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

ggml : reduce hash table reset cost (#8698)

Browse source 
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
This commit is contained in:
slaren 2024-07-27 04:41:55 +02:00 • committed by GitHub
parent 01245f5b16
commit 2b1f616b20
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
46 changed files with 851 additions and 754 deletions
Download patch file Download diff file Expand all files Collapse all files

58
ggml/src/ggml-vulkan.cpp
View file

@ -1961,7 +1961,7 @@ void ggml_vk_instance_init() {
// Make sure at least one device exists
if (devices.empty()) {
std::cerr << "ggml_vulkan: Error: No devices found." << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
// Default to using all dedicated GPUs
@ -2459,7 +2459,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
// Buffer is already mapped
if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
std::cerr << "ggml_vulkan: buffer_write_nc_async dst buffer is host_visible. Use synchronous write." << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
// Check if src is pinned memory
vk_buffer buf;
@ -2527,7 +2527,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
staging = ctx->device->sync_staging;
staging_offset = 0;
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -2563,7 +2563,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, s
// Buffer is already mapped
if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
std::cerr << "ggml_vulkan: buffer_write_async dst buffer is host_visible. Use synchronous write." << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
// Check if src is pinned memory
vk_buffer buf = nullptr;
@ -2602,7 +2602,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context * subctx, vk_buffer& dst, s
staging_buffer = dst->device->sync_staging;
staging_offset = 0;
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -2704,7 +2704,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context * subctx, vk_buffer& src, si
staging_buffer = src->device->sync_staging;
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -2913,7 +2913,7 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_
}
std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
@ -3499,7 +3499,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
if (mmp == nullptr) {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
// Not implemented
@ -4078,7 +4078,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
std::cerr << " and " << ggml_type_name(src1->type);
}
std::cerr << " to " << ggml_type_name(dst->type) << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
op_func(ctx, subctx, src0, src1, dst);
@ -4521,7 +4521,7 @@ static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0
} else if (type == GGML_TYPE_F16) {
val = ggml_fp16_to_fp32(*((const ggml_fp16_t *) data + i2*ne1*ne0 + idx1*ne0 + idx0));
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
fprintf(stderr, "% 7.2f ", val);
} else {
@ -4555,7 +4555,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f16->a_s;
shname = "F16_ALIGNED_S";
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
} else if (shader_size == 1) {
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
@ -4571,7 +4571,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f16->a_m;
shname = "F16_ALIGNED_M";
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
} else if (shader_size == 2) {
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
@ -4587,7 +4587,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f16->a_l;
shname = "F16_ALIGNED_L";
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
} else {
GGML_ASSERT(0);
@ -4668,7 +4668,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
} else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
for (size_t i = 0; i < y_ne; i++) {
@ -4679,7 +4679,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
// y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}
@ -4727,14 +4727,14 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
} else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
src0_type = GGML_TYPE_F16;
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
if (std::is_same<float, Y_TYPE>()) {
src1_type = GGML_TYPE_F32;
} else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
src1_type = GGML_TYPE_F16;
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, src0_type, k, m, batch);
@ -4841,7 +4841,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, int i0, int i1
} else if (tensor->type == GGML_TYPE_F16) {
val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
fprintf(stderr, "% 7.2f ", val);
} else {
@ -5391,7 +5391,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
std::cerr << std::endl;
}
GGML_ASSERT(false);
GGML_ABORT("fatal error");
#endif
if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) {
@ -5486,7 +5486,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
break;
default:
std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
return;
}
@ -6498,7 +6498,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
} else if (tensor->type == GGML_TYPE_I32) {
val = *(const int32_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
fprintf(stderr, "% 7.2f ", val);
} else {
@ -6620,7 +6620,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
}
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
@ -6662,7 +6662,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
}
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
@ -6720,7 +6720,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
}
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
@ -6797,7 +6797,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
break;
default:
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
} else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
if (src1 == nullptr) {
@ -6825,7 +6825,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor *
tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone);
} else {
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
@ -6912,7 +6912,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
}
} else {
std::cerr << "Missing debug code for type " << ggml_type_name(tensor->type) << std::endl;
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
if ((std::isnan(correct) != std::isnan(result)) || (std::isinf(correct) != std::isinf(result)) || !buffer_size_fit) {
@ -6935,7 +6935,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
std::cerr << std::endl;
std::vector<const ggml_tensor *> done;
ggml_vk_print_graph_origin(tensor, done);
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) {
first_error[0] = i0;
@ -7006,7 +7006,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor *
std::cerr << std::endl;
std::vector<const ggml_tensor *> done;
ggml_vk_print_graph_origin(tensor, done);
GGML_ASSERT(false);
GGML_ABORT("fatal error");
} else {
std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " avg_err=" << avg_err << std::endl;
}