor) { tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } +static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * tensor) { + static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); + + return + tensor->nb[0] == GGML_TYPE_SIZE[tensor->type] && + tensor->nb[2] == tensor->nb[1]*tensor->ne[1] && + tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; +} + bool ggml_is_permuted(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); @@ -4239,7 +4253,7 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) { tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } -static inline bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { +bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); return @@ -4369,7 +4383,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { return NULL; } - const size_t mem_size = (params.mem_size + GGML_MEM_ALIGN - 1) & ~(GGML_MEM_ALIGN - 1); + const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN); *ctx = (struct ggml_context) { /*.mem_size =*/ mem_size, @@ -4405,8 +4419,8 @@ void ggml_free(struct ggml_context * ctx) { if (&g_state.contexts[i].context == ctx) { g_state.contexts[i].used = false; - GGML_PRINT_DEBUG("%s: context %d with %d objects has been freed. memory used = %zu\n", - __func__, i, ctx->n_objects, ctx->objects_end->offs + ctx->objects_end->size); + GGML_PRINT_DEBUG("%s: context %d has been freed. memory used = %zu\n", + __func__, i, ggml_used_mem(ctx)); if (ctx->mem_buffer_owned) { GGML_ALIGNED_FREE(ctx->mem_buffer); @@ -4436,6 +4450,10 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) return result; } +bool ggml_get_no_alloc(struct ggml_context * ctx) { + return ctx->no_alloc; +} + void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc) { ctx->no_alloc = no_alloc; } @@ -4454,12 +4472,14 @@ size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) { struct ggml_object * obj = ctx->objects_begin; while (obj != NULL) { - struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs); + if (obj->type == GGML_OBJECT_TENSOR) { + struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs); - const size_t size = ggml_nbytes(tensor); + const size_t size = ggml_nbytes(tensor); - if (max_size < size) { - max_size = size; + if (max_size < size) { + max_size = size; + } } obj = obj->next; @@ -4473,7 +4493,7 @@ size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) { // this is an error prone process, but it is necessary to support inplace // operators when using scratch buffers // TODO: implement a better way -void ggml_scratch_save(struct ggml_context * ctx) { +static void ggml_scratch_save(struct ggml_context * ctx) { // this is needed to allow opt tensors to store their data // TODO: again, need to find a better way ctx->no_alloc_save = ctx->no_alloc; @@ -4483,7 +4503,7 @@ void ggml_scratch_save(struct ggml_context * ctx) { ctx->scratch.data = NULL; } -void ggml_scratch_load(struct ggml_context * ctx) { +static void ggml_scratch_load(struct ggml_context * ctx) { ctx->no_alloc = ctx->no_alloc_save; ctx->scratch = ctx->scratch_save; @@ -4491,12 +4511,7 @@ void ggml_scratch_load(struct ggml_context * ctx) { //////////////////////////////////////////////////////////////////////////////// -struct ggml_tensor * ggml_new_tensor_impl( - struct ggml_context * ctx, - enum ggml_type type, - int n_dims, - const int64_t* ne, - void* data) { +static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml_object_type type, size_t size) { // always insert objects at the end of the context's memory pool struct ggml_object * obj_cur = ctx->objects_end; @@ -4504,77 +4519,81 @@ struct ggml_tensor * ggml_new_tensor_impl( const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size; const size_t cur_end = cur_offs + cur_size; - size_t size_needed = 0; - - if (data == NULL && !ctx->no_alloc) { - size_needed += GGML_TYPE_SIZE[type]*(ne[0]/GGML_BLCK_SIZE[type]); - for (int i = 1; i < n_dims; i++) { - size_needed *= ne[i]; - } - // align to GGML_MEM_ALIGN - size_needed = ((size_needed + GGML_MEM_ALIGN - 1)/GGML_MEM_ALIGN)*GGML_MEM_ALIGN; - } + // align to GGML_MEM_ALIGN + size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN); char * const mem_buffer = ctx->mem_buffer; struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end); - if (ctx->scratch.data == NULL || data != NULL) { - size_needed += GGML_TENSOR_SIZE; + if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) { + GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", + __func__, cur_end + size_needed, ctx->mem_size); + assert(false); + return NULL; + } - if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) { - GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", - __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size); - assert(false); - return NULL; - } + *obj_new = (struct ggml_object) { + .offs = cur_end + GGML_OBJECT_SIZE, + .size = size_needed, + .next = NULL, + .type = type, + }; - *obj_new = (struct ggml_object) { - .offs = cur_end + GGML_OBJECT_SIZE, - .size = size_needed, - .next = NULL, - }; + ggml_assert_aligned(mem_buffer + obj_new->offs); + + if (obj_cur != NULL) { + obj_cur->next = obj_new; } else { - if (ctx->scratch.offs + size_needed > ctx->scratch.size) { - GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", - __func__, ctx->scratch.offs + size_needed, ctx->scratch.size); - assert(false); - return NULL; + // this is the first object in this context + ctx->objects_begin = obj_new; + } + + ctx->objects_end = obj_new; + + //printf("%s: inserted new object at %zu, size = %zu\n", __func__, cur_end, obj_new->size); + + return obj_new; +} + +static struct ggml_tensor * ggml_new_tensor_impl( + struct ggml_context * ctx, + enum ggml_type type, + int n_dims, + const int64_t * ne, + void * data) { + + assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS); + + size_t data_size = 0; + + if (data == NULL && !ctx->no_alloc) { + data_size += GGML_TYPE_SIZE[type]*(ne[0]/GGML_BLCK_SIZE[type]); + for (int i = 1; i < n_dims; i++) { + data_size *= ne[i]; } + } - if (cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE > ctx->mem_size) { - GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", - __func__, cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE, ctx->mem_size); + if (ctx->scratch.data != NULL && data == NULL) { + // allocate tensor data in the scratch buffer + if (ctx->scratch.offs + data_size > ctx->scratch.size) { + GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", + __func__, ctx->scratch.offs + data_size, ctx->scratch.size); assert(false); return NULL; } data = (char * const) ctx->scratch.data + ctx->scratch.offs; - *obj_new = (struct ggml_object) { - .offs = cur_end + GGML_OBJECT_SIZE, - .size = GGML_TENSOR_SIZE, - .next = NULL, - }; - - //printf("scratch offs = %zu, size_needed = %zu\n", ctx->scratch.offs, size_needed); + ctx->scratch.offs += data_size; - ctx->scratch.offs += size_needed; + data_size = 0; } - if (obj_cur != NULL) { - obj_cur->next = obj_new; - } else { - // this is the first object in this context - ctx->objects_begin = obj_new; - } - - ctx->objects_end = obj_new; - - //printf("%s: inserted new object at %zu, size = %zu\n", __func__, cur_end, obj_new->size); + struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + data_size); - struct ggml_tensor * const result = (struct ggml_tensor *)(mem_buffer + obj_new->offs); + // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here - ggml_assert_aligned(result); + struct ggml_tensor * const result = (struct ggml_tensor *)((char *)ctx->mem_buffer + obj_new->offs); *result = (struct ggml_tensor) { /*.type =*/ type, @@ -4583,6 +4602,7 @@ struct ggml_tensor * ggml_new_tensor_impl( /*.ne =*/ { 1, 1, 1, 1 }, /*.nb =*/ { 0, 0, 0, 0 }, /*.op =*/ GGML_OP_NONE, + /*.op_params =*/ { 0 }, /*.is_param =*/ false, /*.grad =*/ NULL, /*.src =*/ { NULL }, @@ -4613,24 +4633,40 @@ struct ggml_tensor * ggml_new_tensor_impl( return result; } +static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) { + GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings + assert(params_size <= GGML_MAX_OP_PARAMS); + memcpy(tensor->op_params, params, params_size); +} + +static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) { + assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); + return ((const int32_t *)(tensor->op_params))[i]; +} + +static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) { + assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); + ((int32_t *)(tensor->op_params))[i] = value; +} + struct ggml_tensor * ggml_new_tensor( struct ggml_context * ctx, - enum ggml_type type, - int n_dims, - const int64_t * ne) { + enum ggml_type type, + int n_dims, + const int64_t * ne) { return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL); } struct ggml_tensor * ggml_new_tensor_1d( struct ggml_context * ctx, - enum ggml_type type, + enum ggml_type type, int64_t ne0) { return ggml_new_tensor(ctx, type, 1, &ne0); } struct ggml_tensor * ggml_new_tensor_2d( struct ggml_context * ctx, - enum ggml_type type, + enum ggml_type type, int64_t ne0, int64_t ne1) { const int64_t ne[2] = { ne0, ne1 }; @@ -4639,7 +4675,7 @@ struct ggml_tensor * ggml_new_tensor_2d( struct ggml_tensor * ggml_new_tensor_3d( struct ggml_context * ctx, - enum ggml_type type, + enum ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2) { @@ -4944,6 +4980,11 @@ float * ggml_get_data_f32(const struct ggml_tensor * tensor) { return (float *)(tensor->data); } +enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) { + GGML_ASSERT(tensor->op == GGML_OP_UNARY); + return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0); +} + const char * ggml_get_name(const struct ggml_tensor * tensor) { return tensor->name; } @@ -4982,9 +5023,11 @@ struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * nam char * const mem_buffer = ctx->mem_buffer; while (obj != NULL) { - struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs); - if (strcmp(cur->name, name) == 0) { - return cur; + if (obj->type == GGML_OBJECT_TENSOR) { + struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs); + if (strcmp(cur->name, name) == 0) { + return cur; + } } obj = obj->next; @@ -4997,7 +5040,7 @@ struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * nam // ggml_dup -struct ggml_tensor * ggml_dup_impl( +static struct ggml_tensor * ggml_dup_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -5012,7 +5055,6 @@ struct ggml_tensor * ggml_dup_impl( result->op = GGML_OP_DUP; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5031,7 +5073,7 @@ struct ggml_tensor * ggml_dup_inplace( // ggml_add -struct ggml_tensor * ggml_add_impl( +static struct ggml_tensor * ggml_add_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5074,7 +5116,7 @@ struct ggml_tensor * ggml_add_inplace( // ggml_add1 -struct ggml_tensor * ggml_add1_impl( +static struct ggml_tensor * ggml_add1_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5114,7 +5156,7 @@ struct ggml_tensor * ggml_add1_inplace( // ggml_acc -struct ggml_tensor * ggml_acc_impl( +static struct ggml_tensor * ggml_acc_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5136,23 +5178,13 @@ struct ggml_tensor * ggml_acc_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 5); - - ((int32_t *) c->data)[0] = nb1; - ((int32_t *) c->data)[1] = nb2; - ((int32_t *) c->data)[2] = nb3; - ((int32_t *) c->data)[3] = offset; - ((int32_t *) c->data)[4] = inplace ? 1 : 0; - - ggml_scratch_load(ctx); + int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ACC; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = c; return result; } @@ -5181,7 +5213,7 @@ struct ggml_tensor * ggml_acc_inplace( // ggml_sub -struct ggml_tensor * ggml_sub_impl( +static struct ggml_tensor * ggml_sub_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5220,7 +5252,7 @@ struct ggml_tensor * ggml_sub_inplace( // ggml_mul -struct ggml_tensor * ggml_mul_impl( +static struct ggml_tensor * ggml_mul_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5267,7 +5299,7 @@ struct ggml_tensor * ggml_mul_inplace( // ggml_div -struct ggml_tensor * ggml_div_impl( +static struct ggml_tensor * ggml_div_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5310,7 +5342,7 @@ struct ggml_tensor * ggml_div_inplace( // ggml_sqr -struct ggml_tensor * ggml_sqr_impl( +static struct ggml_tensor * ggml_sqr_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -5325,7 +5357,6 @@ struct ggml_tensor * ggml_sqr_impl( result->op = GGML_OP_SQR; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5344,7 +5375,7 @@ struct ggml_tensor * ggml_sqr_inplace( // ggml_sqrt -struct ggml_tensor * ggml_sqrt_impl( +static struct ggml_tensor * ggml_sqrt_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -5359,7 +5390,6 @@ struct ggml_tensor * ggml_sqrt_impl( result->op = GGML_OP_SQRT; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5379,7 +5409,7 @@ struct ggml_tensor * ggml_sqrt_inplace( // ggml_log -struct ggml_tensor * ggml_log_impl( +static struct ggml_tensor * ggml_log_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -5394,7 +5424,6 @@ struct ggml_tensor * ggml_log_impl( result->op = GGML_OP_LOG; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5427,7 +5456,6 @@ struct ggml_tensor * ggml_sum( result->op = GGML_OP_SUM; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5454,7 +5482,6 @@ struct ggml_tensor * ggml_sum_rows( result->op = GGML_OP_SUM_ROWS; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5477,7 +5504,6 @@ struct ggml_tensor * ggml_mean( result->op = GGML_OP_MEAN; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5501,7 +5527,6 @@ struct ggml_tensor * ggml_argmax( result->op = GGML_OP_ARGMAX; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -5564,343 +5589,142 @@ struct ggml_tensor * ggml_repeat_back( // ggml_abs -struct ggml_tensor * ggml_abs_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_ABS; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_abs( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_abs_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_ABS); } struct ggml_tensor * ggml_abs_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_abs_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_ABS); } - // ggml_sgn -struct ggml_tensor * ggml_sgn_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_SGN; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_sgn( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_sgn_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_SGN); } struct ggml_tensor * ggml_sgn_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_sgn_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SGN); } // ggml_neg -struct ggml_tensor * ggml_neg_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_NEG; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_neg( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_neg_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_NEG); } struct ggml_tensor * ggml_neg_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_neg_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_NEG); } // ggml_step -struct ggml_tensor * ggml_step_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_STEP; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_step( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_step_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_STEP); } struct ggml_tensor * ggml_step_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_step_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_STEP); } // ggml_tanh -struct ggml_tensor * ggml_tanh_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_TANH; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_tanh( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_tanh_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_TANH); } struct ggml_tensor * ggml_tanh_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_tanh_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_TANH); } // ggml_elu -struct ggml_tensor * ggml_elu_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_ELU; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_elu( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_elu_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_ELU); } struct ggml_tensor * ggml_elu_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_elu_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_ELU); } // ggml_relu -struct ggml_tensor * ggml_relu_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_RELU; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_relu( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_relu_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_RELU); } struct ggml_tensor * ggml_relu_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_relu_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_RELU); } // ggml_gelu -struct ggml_tensor * ggml_gelu_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_GELU; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_gelu( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_gelu_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_GELU); } struct ggml_tensor * ggml_gelu_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_gelu_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_GELU); } // ggml_gelu_quick -struct ggml_tensor * ggml_gelu_quick_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_GELU_QUICK; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_gelu_quick( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_gelu_quick_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_GELU_QUICK); } struct ggml_tensor * ggml_gelu_quick_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_gelu_quick_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_GELU_QUICK); } // ggml_silu -struct ggml_tensor * ggml_silu_impl( - struct ggml_context * ctx, - struct ggml_tensor * a, - bool inplace) { - bool is_node = false; - - if (!inplace && (a->grad)) { - is_node = true; - } - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - result->op = GGML_OP_SILU; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = NULL; - - return result; -} - struct ggml_tensor * ggml_silu( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_silu_impl(ctx, a, false); + return ggml_unary(ctx, a, GGML_UNARY_OP_SILU); } struct ggml_tensor * ggml_silu_inplace( struct ggml_context * ctx, struct ggml_tensor * a) { - return ggml_silu_impl(ctx, a, true); + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SILU); } // ggml_silu_back @@ -5928,7 +5752,7 @@ struct ggml_tensor * ggml_silu_back( // ggml_norm -struct ggml_tensor * ggml_norm_impl( +static struct ggml_tensor * ggml_norm_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -5941,10 +5765,11 @@ struct ggml_tensor * ggml_norm_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + // TODO: maybe store epsilon here? + result->op = GGML_OP_NORM; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; // TODO: maybe store epsilon here? return result; } @@ -5961,9 +5786,10 @@ struct ggml_tensor * ggml_norm_inplace( return ggml_norm_impl(ctx, a, true); } -struct ggml_tensor * ggml_rms_norm_impl( +static struct ggml_tensor * ggml_rms_norm_impl( struct ggml_context * ctx, struct ggml_tensor * a, + float eps, bool inplace) { bool is_node = false; @@ -5973,24 +5799,27 @@ struct ggml_tensor * ggml_rms_norm_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + ggml_set_op_params(result, &eps, sizeof(eps)); + result->op = GGML_OP_RMS_NORM; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; // TODO: maybe store epsilon here? return result; } struct ggml_tensor * ggml_rms_norm( struct ggml_context * ctx, - struct ggml_tensor * a) { - return ggml_rms_norm_impl(ctx, a, false); + struct ggml_tensor * a, + float eps) { + return ggml_rms_norm_impl(ctx, a, eps, false); } struct ggml_tensor * ggml_rms_norm_inplace( struct ggml_context * ctx, - struct ggml_tensor * a) { - return ggml_rms_norm_impl(ctx, a, true); + struct ggml_tensor * a, + float eps) { + return ggml_rms_norm_impl(ctx, a, eps, true); } struct ggml_tensor * ggml_rms_norm_back( @@ -6030,8 +5859,8 @@ struct ggml_tensor * ggml_mul_mat( is_node = true; } - const int64_t ne[4] = { a->ne[1], b->ne[1], a->ne[2], b->ne[3] }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MIN(a->n_dims, b->n_dims), ne); + const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne); result->op = GGML_OP_MUL_MAT; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -6069,7 +5898,7 @@ struct ggml_tensor * ggml_out_prod( // ggml_scale -struct ggml_tensor * ggml_scale_impl( +static struct ggml_tensor * ggml_scale_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -6109,7 +5938,7 @@ struct ggml_tensor * ggml_scale_inplace( // ggml_set -struct ggml_tensor * ggml_set_impl( +static struct ggml_tensor * ggml_set_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -6129,23 +5958,13 @@ struct ggml_tensor * ggml_set_impl( // make a view of the destination struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 5); - - (( int32_t * ) c->data)[0] = nb1; - (( int32_t * ) c->data)[1] = nb2; - (( int32_t * ) c->data)[2] = nb3; - (( int32_t * ) c->data)[3] = offset; - (( int32_t * ) c->data)[4] = inplace ? 1 : 0; - - ggml_scratch_load(ctx); + int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_SET; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = c; return result; } @@ -6209,7 +6028,7 @@ struct ggml_tensor * ggml_set_2d_inplace( // ggml_cpy -struct ggml_tensor * ggml_cpy_impl( +static struct ggml_tensor * ggml_cpy_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -6254,7 +6073,7 @@ struct ggml_tensor * ggml_cpy_inplace( // ggml_cont -struct ggml_tensor * ggml_cont_impl( +static struct ggml_tensor * ggml_cont_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -6270,7 +6089,6 @@ struct ggml_tensor * ggml_cont_impl( result->op = GGML_OP_CONT; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6314,7 +6132,6 @@ struct ggml_tensor * ggml_reshape( result->op = GGML_OP_RESHAPE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6339,7 +6156,6 @@ struct ggml_tensor * ggml_reshape_1d( result->op = GGML_OP_RESHAPE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6365,7 +6181,6 @@ struct ggml_tensor * ggml_reshape_2d( result->op = GGML_OP_RESHAPE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6392,7 +6207,6 @@ struct ggml_tensor * ggml_reshape_3d( result->op = GGML_OP_RESHAPE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6421,13 +6235,33 @@ struct ggml_tensor * ggml_reshape_4d( result->op = GGML_OP_RESHAPE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } // ggml_view_1d +static struct ggml_tensor * ggml_view_tensor_offset( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_dims, + const int64_t * ne, + size_t offset) { + // don't calculate an offset from an unallocated tensor + void * data = NULL; + if (a->data != NULL) { + data = (char *) a->data + offset; + } + + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data); + + ggml_format_name(result, "%s (view)", a->name); + + ggml_set_op_params(result, &offset, sizeof(offset)); + + return result; +} + struct ggml_tensor * ggml_view_1d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -6440,22 +6274,11 @@ struct ggml_tensor * ggml_view_1d( is_node = true; } - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset); - ggml_format_name(result, "%s (view)", a->name); - - ggml_scratch_save(ctx); - - struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); - ggml_set_name(offs, "offset"); - memcpy(offs->data, &offset, 2*sizeof(int32_t)); - - ggml_scratch_load(ctx); + struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset); result->op = GGML_OP_VIEW; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - result->src[2] = offs; return result; } @@ -6478,16 +6301,7 @@ struct ggml_tensor * ggml_view_2d( const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, (char *) a->data + offset); - ggml_format_name(result, "%s (view)", a->name); - - ggml_scratch_save(ctx); - - struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); - ggml_set_name(offs, "offset"); - memcpy(offs->data, &offset, 2*sizeof(int32_t)); - - ggml_scratch_load(ctx); + struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset); result->nb[1] = nb1; result->nb[2] = result->nb[1]*ne1; @@ -6496,8 +6310,6 @@ struct ggml_tensor * ggml_view_2d( result->op = GGML_OP_VIEW; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - result->src[2] = offs; return result; } @@ -6522,16 +6334,7 @@ struct ggml_tensor * ggml_view_3d( const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, (char *) a->data + offset); - ggml_format_name(result, "%s (view)", a->name); - - ggml_scratch_save(ctx); - - struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); - ggml_set_name(offs, "offset"); - memcpy(offs->data, &offset, 2*sizeof(int32_t)); - - ggml_scratch_load(ctx); + struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset); result->nb[1] = nb1; result->nb[2] = nb2; @@ -6540,8 +6343,6 @@ struct ggml_tensor * ggml_view_3d( result->op = GGML_OP_VIEW; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - result->src[2] = offs; return result; } @@ -6568,16 +6369,7 @@ struct ggml_tensor * ggml_view_4d( const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset); - ggml_format_name(result, "%s (view)", a->name); - - ggml_scratch_save(ctx); - - struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); - ggml_set_name(offs, "offset"); - memcpy(offs->data, &offset, 2*sizeof(int32_t)); - - ggml_scratch_load(ctx); + struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset); result->nb[1] = nb1; result->nb[2] = nb2; @@ -6586,8 +6378,6 @@ struct ggml_tensor * ggml_view_4d( result->op = GGML_OP_VIEW; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - result->src[2] = offs; return result; } @@ -6648,22 +6438,9 @@ struct ggml_tensor * ggml_permute( result->op = GGML_OP_PERMUTE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - - if (is_node) { - ggml_scratch_save(ctx); - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 4); - - ((int32_t *) b->data)[0] = axis0; - ((int32_t *) b->data)[1] = axis1; - ((int32_t *) b->data)[2] = axis2; - ((int32_t *) b->data)[3] = axis3; - - ggml_scratch_load(ctx); - - result->src[2] = b; - } + int32_t params[] = { axis0, axis1, axis2, axis3 }; + ggml_set_op_params(result, params, sizeof(params)); return result; } @@ -6691,7 +6468,6 @@ struct ggml_tensor * ggml_transpose( result->op = GGML_OP_TRANSPOSE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6769,7 +6545,6 @@ struct ggml_tensor * ggml_diag( result->op = GGML_OP_DIAG; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6777,7 +6552,7 @@ struct ggml_tensor * ggml_diag( // ggml_diag_mask_inf -struct ggml_tensor * ggml_diag_mask_inf_impl( +static struct ggml_tensor * ggml_diag_mask_inf_impl( struct ggml_context * ctx, struct ggml_tensor * a, int n_past, @@ -6790,19 +6565,12 @@ struct ggml_tensor * ggml_diag_mask_inf_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); - - ((int32_t *) b->data)[0] = n_past; - ((int32_t *) b->data)[1] = inplace ? 1 : 0; - - ggml_scratch_load(ctx); + int32_t params[] = { n_past, inplace ? 1 : 0 }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_DIAG_MASK_INF; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = b; return result; } @@ -6824,7 +6592,7 @@ struct ggml_tensor * ggml_diag_mask_inf_inplace( // ggml_diag_mask_zero -struct ggml_tensor * ggml_diag_mask_zero_impl( +static struct ggml_tensor * ggml_diag_mask_zero_impl( struct ggml_context * ctx, struct ggml_tensor * a, int n_past, @@ -6837,20 +6605,12 @@ struct ggml_tensor * ggml_diag_mask_zero_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); - ggml_set_name(b, "n_past, inplace"); - - ((int32_t *) b->data)[0] = n_past; - ((int32_t *) b->data)[1] = inplace ? 1 : 0; - - ggml_scratch_load(ctx); + int32_t params[] = { n_past, inplace ? 1 : 0 }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_DIAG_MASK_ZERO; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = b; return result; } @@ -6871,7 +6631,7 @@ struct ggml_tensor * ggml_diag_mask_zero_inplace( // ggml_soft_max -struct ggml_tensor * ggml_soft_max_impl( +static struct ggml_tensor * ggml_soft_max_impl( struct ggml_context * ctx, struct ggml_tensor * a, bool inplace) { @@ -6886,7 +6646,6 @@ struct ggml_tensor * ggml_soft_max_impl( result->op = GGML_OP_SOFT_MAX; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; return result; } @@ -6906,7 +6665,7 @@ struct ggml_tensor * ggml_soft_max_inplace( // ggml_soft_max_back -struct ggml_tensor * ggml_soft_max_back_impl( +static struct ggml_tensor * ggml_soft_max_back_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -6943,13 +6702,15 @@ struct ggml_tensor * ggml_soft_max_back_inplace( // ggml_rope -struct ggml_tensor * ggml_rope_impl( +static struct ggml_tensor * ggml_rope_impl( struct ggml_context * ctx, struct ggml_tensor * a, int n_past, int n_dims, int mode, int n_ctx, + float freq_base, + float freq_scale, bool inplace) { GGML_ASSERT(n_past >= 0); bool is_node = false; @@ -6960,21 +6721,14 @@ struct ggml_tensor * ggml_rope_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 4); - - ((int32_t *) b->data)[0] = n_past; - ((int32_t *) b->data)[1] = n_dims; - ((int32_t *) b->data)[2] = mode; - ((int32_t *) b->data)[3] = n_ctx; - - ggml_scratch_load(ctx); + int32_t params[6] = { n_past, n_dims, mode, n_ctx }; + memcpy(params + 4, &freq_base, sizeof(float)); + memcpy(params + 5, &freq_scale, sizeof(float)); + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = b; return result; } @@ -6986,7 +6740,7 @@ struct ggml_tensor * ggml_rope( int n_dims, int mode, int n_ctx) { - return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, false); + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, false); } struct ggml_tensor * ggml_rope_inplace( @@ -6996,7 +6750,31 @@ struct ggml_tensor * ggml_rope_inplace( int n_dims, int mode, int n_ctx) { - return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, true); + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, true); +} + +struct ggml_tensor * ggml_rope_custom( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode, + int n_ctx, + float freq_base, + float freq_scale) { + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, false); +} + +struct ggml_tensor * ggml_rope_custom_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode, + int n_ctx, + float freq_base, + float freq_scale) { + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, true); } // ggml_rope_back @@ -7006,7 +6784,8 @@ struct ggml_tensor * ggml_rope_back( struct ggml_tensor * a, int n_past, int n_dims, - int mode) { + int mode, + int n_ctx) { GGML_ASSERT(n_past >= 0); GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet"); @@ -7018,21 +6797,12 @@ struct ggml_tensor * ggml_rope_back( struct ggml_tensor * result = ggml_dup_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); - ggml_set_name(b, "n_past, n_dims, mode"); - - ((int32_t *) b->data)[0] = n_past; - ((int32_t *) b->data)[1] = n_dims; - ((int32_t *) b->data)[2] = mode; - - ggml_scratch_load(ctx); + int32_t params[] = { n_past, n_dims, mode, n_ctx }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE_BACK; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = b; return result; } @@ -7057,21 +6827,13 @@ struct ggml_tensor * ggml_alibi( //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_tensor * result = ggml_view_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); - - ((int32_t *) b->data)[0] = n_past; - ((int32_t *) b->data)[1] = n_head; - GGML_ASSERT(sizeof(float) == sizeof(int32_t)); - (((float *) b->data)[2]) = bias_max; - - ggml_scratch_load(ctx); + int32_t op_params[3] = { n_past, n_head }; + memcpy(op_params + 2, &bias_max, sizeof(float)); + ggml_set_op_params(result, op_params, sizeof(op_params)); result->op = GGML_OP_ALIBI; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = b; return result; } @@ -7093,19 +6855,12 @@ struct ggml_tensor * ggml_clamp( // TODO: when implement backward, fix this: struct ggml_tensor * result = ggml_view_tensor(ctx, a); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 2); - - ((float *) b->data)[0] = min; - ((float *) b->data)[1] = max; - - ggml_scratch_load(ctx); + float params[] = { min, max }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_CLAMP; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = b; return result; } @@ -7136,30 +6891,25 @@ GGML_API struct ggml_tensor * ggml_conv_1d( ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0), a->ne[2], 1, 1, }; - struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); - ggml_scratch_save(ctx); - struct ggml_tensor* c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); - ((int32_t*)c->data)[0] = s0; - ((int32_t*)c->data)[1] = p0; - ((int32_t*)c->data)[2] = d0; - ggml_scratch_load(ctx); + int32_t params[] = { s0, p0, d0 }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_CONV_1D; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = c; return result; } // ggml_conv_2d -struct ggml_tensor* ggml_conv_2d( - struct ggml_context* ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, +struct ggml_tensor * ggml_conv_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, int s0, int s1, int p0, @@ -7167,7 +6917,6 @@ struct ggml_tensor* ggml_conv_2d( int d0, int d1) { - GGML_ASSERT(b->ne[3] == 1); GGML_ASSERT(a->ne[2] == b->ne[2]); bool is_node = false; @@ -7179,25 +6928,17 @@ struct ggml_tensor* ggml_conv_2d( const int64_t ne[4] = { ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0), ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1), - a->ne[3], 1, + a->ne[3], b->ne[3], }; - struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - ggml_scratch_save(ctx); - struct ggml_tensor* c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 6); - ((int32_t*)c->data)[0] = s0; - ((int32_t*)c->data)[1] = s1; - ((int32_t*)c->data)[2] = p0; - ((int32_t*)c->data)[3] = p1; - ((int32_t*)c->data)[4] = d0; - ((int32_t*)c->data)[5] = d1; - ggml_scratch_load(ctx); + int32_t params[] = { s0, s1, p0, p1, d0, d1 }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_CONV_2D; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = c; return result; @@ -7205,7 +6946,7 @@ struct ggml_tensor* ggml_conv_2d( // ggml_conv_1d_ph -struct ggml_tensor* ggml_conv_1d_ph( +struct ggml_tensor * ggml_conv_1d_ph( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -7214,6 +6955,83 @@ struct ggml_tensor* ggml_conv_1d_ph( return ggml_conv_1d(ctx, a, b, s, a->ne[0] / 2, d); } + +// ggml_pool_* + +static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, int p) { + return (ins + 2 * p - ks) / s + 1; +} + +// ggml_pool_1d + +struct ggml_tensor * ggml_pool_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_op_pool op, + int k0, + int s0, + int p0) { + + bool is_node = false; + + if (a->grad) { + GGML_ASSERT(false); // TODO: implement backward + is_node = true; + } + + const int64_t ne[3] = { + ggml_calc_pool_output_size(a->ne[0], k0, s0, p0), + a->ne[1], + }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); + + int32_t params[] = { op, k0, s0, p0 }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_POOL_1D; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + + return result; +} + +// ggml_pool_2d + +struct ggml_tensor * ggml_pool_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_op_pool op, + int k0, + int k1, + int s0, + int s1, + int p0, + int p1) { + + bool is_node = false; + + if (a->grad) { + GGML_ASSERT(false); // TODO: implement backward + is_node = true; + } + + const int64_t ne[3] = { + ggml_calc_pool_output_size(a->ne[0], k0, s0, p0), + ggml_calc_pool_output_size(a->ne[1], k1, s1, p1), + a->ne[2], + }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne); + + int32_t params[] = { op, k0, k1, s0, s1, p0, p1 }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_POOL_2D; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + + return result; +} + // ggml_flash_attn struct ggml_tensor * ggml_flash_attn( @@ -7232,14 +7050,16 @@ struct ggml_tensor * ggml_flash_attn( } //struct ggml_tensor * result = ggml_dup_tensor(ctx, q); - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, q->ne); + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, q->n_dims, q->ne); + + int32_t t = masked ? 1 : 0; + ggml_set_op_params(result, &t, sizeof(t)); result->op = GGML_OP_FLASH_ATTN; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = q; result->src[1] = k; result->src[2] = v; - result->src[3] = ggml_new_i32(ctx, masked ? 1 : 0); return result; } @@ -7263,7 +7083,7 @@ struct ggml_tensor * ggml_flash_ff( } //struct ggml_tensor * result = ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, a->ne); + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne); result->op = GGML_OP_FLASH_FF; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -7329,13 +7149,15 @@ struct ggml_tensor * ggml_flash_attn_back( struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + int32_t masked_i = masked ? 1 : 0; + ggml_set_op_params(result, &masked_i, sizeof(masked_i)); + result->op = GGML_OP_FLASH_ATTN_BACK; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = q; result->src[1] = k; result->src[2] = v; result->src[3] = d; - result->src[4] = ggml_new_i32(ctx, masked ? 1 : 0); return result; } @@ -7368,21 +7190,12 @@ struct ggml_tensor * ggml_win_part( struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - ggml_scratch_save(ctx); - - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); - - ((int32_t *) b->data)[0] = npx; - ((int32_t *) b->data)[1] = npy; - ((int32_t *) b->data)[2] = w; - - ggml_scratch_load(ctx); + int32_t params[] = { npx, npy, w }; + ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_WIN_PART; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - result->src[2] = b; return result; } @@ -7407,26 +7220,57 @@ struct ggml_tensor * ggml_win_unpart( const int64_t ne[4] = { a->ne[0], w0, h0, 1, }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne); - ggml_scratch_save(ctx); + int32_t params[] = { w }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_WIN_UNPART; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; - struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1); + return result; +} - ((int32_t *) b->data)[0] = w; +// gmml_unary - ggml_scratch_load(ctx); +static struct ggml_tensor * ggml_unary_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_unary_op op, + bool inplace) { + bool is_node = false; - result->op = GGML_OP_WIN_UNPART; + if (!inplace && (a->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_set_op_params_i32(result, 0, (int32_t) op); + + result->op = GGML_OP_UNARY; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[1] = NULL; - result->src[2] = b; return result; } +struct ggml_tensor * ggml_unary( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_unary_op op) { + return ggml_unary_impl(ctx, a, op, false); +} + +struct ggml_tensor * ggml_unary_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + enum ggml_unary_op op) { + return ggml_unary_impl(ctx, a, op, true); +} + // ggml_map_unary -struct ggml_tensor * ggml_map_unary_impl_f32( +static struct ggml_tensor * ggml_map_unary_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, const ggml_unary_op_f32_t fun, @@ -7437,19 +7281,13 @@ struct ggml_tensor * ggml_map_unary_impl_f32( is_node = true; } - struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_scratch_save(ctx); - - struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); - *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_load(ctx); + ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); result->op = GGML_OP_MAP_UNARY; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[2] = addr_tensor; return result; } @@ -7470,7 +7308,7 @@ struct ggml_tensor * ggml_map_unary_inplace_f32( // ggml_map_binary -struct ggml_tensor * ggml_map_binary_impl_f32( +static struct ggml_tensor * ggml_map_binary_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -7484,20 +7322,14 @@ struct ggml_tensor * ggml_map_binary_impl_f32( is_node = true; } - struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_scratch_save(ctx); + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); - *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; - - ggml_scratch_load(ctx); + ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); result->op = GGML_OP_MAP_BINARY; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = addr_tensor; return result; } @@ -7518,9 +7350,9 @@ struct ggml_tensor * ggml_map_binary_inplace_f32( return ggml_map_binary_impl_f32(ctx, a, b, fun, true); } -// ggml_map_custom1 +// ggml_map_custom1_f32 -struct ggml_tensor * ggml_map_custom1_impl_f32( +static struct ggml_tensor * ggml_map_custom1_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, const ggml_custom1_op_f32_t fun, @@ -7531,19 +7363,13 @@ struct ggml_tensor * ggml_map_custom1_impl_f32( is_node = true; } - struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_scratch_save(ctx); - - struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); - *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_load(ctx); + ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_CUSTOM1; + result->op = GGML_OP_MAP_CUSTOM1_F32; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; - result->src[2] = addr_tensor; return result; } @@ -7562,9 +7388,9 @@ struct ggml_tensor * ggml_map_custom1_inplace_f32( return ggml_map_custom1_impl_f32(ctx, a, fun, true); } -// ggml_map_custom2 +// ggml_map_custom2_f32 -struct ggml_tensor * ggml_map_custom2_impl_f32( +static struct ggml_tensor * ggml_map_custom2_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -7576,20 +7402,14 @@ struct ggml_tensor * ggml_map_custom2_impl_f32( is_node = true; } - struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_scratch_save(ctx); - - struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); - *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_load(ctx); + ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_CUSTOM2; + result->op = GGML_OP_MAP_CUSTOM2_F32; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = addr_tensor; return result; } @@ -7610,9 +7430,9 @@ struct ggml_tensor * ggml_map_custom2_inplace_f32( return ggml_map_custom2_impl_f32(ctx, a, b, fun, true); } -// ggml_map_custom3 +// ggml_map_custom3_f32 -struct ggml_tensor * ggml_map_custom3_impl_f32( +static struct ggml_tensor * ggml_map_custom3_impl_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -7625,21 +7445,15 @@ struct ggml_tensor * ggml_map_custom3_impl_f32( is_node = true; } - struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_scratch_save(ctx); - - struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); - *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - ggml_scratch_load(ctx); + ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - result->op = GGML_OP_MAP_CUSTOM3; + result->op = GGML_OP_MAP_CUSTOM3_F32; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = addr_tensor; - result->src[3] = c; + result->src[2] = c; return result; } @@ -7662,6 +7476,190 @@ struct ggml_tensor * ggml_map_custom3_inplace_f32( return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true); } +// ggml_map_custom1 +struct ggml_map_custom1_op_params { + ggml_custom1_op_t fun; + int n_tasks; + void * userdata; +}; + +static struct ggml_tensor * ggml_map_custom1_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_custom1_op_t fun, + int n_tasks, + void * userdata, + bool inplace) { + GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0); + + bool is_node = false; + + if (!inplace && a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + struct ggml_map_custom1_op_params params = { + /*.fun =*/ fun, + /*.n_tasks =*/ n_tasks, + /*.userdata =*/ userdata + }; + ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); + + result->op = GGML_OP_MAP_CUSTOM1; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + + return result; +} + +struct ggml_tensor * ggml_map_custom1( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_custom1_op_t fun, + int n_tasks, + void * userdata) { + return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false); +} + +struct ggml_tensor * ggml_map_custom1_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_custom1_op_t fun, + int n_tasks, + void * userdata) { + return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true); +} + +// ggml_map_custom2 + +struct ggml_map_custom2_op_params { + ggml_custom2_op_t fun; + int n_tasks; + void * userdata; +}; + +static struct ggml_tensor * ggml_map_custom2_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_custom2_op_t fun, + int n_tasks, + void * userdata, + bool inplace) { + GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + struct ggml_map_custom2_op_params params = { + /*.fun =*/ fun, + /*.n_tasks =*/ n_tasks, + /*.userdata =*/ userdata + }; + ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); + + result->op = GGML_OP_MAP_CUSTOM2; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + result->src[1] = b; + + return result; +} + +struct ggml_tensor * ggml_map_custom2( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_custom2_op_t fun, + int n_tasks, + void * userdata) { + return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false); +} + +struct ggml_tensor * ggml_map_custom2_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_custom2_op_t fun, + int n_tasks, + void * userdata) { + return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true); +} + +// ggml_map_custom3 + +struct ggml_map_custom3_op_params { + ggml_custom3_op_t fun; + int n_tasks; + void * userdata; +}; + +static struct ggml_tensor * ggml_map_custom3_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + const ggml_custom3_op_t fun, + int n_tasks, + void * userdata, + bool inplace) { + GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad || c->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + struct ggml_map_custom3_op_params params = { + /*.fun =*/ fun, + /*.n_tasks =*/ n_tasks, + /*.userdata =*/ userdata + }; + ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); + + result->op = GGML_OP_MAP_CUSTOM3; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + result->src[1] = b; + result->src[2] = c; + + return result; +} + +struct ggml_tensor * ggml_map_custom3( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + const ggml_custom3_op_t fun, + int n_tasks, + void * userdata) { + return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false); +} + +struct ggml_tensor * ggml_map_custom3_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + const ggml_custom3_op_t fun, + int n_tasks, + void * userdata) { + return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true); +} + + + // ggml_cross_entropy_loss struct ggml_tensor * ggml_cross_entropy_loss( @@ -8867,21 +8865,17 @@ static void ggml_compute_forward_acc_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { GGML_ASSERT(ggml_are_same_shape(src0, dst)); GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - GGML_ASSERT(opt0->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(opt0) == 5); - // view src0 and dst with these strides and data offset inbytes during acc // nb0 is implicitely element_size because src0 and dst are contiguous - size_t nb1 = ((int32_t *) opt0->data)[0]; - size_t nb2 = ((int32_t *) opt0->data)[1]; - size_t nb3 = ((int32_t *) opt0->data)[2]; - size_t offset = ((int32_t *) opt0->data)[3]; - bool inplace = (bool) ((int32_t *) opt0->data)[4]; + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; if (!inplace && (params->type == GGML_TASK_INIT)) { // memcpy needs to be synchronized across threads to avoid race conditions. @@ -8950,13 +8944,12 @@ static void ggml_compute_forward_acc( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_acc_f32(params, src0, src1, opt0, dst); + ggml_compute_forward_acc_f32(params, src0, src1, dst); } break; case GGML_TYPE_F16: case GGML_TYPE_Q4_0: @@ -9388,7 +9381,7 @@ static void ggml_compute_forward_sum_f32( for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { for (int64_t i01 = 0; i01 < ne01; i01++) { - ggml_vec_sum_ggf(ne00, + ggml_vec_sum_f32_ggf(ne00, &row_sum, (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); sum += row_sum; @@ -9398,6 +9391,38 @@ static void ggml_compute_forward_sum_f32( ((float *) dst->data)[0] = sum; } +static void ggml_compute_forward_sum_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_is_scalar(dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + assert(src0->nb[0] == sizeof(ggml_fp16_t)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne); + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb); + + float sum = 0; + float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f16_ggf(ne00, + &row_sum, + (ggml_fp16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); + sum += row_sum; + } + } + } + ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum); +} + static void ggml_compute_forward_sum( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -9407,6 +9432,10 @@ static void ggml_compute_forward_sum( { ggml_compute_forward_sum_f32(params, src0, dst); } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_sum_f16(params, src0, dst); + } break; default: { GGML_ASSERT(false); @@ -9439,8 +9468,8 @@ static void ggml_compute_forward_sum_rows_f32( for (int64_t i3 = 0; i3 < ne03; i3++) { for (int64_t i2 = 0; i2 < ne02; i2++) { for (int64_t i1 = 0; i1 < ne01; i1++) { - float* src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); - float* dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); + float * src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); + float * dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); float row_sum = 0; ggml_vec_sum_f32(ne00, &row_sum, src_row); dst_row[0] = row_sum; @@ -10002,8 +10031,8 @@ static void ggml_compute_forward_gelu_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { @@ -10061,8 +10090,8 @@ static void ggml_compute_forward_gelu_quick_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { @@ -10120,8 +10149,8 @@ static void ggml_compute_forward_silu_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { @@ -10173,7 +10202,6 @@ static void ggml_compute_forward_silu( } } - // ggml_compute_forward_silu_back static void ggml_compute_forward_silu_back_f32( @@ -10181,9 +10209,9 @@ static void ggml_compute_forward_silu_back_f32( const struct ggml_tensor * src0, const struct ggml_tensor * grad, struct ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(grad)); - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(grad)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); + GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); GGML_ASSERT(ggml_are_same_shape(src0, grad)); @@ -10323,7 +10351,8 @@ static void ggml_compute_forward_rms_norm_f32( GGML_TENSOR_UNARY_OP_LOCALS; - const float eps = 1e-6f; // TODO: make this a parameter + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); // TODO: optimize for (int64_t i03 = 0; i03 < ne03; i03++) { @@ -10543,7 +10572,6 @@ static void ggml_compute_forward_rms_norm_back( } } - // ggml_compute_forward_mul_mat #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) @@ -10587,17 +10615,19 @@ static void ggml_compute_forward_mul_mat( const int ith = params->ith; const int nth = params->nth; - GGML_ASSERT(ne02 == ne12); - GGML_ASSERT(ne03 == ne13); - GGML_ASSERT(ne2 == ne12); - GGML_ASSERT(ne3 == ne13); - const enum ggml_type type = src0->type; + const bool src1_cont = ggml_is_contiguous(src1); + ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot; enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float; + GGML_ASSERT(ne0 == ne01); + GGML_ASSERT(ne1 == ne11); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == GGML_TYPE_SIZE[type]); GGML_ASSERT(nb10 == sizeof(float)); @@ -10608,16 +10638,16 @@ static void ggml_compute_forward_mul_mat( GGML_ASSERT(nb1 <= nb2); GGML_ASSERT(nb2 <= nb3); - GGML_ASSERT(ne0 == ne01); - GGML_ASSERT(ne1 == ne11); - GGML_ASSERT(ne2 == ne02); - GGML_ASSERT(ne3 == ne03); - // nb01 >= nb00 - src0 is not transposed // compute by src0 rows #if defined(GGML_USE_CLBLAST) if (ggml_cl_can_mul_mat(src0, src1, dst)) { + // TODO: handle case when src0 is broadcast-able into src1 across 2nd,3rd dimension + // ref: https://github.com/ggerganov/ggml/pull/224 + GGML_ASSERT(ne02 == ne12); + GGML_ASSERT(ne03 == ne13); + if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) { ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize); } @@ -10627,6 +10657,11 @@ static void ggml_compute_forward_mul_mat( #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { + // TODO: handle case when src0 is broadcast-able into src1 across 2nd,3rd dimension + // ref: https://github.com/ggerganov/ggml/pull/224 + GGML_ASSERT(ne02 == ne12); + GGML_ASSERT(ne03 == ne13); + if (params->ith != 0) { return; } @@ -10647,7 +10682,7 @@ static void ggml_compute_forward_mul_mat( float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); if (type != GGML_TYPE_F32) { - float * const wdata = params->wdata; + float * const wdata = params->wdata; ggml_to_float_t const to_float = type_traits[type].to_float; size_t id = 0; @@ -10696,60 +10731,95 @@ static void ggml_compute_forward_mul_mat( return; } - // parallelize by src0 rows using ggml_vec_dot_q + const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type]; - // total rows in src0 - const int nr = ne01*ne02*ne03; + const int64_t nr0 = ne01; // src0 rows + const int64_t nr1 = ne11*ne12*ne13; // src1 rows - // rows per thread - const int dr = (nr + nth - 1)/nth; + //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1); - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); + // distribute the thread work across the inner or outer loop based on which one is larger - void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; - const size_t row_size = ne00*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type]; + const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows + const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows - for (int ir = ir0; ir < ir1; ++ir) { - // src0 indices - const int i03 = ir/(ne02*ne01); - const int i02 = (ir - i03*ne02*ne01)/ne01; - const int i01 = (ir - i03*ne02*ne01 - i02*ne01); + const int64_t ith0 = ith % nth0; + const int64_t ith1 = ith / nth0; - const int i13 = i03; - const int i12 = i02; + const int64_t dr0 = (nr0 + nth0 - 1)/nth0; + const int64_t dr1 = (nr1 + nth1 - 1)/nth1; - const int i0 = i01; - const int i2 = i02; - const int i3 = i03; + const int64_t ir010 = dr0*ith0; + const int64_t ir011 = MIN(ir010 + dr0, nr0); - void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03)); - char * src1_col = ((char *) wdata + ( (0 + i12*ne11 + i13*ne12*ne11)*row_size)); + const int64_t ir110 = dr1*ith1; + const int64_t ir111 = MIN(ir110 + dr1, nr1); - float * dst_col = (float *) ((char *) dst->data + (i0*nb0 + 0*nb1 + i2*nb2 + i3*nb3)); + //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111); - for (int64_t ic = 0; ic < ne11; ++ic) { - vec_dot(ne00, &dst_col[ic*ne0], src0_row, (void *) (src1_col + ic*row_size)); - } + // threads with no work simply yield (not sure if it helps) + if (ir010 >= ir011 || ir110 >= ir111) { + sched_yield(); + return; } - //int64_t t1 = ggml_time_us(); - //static int64_t acc = 0; - //acc += t1 - t0; - //if (t1 - t0 > 10) { - // printf("\n"); - // printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03); - // printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03); - // printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13); + assert(ne12 % ne02 == 0); + assert(ne13 % ne03 == 0); - // printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc); - //} -} + // broadcast factors + const int64_t r2 = ne12/ne02; + const int64_t r3 = ne13/ne03; + // block-tiling attempt + const int64_t blck_0 = 16; + const int64_t blck_1 = 16; -// ggml_compute_forward_out_prod + // attempt to reduce false-sharing (does not seem to make a difference) + float tmp[16]; + + for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) { + for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) { + for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) { + const int64_t i13 = (ir1/(ne12*ne11)); + const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11; + const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11); + + // broadcast src0 into src1 + const int64_t i03 = i13/r3; + const int64_t i02 = i12/r2; + + const int64_t i1 = i11; + const int64_t i2 = i12; + const int64_t i3 = i13; + + const char * src0_row = (const char *) src0->data + (0 + i02*nb02 + i03*nb03); + // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides + // if it is, then we have either copied the data to params->wdata and made it contiguous or we are using + // the original src1 data pointer, so we should index using the indices directly + // TODO: this is a bit of a hack, we should probably have a better way to handle this + const char * src1_col = (const char *) wdata + + (src1_cont || src1->type != vec_dot_type + ? (i11 + i12*ne11 + i13*ne12*ne11)*row_size + : (i11*nb11 + i12*nb12 + i13*nb13)); + + float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3)); + + //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) { + // vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col); + //} + + for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) { + vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col); + } + memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float)); + } + } + } +} + +// ggml_compute_forward_out_prod static void ggml_compute_forward_out_prod_f32( const struct ggml_compute_params * params, @@ -10959,21 +11029,17 @@ static void ggml_compute_forward_set_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { GGML_ASSERT(ggml_are_same_shape(src0, dst)); GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - GGML_ASSERT(opt0->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(opt0) == 5); - // view src0 and dst with these strides and data offset inbytes during set // nb0 is implicitely element_size because src0 and dst are contiguous - size_t nb1 = ((int32_t *) opt0->data)[0]; - size_t nb2 = ((int32_t *) opt0->data)[1]; - size_t nb3 = ((int32_t *) opt0->data)[2]; - size_t offset = ((int32_t *) opt0->data)[3]; - bool inplace = (bool) ((int32_t *) opt0->data)[4]; + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; if (!inplace && (params->type == GGML_TASK_INIT)) { // memcpy needs to be synchronized across threads to avoid race conditions. @@ -11033,13 +11099,12 @@ static void ggml_compute_forward_set( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_set_f32(params, src0, src1, opt0, dst); + ggml_compute_forward_set_f32(params, src0, src1, dst); } break; case GGML_TYPE_F16: case GGML_TYPE_Q4_0: @@ -11435,17 +11500,14 @@ static void ggml_compute_forward_diag( static void ggml_compute_forward_diag_mask_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst, const float value) { - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(src1) == 2); const int ith = params->ith; const int nth = params->nth; - const int n_past = ((int32_t *) src1->data)[0]; - const bool inplace = (bool)((int32_t *) src1->data)[1]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const bool inplace = (bool)((int32_t *) dst->op_params)[1]; GGML_ASSERT(n_past >= 0); @@ -11488,12 +11550,11 @@ static void ggml_compute_forward_diag_mask_f32( static void ggml_compute_forward_diag_mask_inf( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_diag_mask_f32(params, src0, src1, dst, -INFINITY); + ggml_compute_forward_diag_mask_f32(params, src0, dst, -INFINITY); } break; default: { @@ -11505,12 +11566,11 @@ static void ggml_compute_forward_diag_mask_inf( static void ggml_compute_forward_diag_mask_zero( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_diag_mask_f32(params, src0, src1, dst, 0); + ggml_compute_forward_diag_mask_f32(params, src0, dst, 0); } break; default: { @@ -11708,20 +11768,17 @@ static void ggml_compute_forward_soft_max_back( static void ggml_compute_forward_alibi_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { assert(params->ith == 0); - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(src1) == 3); - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const int n_past = ((int32_t *) src1->data)[0]; - const int n_head = ((int32_t *) src1->data)[1]; - const float max_bias = ((float *) src1->data)[2]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_head = ((int32_t *) dst->op_params)[1]; + float max_bias; + memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); assert(n_past >= 0); @@ -11774,20 +11831,17 @@ static void ggml_compute_forward_alibi_f32( static void ggml_compute_forward_alibi_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { assert(params->ith == 0); - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(src1) == 3); - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const int n_past = ((int32_t *) src1->data)[0]; - const int n_head = ((int32_t *) src1->data)[1]; - const float max_bias = ((float *) src1->data)[2]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_head = ((int32_t *) dst->op_params)[1]; + float max_bias; + memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); assert(n_past >= 0); @@ -11840,16 +11894,15 @@ static void ggml_compute_forward_alibi_f16( static void ggml_compute_forward_alibi( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_alibi_f16(params, src0, src1, dst); + ggml_compute_forward_alibi_f16(params, src0, dst); } break; case GGML_TYPE_F32: { - ggml_compute_forward_alibi_f32(params, src0, src1, dst); + ggml_compute_forward_alibi_f32(params, src0, dst); } break; case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: @@ -11879,19 +11932,17 @@ static void ggml_compute_forward_alibi( static void ggml_compute_forward_clamp_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { assert(params->ith == 0); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(ggml_nelements(src1) == 2); - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const float min = ((float *) src1->data)[0]; - const float max = ((float *) src1->data)[1]; + float min; + float max; + memcpy(&min, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); const int ith = params->ith; const int nth = params->nth; @@ -11921,12 +11972,11 @@ static void ggml_compute_forward_clamp_f32( static void ggml_compute_forward_clamp( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_clamp_f32(params, src0, src1, dst); + ggml_compute_forward_clamp_f32(params, src0, dst); } break; case GGML_TYPE_F16: case GGML_TYPE_Q4_0: @@ -11956,19 +12006,21 @@ static void ggml_compute_forward_clamp( static void ggml_compute_forward_rope_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(src1) == 4); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const int n_past = ((int32_t *) src1->data)[0]; - const int n_dims = ((int32_t *) src1->data)[1]; - const int mode = ((int32_t *) src1->data)[2]; - const int n_ctx = ((int32_t *) src1->data)[3]; + float freq_base; + float freq_scale; + + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); assert(n_past >= 0); @@ -11997,7 +12049,7 @@ static void ggml_compute_forward_rope_f32( // row index used to determine which thread to use int ir = 0; - const float theta_scale = powf(10000.0, -2.0f/n_dims); + const float theta_scale = powf(freq_base, -2.0f/n_dims); const bool is_neox = mode & 2; const bool is_glm = mode & 4; @@ -12009,7 +12061,7 @@ static void ggml_compute_forward_rope_f32( if (ir++ < ir0) continue; if (ir > ir1) break; - float theta = (float)p; + float theta = freq_scale * (float)p; if (is_glm) { theta = MIN(p, n_ctx - 2); @@ -12083,19 +12135,21 @@ static void ggml_compute_forward_rope_f32( static void ggml_compute_forward_rope_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_nelements(src1) == 4); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const int n_past = ((int32_t *) src1->data)[0]; - const int n_dims = ((int32_t *) src1->data)[1]; - const int mode = ((int32_t *) src1->data)[2]; - const int n_ctx = ((int32_t *) src1->data)[3]; + float freq_base; + float freq_scale; + + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); assert(n_past >= 0); @@ -12124,7 +12178,7 @@ static void ggml_compute_forward_rope_f16( // row index used to determine which thread to use int ir = 0; - const float theta_scale = powf(10000.0, -2.0f/n_dims); + const float theta_scale = powf(freq_base, -2.0f/n_dims); const bool is_neox = mode & 2; const bool is_glm = mode & 4; @@ -12136,7 +12190,7 @@ static void ggml_compute_forward_rope_f16( if (ir++ < ir0) continue; if (ir > ir1) break; - float theta = (float)p; + float theta = freq_scale * (float)p; if (is_glm) { theta = MIN(p, n_ctx - 2); @@ -12197,7 +12251,7 @@ static void ggml_compute_forward_rope_f16( const float x0 = GGML_FP16_TO_FP32(src[0]); const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } @@ -12210,16 +12264,15 @@ static void ggml_compute_forward_rope_f16( static void ggml_compute_forward_rope( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f16(params, src0, dst); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_f32(params, src0, src1, dst); + ggml_compute_forward_rope_f32(params, src0, dst); } break; default: { @@ -12233,10 +12286,7 @@ static void ggml_compute_forward_rope( static void ggml_compute_forward_rope_back_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 3); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; @@ -12246,9 +12296,9 @@ static void ggml_compute_forward_rope_back_f32( // dx = rope_back(dy, src1) // src0 is dy, src1 contains options - const int n_past = ((int32_t *) src1->data)[0]; - const int n_dims = ((int32_t *) src1->data)[1]; - const int mode = ((int32_t *) src1->data)[2]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; assert(n_past >= 0); @@ -12332,10 +12382,7 @@ static void ggml_compute_forward_rope_back_f32( static void ggml_compute_forward_rope_back_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 3); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; @@ -12345,9 +12392,9 @@ static void ggml_compute_forward_rope_back_f16( // dx = rope_back(dy, src1) // src0 is dy, src1 contains options - const int n_past = ((int32_t *) src1->data)[0]; - const int n_dims = ((int32_t *) src1->data)[1]; - const int mode = ((int32_t *) src1->data)[2]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; assert(n_past >= 0); @@ -12431,16 +12478,15 @@ static void ggml_compute_forward_rope_back_f16( static void ggml_compute_forward_rope_back( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * src1, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_back_f16(params, src0, src1, dst); + ggml_compute_forward_rope_back_f16(params, src0, dst); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_back_f32(params, src0, src1, dst); + ggml_compute_forward_rope_back_f32(params, src0, dst); } break; default: { @@ -12637,7 +12683,7 @@ static void ggml_compute_forward_conv_1d_s1_ph( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: { @@ -12840,7 +12886,7 @@ static void ggml_compute_forward_conv_1d_s2_ph( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: { @@ -12860,14 +12906,13 @@ static void ggml_compute_forward_conv_1d_s2_ph( // ggml_compute_forward_conv_1d static void ggml_compute_forward_conv_1d( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - const struct ggml_tensor * opt0, - struct ggml_tensor * dst) { - const int32_t s0 = ((const int32_t*)(opt0->data))[0]; - const int32_t p0 = ((const int32_t*)(opt0->data))[1]; - const int32_t d0 = ((const int32_t*)(opt0->data))[2]; + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; GGML_ASSERT(d0 == 1); // dilation not supported GGML_ASSERT(p0 == src0->ne[0]/2); // only half padding supported if (s0 == 1) { @@ -12879,9 +12924,9 @@ static void ggml_compute_forward_conv_1d( }; } -// ggml_compute_forward_conv_2d_sk_p0 +// ggml_compute_forward_conv_2d -static void ggml_compute_forward_conv_2d_sk_p0_f16_f32( +static void ggml_compute_forward_conv_2d_f16_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12904,11 +12949,17 @@ static void ggml_compute_forward_conv_2d_sk_p0_f16_f32( // size of the convolution row - the kernel size unrolled across all channels const int ew0 = nk0*nk1*ne02; + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { - // TODO: fix this memset (wsize is overestimated) memset(params->wdata, 0, params->wsize); // prepare source data (src1) @@ -12923,8 +12974,13 @@ static void ggml_compute_forward_conv_2d_sk_p0_f16_f32( for (int i0 = 0; i0 < ne0; i0++) { for (int ik1 = 0; ik1 < nk1; ik1++) { for (int ik0 = 0; ik0 < nk0; ik0++) { - dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] = - GGML_FP32_TO_FP16(src[(i1*nk1 + ik1)*ne10 + (i0*nk0 + ik0)]); + const int idx0 = i0*s0 + ik0*d0 - p0; + const int idx1 = i1*s1 + ik1*d1 - p1; + + if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) { + dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] = + GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]); + } } } } @@ -12951,32 +13007,34 @@ static void ggml_compute_forward_conv_2d_sk_p0_f16_f32( ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - for (int i2 = ip0; i2 < ip1; i2++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2); - - for (int i1 = 0; i1 < ne1; ++i1) { - for (int i0 = 0; i0 < ne0; ++i0) { - ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0, - (ggml_fp16_t *) ((char *) src0->data + i2*nb03), - (ggml_fp16_t *) wdata + (i1*ne0 + i0)*ew0); + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ip0; i2 < ip1; i2++) { + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2); + + for (int i1 = 0; i1 < ne1; ++i1) { + for (int i0 = 0; i0 < ne0; ++i0) { + ggml_vec_dot_f16(ew0, dst_data + i1*ne0 + i0, + (ggml_fp16_t *) ((char *) src0->data + i2*nb03), + (ggml_fp16_t *) wdata + i3*nb3 + (i1*ne0 + i0)*ew0); + } } } } } -static void ggml_compute_forward_conv_2d_sk_p0( +static void ggml_compute_forward_conv_2d( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_conv_2d_sk_p0_f16_f32(params, src0, src1, dst); + ggml_compute_forward_conv_2d_f16_f32(params, src0, src1, dst); } break; case GGML_TYPE_F32: { - //ggml_compute_forward_conv_2d_sk_p0_f32(params, src0, src1, dst); + //ggml_compute_forward_conv_2d_f32(params, src0, src1, dst); GGML_ASSERT(false); } break; default: @@ -12986,31 +13044,162 @@ static void ggml_compute_forward_conv_2d_sk_p0( } } -// ggml_compute_forward_conv_2d +// ggml_compute_forward_pool_1d_sk_p0 -static void ggml_compute_forward_conv_2d( - const struct ggml_compute_params* params, - const struct ggml_tensor* src0, - const struct ggml_tensor* src1, - const struct ggml_tensor* opt0, - struct ggml_tensor* dst) { - const int32_t s0 = ((const int32_t*)(opt0->data))[0]; - const int32_t s1 = ((const int32_t*)(opt0->data))[1]; - const int32_t p0 = ((const int32_t*)(opt0->data))[2]; - const int32_t p1 = ((const int32_t*)(opt0->data))[3]; - const int32_t d0 = ((const int32_t*)(opt0->data))[4]; - const int32_t d1 = ((const int32_t*)(opt0->data))[5]; - GGML_ASSERT(d0 == 1); // dilation not supported - GGML_ASSERT(d1 == 1); +static void ggml_compute_forward_pool_1d_sk_p0( + const struct ggml_compute_params * params, + const enum ggml_op_pool op, + const struct ggml_tensor * src, + const int k, + struct ggml_tensor * dst) { + assert(src->type == GGML_TYPE_F32); + assert(params->ith == 0); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const char * cdata = (const char *)src->data; + const char * const data_end = cdata + ggml_nbytes(src); + float * drow = (float *)dst->data; + + const int64_t rs = dst->ne[0]; + + while (cdata < data_end) { + const float * const srow = (const float *)cdata; + + int j = 0; + + for (int64_t i = 0; i < rs; ++i) { + switch (op) { + case GGML_OP_POOL_AVG: drow[i] = 0; break; + case GGML_OP_POOL_MAX: drow[i] = -FLT_MAX; break; + case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; + } + for (int ki = 0; ki < k; ++ki) { + switch (op) { + case GGML_OP_POOL_AVG: drow[i] += srow[j]; break; + case GGML_OP_POOL_MAX: if (srow[j] > drow[i]) drow[i] = srow[j]; break; + case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; + } + ++j; + } + switch (op) { + case GGML_OP_POOL_AVG: drow[i] /= k; break; + case GGML_OP_POOL_MAX: break; + case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; + } + } + + cdata += src->nb[1]; + drow += rs; + } +} + +// ggml_compute_forward_pool_1d + +static void ggml_compute_forward_pool_1d( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int s0 = opts[2]; + const int p0 = opts[3]; GGML_ASSERT(p0 == 0); // padding not supported - GGML_ASSERT(p1 == 0); + GGML_ASSERT(k0 == s0); // only s = k supported + + ggml_compute_forward_pool_1d_sk_p0(params, op, src0, k0, dst); +} + +// ggml_compute_forward_pool_2d_sk_p0 - if (s0 == src0->ne[0] && s1 == src0->ne[1]) { - ggml_compute_forward_conv_2d_sk_p0(params, src0, src1, dst); +static void ggml_compute_forward_pool_2d_sk_p0( + const struct ggml_compute_params * params, + const enum ggml_op_pool op, + const struct ggml_tensor * src, + const int k0, + const int k1, + struct ggml_tensor * dst) { + assert(src->type == GGML_TYPE_F32); + assert(params->ith == 0); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; } - else { - GGML_ASSERT(false); // only stride equal to kernel size is supported - }; + + const char * cdata = (const char*)src->data; + const char * const data_end = cdata + ggml_nbytes(src); + + const int64_t px = dst->ne[0]; + const int64_t py = dst->ne[1]; + const int64_t pa = px * py; + + float * dplane = (float *)dst->data; + + const int ka = k0 * k1; + + while (cdata < data_end) { + for (int oy = 0; oy < py; ++oy) { + float * const drow = dplane + oy * px; + for (int ox = 0; ox < px; ++ox) { + float * const out = drow + ox; + switch (op) { + case GGML_OP_POOL_AVG: *out = 0; break; + case GGML_OP_POOL_MAX: *out = -FLT_MAX; break; + case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; + } + + const int ix = ox * k0; + const int iy = oy * k1; + + for (int ky = 0; ky < k1; ++ky) { + const float * const srow = (const float *)(cdata + src->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + switch (op) { + case GGML_OP_POOL_AVG: *out += srow[j]; break; + case GGML_OP_POOL_MAX: if (srow[j] > *out) *out = srow[j]; break; + case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; + } + } + } + switch (op) { + case GGML_OP_POOL_AVG: *out /= ka; break; + case GGML_OP_POOL_MAX: break; + case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; + } + } + } + + cdata += src->nb[2]; + dplane += pa; + } +} + +// ggml_compute_forward_pool_2d + +static void ggml_compute_forward_pool_2d( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + GGML_ASSERT(p0 == 0); + GGML_ASSERT(p1 == 0); // padding not supported + GGML_ASSERT(k0 == s0); + GGML_ASSERT(k1 == s1); // only s = k supported + + ggml_compute_forward_pool_2d_sk_p0(params, op, src0, k0, k1, dst); } @@ -13022,7 +13211,7 @@ static void ggml_compute_forward_flash_attn_f32( const struct ggml_tensor * k, const struct ggml_tensor * v, const bool masked, - struct ggml_tensor * dst) { + struct ggml_tensor * dst) { int64_t t0 = ggml_perf_time_us(); UNUSED(t0); @@ -13200,7 +13389,7 @@ static void ggml_compute_forward_flash_attn_f16( const struct ggml_tensor * k, const struct ggml_tensor * v, const bool masked, - struct ggml_tensor * dst) { + struct ggml_tensor * dst) { int64_t t0 = ggml_perf_time_us(); UNUSED(t0); @@ -13965,7 +14154,6 @@ static void ggml_compute_forward_flash_attn_back( static void ggml_compute_forward_win_part_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; @@ -13974,9 +14162,9 @@ static void ggml_compute_forward_win_part_f32( GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne); GGML_TENSOR_LOCALS(int64_t, ne, dst, ne); - const int32_t nep0 = ((const int32_t *)(opt0->data))[0]; - const int32_t nep1 = ((const int32_t *)(opt0->data))[1]; - const int32_t w = ((const int32_t *)(opt0->data))[2]; + const int32_t nep0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t nep1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t w = ((const int32_t *)(dst->op_params))[2]; assert(ne00 == ne0); assert(ne3 == nep0*nep1); @@ -14010,12 +14198,11 @@ static void ggml_compute_forward_win_part_f32( static void ggml_compute_forward_win_part( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_win_part_f32(params, src0, opt0, dst); + ggml_compute_forward_win_part_f32(params, src0, dst); } break; default: { @@ -14029,7 +14216,6 @@ static void ggml_compute_forward_win_part( static void ggml_compute_forward_win_unpart_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; @@ -14038,7 +14224,7 @@ static void ggml_compute_forward_win_unpart_f32( GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne); GGML_TENSOR_LOCALS(int64_t, ne, dst, ne); - const int32_t w = ((const int32_t *)(opt0->data))[0]; + const int32_t w = ((const int32_t *)(dst->op_params))[0]; // padding const int px = (w - ne1%w)%w; @@ -14072,12 +14258,67 @@ static void ggml_compute_forward_win_unpart_f32( static void ggml_compute_forward_win_unpart( const struct ggml_compute_params * params, const struct ggml_tensor * src0, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_win_unpart_f32(params, src0, opt0, dst); + ggml_compute_forward_win_unpart_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +//gmml_compute_forward_unary + +static void ggml_compute_forward_unary( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + const enum ggml_unary_op op = ggml_get_unary_op(dst); + + switch (op) { + case GGML_UNARY_OP_ABS: + { + ggml_compute_forward_abs(params, src0, dst); + } break; + case GGML_UNARY_OP_SGN: + { + ggml_compute_forward_sgn(params, src0, dst); + } break; + case GGML_UNARY_OP_NEG: + { + ggml_compute_forward_neg(params, src0, dst); + } break; + case GGML_UNARY_OP_STEP: + { + ggml_compute_forward_step(params, src0, dst); + } break; + case GGML_UNARY_OP_TANH: + { + ggml_compute_forward_tanh(params, src0, dst); + } break; + case GGML_UNARY_OP_ELU: + { + ggml_compute_forward_elu(params, src0, dst); + } break; + case GGML_UNARY_OP_RELU: + { + ggml_compute_forward_relu(params, src0, dst); + } break; + case GGML_UNARY_OP_GELU: + { + ggml_compute_forward_gelu(params, src0, dst); + } break; + case GGML_UNARY_OP_GELU_QUICK: + { + ggml_compute_forward_gelu_quick(params, src0, dst); + } break; + case GGML_UNARY_OP_SILU: + { + ggml_compute_forward_silu(params, src0, dst); } break; default: { @@ -14195,24 +14436,6 @@ static void ggml_compute_forward_map_custom1_f32( fun(dst, a); } - -static void ggml_compute_forward_map_custom1( - const struct ggml_compute_params * params, - const struct ggml_tensor * a, - struct ggml_tensor * dst, - const ggml_custom1_op_f32_t fun) { - switch (a->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_map_custom1_f32(params, a, dst, fun); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - // ggml_compute_forward_map_custom2 static void ggml_compute_forward_map_custom2_f32( @@ -14231,24 +14454,6 @@ static void ggml_compute_forward_map_custom2_f32( } -static void ggml_compute_forward_map_custom2( - const struct ggml_compute_params * params, - const struct ggml_tensor * a, - const struct ggml_tensor * b, - struct ggml_tensor * dst, - const ggml_custom2_op_f32_t fun) { - switch (a->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_map_custom2_f32(params, a, b, dst, fun); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - // ggml_compute_forward_map_custom3 static void ggml_compute_forward_map_custom3_f32( @@ -14267,24 +14472,52 @@ static void ggml_compute_forward_map_custom3_f32( fun(dst, a, b, c); } +// ggml_compute_forward_map_custom1 + +static void ggml_compute_forward_map_custom1( + const struct ggml_compute_params * params, + const struct ggml_tensor * a, + struct ggml_tensor * dst) { + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) dst->op_params; + + p->fun(dst, a, params->ith, params->nth, p->userdata); +} + +// ggml_compute_forward_map_custom2 + +static void ggml_compute_forward_map_custom2( + const struct ggml_compute_params * params, + const struct ggml_tensor * a, + const struct ggml_tensor * b, + struct ggml_tensor * dst) { + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) dst->op_params; + + p->fun(dst, a, b, params->ith, params->nth, p->userdata); +} + +// ggml_compute_forward_map_custom3 static void ggml_compute_forward_map_custom3( const struct ggml_compute_params * params, const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, - struct ggml_tensor * dst, - const ggml_custom3_op_f32_t fun) { - switch (a->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_map_custom3_f32(params, a, b, c, dst, fun); - } break; - default: - { - GGML_ASSERT(false); - } break; + struct ggml_tensor * dst) { + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; } + + struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) dst->op_params; + + p->fun(dst, a, b, c, params->ith, params->nth, p->userdata); } // ggml_compute_forward_cross_entropy_loss @@ -14596,7 +14829,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } break; case GGML_OP_ACC: { - ggml_compute_forward_acc(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor); + ggml_compute_forward_acc(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_SUB: { @@ -14646,46 +14879,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_repeat_back(params, tensor->src[0], tensor); } break; - case GGML_OP_ABS: - { - ggml_compute_forward_abs(params, tensor->src[0], tensor); - } break; - case GGML_OP_SGN: - { - ggml_compute_forward_sgn(params, tensor->src[0], tensor); - } break; - case GGML_OP_NEG: - { - ggml_compute_forward_neg(params, tensor->src[0], tensor); - } break; - case GGML_OP_STEP: - { - ggml_compute_forward_step(params, tensor->src[0], tensor); - } break; - case GGML_OP_TANH: - { - ggml_compute_forward_tanh(params, tensor->src[0], tensor); - } break; - case GGML_OP_ELU: - { - ggml_compute_forward_elu(params, tensor->src[0], tensor); - } break; - case GGML_OP_RELU: - { - ggml_compute_forward_relu(params, tensor->src[0], tensor); - } break; - case GGML_OP_GELU: - { - ggml_compute_forward_gelu(params, tensor->src[0], tensor); - } break; - case GGML_OP_GELU_QUICK: - { - ggml_compute_forward_gelu_quick(params, tensor->src[0], tensor); - } break; - case GGML_OP_SILU: - { - ggml_compute_forward_silu(params, tensor->src[0], tensor); - } break; case GGML_OP_SILU_BACK: { ggml_compute_forward_silu_back(params, tensor->src[0], tensor->src[1], tensor); @@ -14716,7 +14909,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } break; case GGML_OP_SET: { - ggml_compute_forward_set(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor); + ggml_compute_forward_set(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_CPY: { @@ -14756,11 +14949,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } break; case GGML_OP_DIAG_MASK_INF: { - ggml_compute_forward_diag_mask_inf(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_diag_mask_inf(params, tensor->src[0], tensor); } break; case GGML_OP_DIAG_MASK_ZERO: { - ggml_compute_forward_diag_mask_zero(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_diag_mask_zero(params, tensor->src[0], tensor); } break; case GGML_OP_SOFT_MAX: { @@ -14772,31 +14965,39 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } break; case GGML_OP_ROPE: { - ggml_compute_forward_rope(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_rope(params, tensor->src[0], tensor); } break; case GGML_OP_ROPE_BACK: { - ggml_compute_forward_rope_back(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_rope_back(params, tensor->src[0], tensor); } break; case GGML_OP_ALIBI: { - ggml_compute_forward_alibi(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_alibi(params, tensor->src[0], tensor); } break; case GGML_OP_CLAMP: { - ggml_compute_forward_clamp(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_clamp(params, tensor->src[0], tensor); } break; case GGML_OP_CONV_1D: { - ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor); + ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_CONV_2D: { - ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor); + ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor); + } break; + case GGML_OP_POOL_1D: + { + ggml_compute_forward_pool_1d(params, tensor->src[0], tensor); + } break; + case GGML_OP_POOL_2D: + { + ggml_compute_forward_pool_2d(params, tensor->src[0], tensor); } break; case GGML_OP_FLASH_ATTN: { - const int32_t t = ggml_get_i32_1d(tensor->src[3], 0); + const int32_t t = ggml_get_op_params_i32(tensor, 0); GGML_ASSERT(t == 0 || t == 1); const bool masked = t != 0; ggml_compute_forward_flash_attn(params, tensor->src[0], tensor->src[1], tensor->src[2], masked, tensor); @@ -14807,47 +15008,71 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } break; case GGML_OP_FLASH_ATTN_BACK: { - int32_t t = ggml_get_i32_1d(tensor->src[4], 0); + int32_t t = ggml_get_op_params_i32(tensor, 0); GGML_ASSERT(t == 0 || t == 1); bool masked = t != 0; ggml_compute_forward_flash_attn_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], masked, tensor); } break; case GGML_OP_WIN_PART: { - ggml_compute_forward_win_part(params, tensor->src[0], tensor->src[2], tensor); + ggml_compute_forward_win_part(params, tensor->src[0], tensor); } break; case GGML_OP_WIN_UNPART: { - ggml_compute_forward_win_unpart(params, tensor->src[0], tensor->src[2], tensor); + ggml_compute_forward_win_unpart(params, tensor->src[0], tensor); + } break; + case GGML_OP_UNARY: + { + ggml_compute_forward_unary(params, tensor->src[0], tensor); } break; case GGML_OP_MAP_UNARY: { - const ggml_unary_op_f32_t fun = *((ggml_unary_op_f32_t *)tensor->src[2]->data); + ggml_unary_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); ggml_compute_forward_map_unary(params, tensor->src[0], tensor, fun); } break; case GGML_OP_MAP_BINARY: { - const ggml_binary_op_f32_t fun = *((ggml_binary_op_f32_t *)tensor->src[2]->data); + ggml_binary_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun); } break; + case GGML_OP_MAP_CUSTOM1_F32: + { + ggml_custom1_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_custom1_f32(params, tensor->src[0], tensor, fun); + } + break; + case GGML_OP_MAP_CUSTOM2_F32: + { + ggml_custom2_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_custom2_f32(params, tensor->src[0], tensor->src[1], tensor, fun); + } + break; + case GGML_OP_MAP_CUSTOM3_F32: + { + ggml_custom3_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_custom3_f32(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun); + } + break; case GGML_OP_MAP_CUSTOM1: { - const ggml_custom1_op_f32_t fun = *((ggml_custom1_op_f32_t *)tensor->src[2]->data); - ggml_compute_forward_map_custom1(params, tensor->src[0], tensor, fun); + ggml_compute_forward_map_custom1(params, tensor->src[0], tensor); } break; case GGML_OP_MAP_CUSTOM2: { - const ggml_custom2_op_f32_t fun = *((ggml_custom2_op_f32_t *)tensor->src[2]->data); - ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor, fun); + ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_MAP_CUSTOM3: { - const ggml_custom3_op_f32_t fun = *((ggml_custom3_op_f32_t *)tensor->src[2]->data); - ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[3], tensor, fun); + ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor); } break; case GGML_OP_CROSS_ENTROPY_LOSS: @@ -14911,12 +15136,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); } if (src1->grad) { - GGML_ASSERT(ggml_nelements(tensor->src[2]) == 5); - GGML_ASSERT(tensor->src[2]->type == GGML_TYPE_I32); - const size_t nb1 = (( int32_t * ) tensor->src[2]->data)[0]; - const size_t nb2 = (( int32_t * ) tensor->src[2]->data)[1]; - const size_t nb3 = (( int32_t * ) tensor->src[2]->data)[2]; - const size_t offset = (( int32_t * ) tensor->src[2]->data)[3]; + const size_t nb1 = ((int32_t *) tensor->op_params)[0]; + const size_t nb2 = ((int32_t *) tensor->op_params)[1]; + const size_t nb3 = ((int32_t *) tensor->op_params)[2]; + const size_t offset = ((int32_t *) tensor->op_params)[3]; struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx, tensor->grad, @@ -15065,73 +15288,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor inplace); } } break; - case GGML_OP_ABS: - { - if (src0->grad) { - src0->grad = - ggml_add_impl(ctx, - src0->grad, - ggml_mul(ctx, - ggml_sgn(ctx, src0), - tensor->grad), - inplace); - } - } break; - case GGML_OP_SGN: - { - if (src0->grad) { - // noop - } - } break; - case GGML_OP_NEG: - { - if (src0->grad) { - src0->grad = ggml_sub_impl(ctx, src0->grad, tensor->grad, inplace); - } - } break; - case GGML_OP_STEP: - { - if (src0->grad) { - // noop - } - } break; - case GGML_OP_TANH: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_ELU: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_RELU: - { - if (src0->grad) { - src0->grad = ggml_sub_impl(ctx, - src0->grad, - ggml_mul(ctx, - ggml_step(ctx, src0), - tensor->grad), - inplace); - } - } break; - case GGML_OP_GELU: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_GELU_QUICK: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_SILU: - { - // necessary for llama - if (src0->grad) { - src0->grad = ggml_add_impl(ctx, - src0->grad, - ggml_silu_back(ctx, src0, tensor->grad), - inplace); - } - } break; case GGML_OP_SILU_BACK: { GGML_ASSERT(false); // TODO: not implemented @@ -15224,12 +15380,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } break; case GGML_OP_SET: { - GGML_ASSERT(ggml_nelements(tensor->src[2]) == 5); - GGML_ASSERT(tensor->src[2]->type == GGML_TYPE_I32); - const size_t nb1 = (( int32_t * ) tensor->src[2]->data)[0]; - const size_t nb2 = (( int32_t * ) tensor->src[2]->data)[1]; - const size_t nb3 = (( int32_t * ) tensor->src[2]->data)[2]; - const size_t offset = (( int32_t * ) tensor->src[2]->data)[3]; + const size_t nb1 = ((int32_t *) tensor->op_params)[0]; + const size_t nb2 = ((int32_t *) tensor->op_params)[1]; + const size_t nb3 = ((int32_t *) tensor->op_params)[2]; + const size_t offset = ((int32_t *) tensor->op_params)[3]; struct ggml_tensor * tensor_grad_view = NULL; @@ -15306,8 +15460,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor if (src0->grad) { size_t offset; - GGML_ASSERT(sizeof(offset) <= ggml_nbytes(tensor->src[2])); - memcpy(&offset, tensor->src[2]->data, sizeof(offset)); + memcpy(&offset, tensor->op_params, sizeof(offset)); size_t nb1 = tensor->nb[1]; size_t nb2 = tensor->nb[2]; @@ -15334,7 +15487,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { // necessary for llama if (src0->grad) { - int32_t * axes = (int32_t *) tensor->src[2]->data; + int32_t * axes = (int32_t *) tensor->op_params; int axis0 = axes[0] & 0x3; int axis1 = axes[1] & 0x3; int axis2 = axes[2] & 0x3; @@ -15390,33 +15543,23 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { // necessary for llama if (src0->grad) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 2); - const int n_past = ((int32_t *) src1->data)[0]; + const int n_past = ((int32_t *) tensor->op_params)[0]; src0->grad = ggml_add_impl(ctx, src0->grad, ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), inplace); } - if (src1->grad) { - // noop - } } break; case GGML_OP_DIAG_MASK_ZERO: { // necessary for llama if (src0->grad) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 2); - const int n_past = ((int32_t *) src1->data)[0]; + const int n_past = ((int32_t *) tensor->op_params)[0]; src0->grad = ggml_add_impl(ctx, src0->grad, ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), inplace); } - if (src1->grad) { - // noop - } } break; case GGML_OP_SOFT_MAX: { @@ -15437,33 +15580,28 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { // necessary for llama if (src0->grad) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 4); - const int n_past = ((int32_t *) src1->data)[0]; - const int n_dims = ((int32_t *) src1->data)[1]; - const int mode = ((int32_t *) src1->data)[2]; + const int n_past = ((int32_t *) tensor->op_params)[0]; + const int n_dims = ((int32_t *) tensor->op_params)[1]; + const int mode = ((int32_t *) tensor->op_params)[2]; + const int n_ctx = ((int32_t *) tensor->op_params)[3]; src0->grad = ggml_add_impl(ctx, src0->grad, ggml_rope_back(ctx, tensor->grad, n_past, n_dims, - mode), + mode, + n_ctx), inplace); } - if (src1->grad) { - // noop - } } break; case GGML_OP_ROPE_BACK: { if (src0->grad) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 4); - const int n_past = ((int32_t *) src1->data)[0]; - const int n_dims = ((int32_t *) src1->data)[1]; - const int mode = ((int32_t *) src1->data)[2]; - const int n_ctx = ((int32_t *) src1->data)[3]; + const int n_past = ((int32_t *) tensor->op_params)[0]; + const int n_dims = ((int32_t *) tensor->op_params)[1]; + const int mode = ((int32_t *) tensor->op_params)[2]; + const int n_ctx = ((int32_t *) tensor->op_params)[3]; src0->grad = ggml_add_impl(ctx, src0->grad, ggml_rope(ctx, @@ -15474,9 +15612,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor n_ctx), inplace); } - if (src1->grad) { - // noop - } } break; case GGML_OP_ALIBI: { @@ -15494,11 +15629,19 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; + case GGML_OP_POOL_1D: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_OP_POOL_2D: + { + GGML_ASSERT(false); // TODO: not implemented + } break; case GGML_OP_FLASH_ATTN: { struct ggml_tensor * flash_grad = NULL; if (src0->grad || src1->grad || tensor->src[2]->grad) { - int32_t t = ggml_get_i32_1d(tensor->src[3], 0); + int32_t t = ggml_get_op_params_i32(tensor, 0); GGML_ASSERT(t == 0 || t == 1); bool masked = t != 0; flash_grad = @@ -15661,8 +15804,85 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } break; case GGML_OP_WIN_PART: case GGML_OP_WIN_UNPART: + case GGML_OP_UNARY: + { + switch (ggml_get_unary_op(tensor)) { + case GGML_UNARY_OP_ABS: + { + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, + src0->grad, + ggml_mul(ctx, + ggml_sgn(ctx, src0), + tensor->grad), + inplace); + } + } break; + case GGML_UNARY_OP_SGN: + { + if (src0->grad) { + // noop + } + } break; + case GGML_UNARY_OP_NEG: + { + if (src0->grad) { + src0->grad = ggml_sub_impl(ctx, src0->grad, tensor->grad, inplace); + } + } break; + case GGML_UNARY_OP_STEP: + { + if (src0->grad) { + // noop + } + } break; + case GGML_UNARY_OP_TANH: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_UNARY_OP_ELU: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_UNARY_OP_RELU: + { + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_mul(ctx, + ggml_step(ctx, src0), + tensor->grad), + inplace); + } + } break; + case GGML_UNARY_OP_GELU: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_UNARY_OP_GELU_QUICK: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_UNARY_OP_SILU: + { + // necessary for llama + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_silu_back(ctx, src0, tensor->grad), + inplace); + } + } break; + default: + GGML_ASSERT(false); + } + } break; case GGML_OP_MAP_UNARY: case GGML_OP_MAP_BINARY: + case GGML_OP_MAP_CUSTOM1_F32: + case GGML_OP_MAP_CUSTOM2_F32: + case GGML_OP_MAP_CUSTOM3_F32: case GGML_OP_MAP_CUSTOM1: case GGML_OP_MAP_CUSTOM2: case GGML_OP_MAP_CUSTOM3: @@ -15696,6 +15916,34 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } } +static_assert(GGML_GRAPH_HASHTABLE_SIZE > GGML_MAX_NODES * 2, "GGML_GRAPH_HT_SIZE is too small"); + +static size_t hash(void * p) { + return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE; +} + +static bool hash_insert(void * hash_table[], void * p) { + size_t h = hash(p); + + // linear probing + size_t i = h; + while (hash_table[i] != NULL && hash_table[i] != p) { + i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE; + if (i == h) { + // hash table is full + GGML_ASSERT(false); + } + } + + if (hash_table[i] == p) { + return true; + } + + // insert + hash_table[i] = p; + return false; +} + static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) { if (node->grad == NULL) { // this usually happens when we generate intermediate nodes from constants in the backward pass @@ -15706,16 +15954,8 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * } // check if already visited - for (int i = 0; i < cgraph->n_nodes; i++) { - if (cgraph->nodes[i] == node) { - return; - } - } - - for (int i = 0; i < cgraph->n_leafs; i++) { - if (cgraph->leafs[i] == node) { - return; - } + if (hash_insert(cgraph->visited_hash_table, node)) { + return; } for (int i = 0; i < GGML_MAX_SRC; ++i) { @@ -15778,6 +16018,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) { /*.nodes =*/ { NULL }, /*.grads =*/ { NULL }, /*.leafs =*/ { NULL }, + /*.hash_table =*/ { NULL }, /*.perf_runs =*/ 0, /*.perf_cycles =*/ 0, /*.perf_time_us =*/ 0, @@ -15819,13 +16060,42 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg if (node->is_param) { GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node); - ggml_build_forward_impl(&result, node->grad, true); + ggml_build_forward_expand(&result, node->grad); } } return result; } +struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, GGML_GRAPH_SIZE); + struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs); + + *cgraph = (struct ggml_cgraph) { + /*.n_nodes =*/ 0, + /*.n_leafs =*/ 0, + /*.nodes =*/ { NULL }, + /*.grads =*/ { NULL }, + /*.leafs =*/ { NULL }, + /*.hash_table =*/ { NULL }, + /*.perf_runs =*/ 0, + /*.perf_cycles =*/ 0, + /*.perf_time_us =*/ 0, + }; + + return cgraph; +} + +struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor) { + struct ggml_cgraph * cgraph = ggml_new_graph(ctx); + ggml_build_forward_impl(cgraph, tensor, false); + return cgraph; +} + +size_t ggml_graph_overhead(void) { + return GGML_OBJECT_SIZE + GGML_PAD(GGML_GRAPH_SIZE, GGML_MEM_ALIGN); +} + // // thread data // @@ -15891,7 +16161,7 @@ typedef pthread_t ggml_thread_t; // Android's libc implementation "bionic" does not support setting affinity #if defined(__linux__) && !defined(__BIONIC__) -void set_numa_thread_affinity(int thread_n, int n_threads) { +static void set_numa_thread_affinity(int thread_n, int n_threads) { if (!ggml_is_numa()) { return; } @@ -15916,7 +16186,7 @@ void set_numa_thread_affinity(int thread_n, int n_threads) { CPU_FREE(cpus); } -void clear_numa_thread_affinity(void) { +static void clear_numa_thread_affinity(void) { if (!ggml_is_numa()) { return; } @@ -15940,8 +16210,8 @@ void clear_numa_thread_affinity(void) { #else // TODO: Windows etc. // (the linux implementation may also work on BSD, someone should test) -void set_numa_thread_affinity(int thread_n, int n_threads) { UNUSED(thread_n); UNUSED(n_threads); } -void clear_numa_thread_affinity(void) {} +static void set_numa_thread_affinity(int thread_n, int n_threads) { UNUSED(thread_n); UNUSED(n_threads); } +static void clear_numa_thread_affinity(void) {} #endif struct ggml_compute_state_shared { @@ -16011,8 +16281,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { if (GGML_OP_HAS_FINALIZE[node->op]) { params.nth = n_tasks_arr[node_n]; ggml_compute_forward(¶ms, node); - ggml_graph_compute_perf_stats_node(node, state->shared); } + ggml_graph_compute_perf_stats_node(node, state->shared); } // distribute new work or execute it direct if 1T @@ -16042,8 +16312,9 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { if (GGML_OP_HAS_FINALIZE[node->op]) { params.type = GGML_TASK_FINALIZE; ggml_compute_forward(¶ms, node); - ggml_graph_compute_perf_stats_node(node, state->shared); } + + ggml_graph_compute_perf_stats_node(node, state->shared); } else { break; } @@ -16152,21 +16423,34 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { case GGML_OP_ARGMAX: case GGML_OP_REPEAT: case GGML_OP_REPEAT_BACK: - case GGML_OP_ABS: - case GGML_OP_SGN: - case GGML_OP_NEG: - case GGML_OP_STEP: - case GGML_OP_TANH: - case GGML_OP_ELU: - case GGML_OP_RELU: - { + { n_tasks = 1; } break; - case GGML_OP_MUL: - case GGML_OP_GELU: - case GGML_OP_GELU_QUICK: - case GGML_OP_SILU: + + case GGML_OP_UNARY: + { + switch (ggml_get_unary_op(node)) { + case GGML_UNARY_OP_ABS: + case GGML_UNARY_OP_SGN: + case GGML_UNARY_OP_NEG: + case GGML_UNARY_OP_STEP: + case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_ELU: + case GGML_UNARY_OP_RELU: + { + n_tasks = 1; + } break; + + case GGML_UNARY_OP_GELU: + case GGML_UNARY_OP_GELU_QUICK: + case GGML_UNARY_OP_SILU: + { + n_tasks = n_threads; + } break; + } + } break; case GGML_OP_SILU_BACK: + case GGML_OP_MUL: case GGML_OP_NORM: case GGML_OP_RMS_NORM: case GGML_OP_RMS_NORM_BACK: @@ -16231,10 +16515,10 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { case GGML_OP_GET_ROWS: case GGML_OP_GET_ROWS_BACK: case GGML_OP_DIAG: - case GGML_OP_DIAG_MASK_ZERO: { n_tasks = 1; } break; + case GGML_OP_DIAG_MASK_ZERO: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: case GGML_OP_SOFT_MAX_BACK: @@ -16284,8 +16568,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { { n_tasks = n_threads; - GGML_ASSERT(node->src[1]->ne[3] == 1); - const int64_t ne00 = node->src[0]->ne[0]; // W const int64_t ne01 = node->src[0]->ne[1]; // H const int64_t ne02 = node->src[0]->ne[2]; // C @@ -16295,19 +16577,22 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { const int64_t ne11 = node->src[1]->ne[1]; // H const int64_t ne12 = node->src[1]->ne[2]; // C + const int64_t ne0 = node->ne[0]; + const int64_t ne1 = node->ne[1]; + const int64_t ne2 = node->ne[2]; const int64_t nk = ne00*ne01; + const int64_t ew0 = nk * ne02; - UNUSED(ne02); UNUSED(ne03); - UNUSED(nk); + UNUSED(ne2); size_t cur = 0; if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(ggml_fp16_t)*(ne10*ne11*ne12); + node->src[1]->type == GGML_TYPE_F32) { + cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0); } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { + node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)* (ne10*ne11*ne12); } else { GGML_ASSERT(false); @@ -16315,6 +16600,11 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { work_size = MAX(work_size, cur); } break; + case GGML_OP_POOL_1D: + case GGML_OP_POOL_2D: + { + n_tasks = 1; + } break; case GGML_OP_FLASH_ATTN: { n_tasks = n_threads; @@ -16378,11 +16668,38 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { case GGML_OP_WIN_UNPART: case GGML_OP_MAP_UNARY: case GGML_OP_MAP_BINARY: + case GGML_OP_MAP_CUSTOM1_F32: + case GGML_OP_MAP_CUSTOM2_F32: + case GGML_OP_MAP_CUSTOM3_F32: + { + n_tasks = 1; + } break; case GGML_OP_MAP_CUSTOM1: + { + struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; case GGML_OP_MAP_CUSTOM2: + { + struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; case GGML_OP_MAP_CUSTOM3: { - n_tasks = 1; + struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } } break; case GGML_OP_CROSS_ENTROPY_LOSS: { @@ -16521,10 +16838,9 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) { void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads); - struct ggml_tensor * buf = ggml_new_tensor_1d(ctx, GGML_TYPE_I8, cplan.work_size); - GGML_ASSERT(buf); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size); - cplan.work_data = buf->data; + cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; ggml_graph_compute(cgraph, &cplan); } @@ -16579,9 +16895,6 @@ static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char } void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { - //assert(cgraph->work == NULL); - //assert(cgraph->work_size == 0); - uint64_t size_eval = 0; // compute size of intermediate results @@ -16678,7 +16991,8 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { fwrite(&nb, sizeof(uint64_t), 1, fout); } - fwrite(tensor->name, sizeof(char), GGML_MAX_NAME, fout); + fwrite(tensor->name, sizeof(char), GGML_MAX_NAME, fout); + fwrite(tensor->op_params, sizeof(char), GGML_MAX_OP_PARAMS, fout); // dump the data // TODO: pad this to 32 byte boundary @@ -16711,7 +17025,8 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { fwrite(&nb, sizeof(uint64_t), 1, fout); } - fwrite(tensor->name, sizeof(char), GGML_MAX_NAME, fout); + fwrite(tensor->name, sizeof(char), GGML_MAX_NAME, fout); + fwrite(tensor->op_params, sizeof(char), GGML_MAX_OP_PARAMS, fout); // output the op arguments { @@ -16892,7 +17207,8 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** tensor->op = (enum ggml_op) op; - memcpy(tensor->name, ptr, GGML_MAX_NAME); ptr += GGML_MAX_NAME; + memcpy(tensor->name, ptr, GGML_MAX_NAME); ptr += GGML_MAX_NAME; + memcpy(tensor->op_params, ptr, GGML_MAX_OP_PARAMS); ptr += GGML_MAX_OP_PARAMS; tensor->data = (void *) ptr; @@ -16937,7 +17253,8 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** nb[j] = nb_cur; } - const char * ptr_name = ptr; ptr += GGML_MAX_NAME; + const char * ptr_name = ptr; ptr += GGML_MAX_NAME; + const char * ptr_op_params = ptr; ptr += GGML_MAX_OP_PARAMS; const int32_t * ptr_arg_idx = (const int32_t *) ptr; ptr += GGML_MAX_SRC*sizeof(int32_t); @@ -16974,8 +17291,8 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** { tensor = ggml_view_4d(*ctx_eval, args[0], ne[0], ne[1], ne[2], ne[3], 0, 0, 0, 0); - uint64_t offs; - memcpy(&offs, args[2]->data, sizeof(offs)); + size_t offs; + memcpy(&offs, ptr_op_params, sizeof(offs)); tensor->data = ((char *) tensor->data) + offs; } break; @@ -16995,7 +17312,8 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** } break; } - memcpy(tensor->name, ptr_name, GGML_MAX_NAME); + memcpy(tensor->name, ptr_name, GGML_MAX_NAME); + memcpy(tensor->op_params, ptr_op_params, GGML_MAX_OP_PARAMS); for (int j = 0; j < GGML_MAX_DIMS; ++j) { tensor->nb[j] = nb[j]; @@ -17020,9 +17338,6 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { GGML_PRINT("=== GRAPH ===\n"); - GGML_PRINT_DEBUG("n_threads = %d\n", cgraph->n_threads); - GGML_PRINT_DEBUG("total work size = %zu bytes\n", cgraph->work_size); - GGML_PRINT("n_nodes = %d\n", cgraph->n_nodes); for (int i = 0; i < cgraph->n_nodes; i++) { struct ggml_tensor * node = cgraph->nodes[i]; @@ -17032,7 +17347,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n", i, node->ne[0], node->ne[1], node->ne[2], - GGML_OP_NAME[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs, + ggml_op_name(node->op), node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs, (double) node->perf_cycles / (double) ggml_cycles_per_ms(), (double) node->perf_cycles / (double) ggml_cycles_per_ms() / (double) node->perf_runs, (double) node->perf_time_us / 1000.0, @@ -17046,7 +17361,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n", i, node->ne[0], node->ne[1], - GGML_OP_NAME[node->op]); + ggml_op_name(node->op)); } for (int i = 0; i < GGML_OP_COUNT; i++) { @@ -17054,7 +17369,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { continue; } - GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_NAME[i], (double) perf_total_per_op_us[i] / 1000.0); + GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", ggml_op_name(i), (double) perf_total_per_op_us[i] / 1000.0); } GGML_PRINT("========================================\n"); @@ -17148,13 +17463,13 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph } if (node->n_dims == 2) { - fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], GGML_OP_SYMBOL[node->op]); + fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], ggml_op_symbol(node->op)); } else { - fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], GGML_OP_SYMBOL[node->op]); + fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op)); } if (node->grad) { - fprintf(fp, " | <g>%s\"; ]\n", GGML_OP_SYMBOL[node->grad->op]); + fprintf(fp, " | <g>%s\"; ]\n", ggml_op_symbol(node->grad->op)); } else { fprintf(fp, "\"; ]\n"); } |