LLM inference in C/C++
C++ ggml
Implemented vulkan cross_entropy_loss and cross_entropy_loss_back (#27216)
Pranav Uttarkarbf942164697d
diff --git a/docs/ops.md b/docs/ops.mdindex 2c179dd01f..54a3602833 100644--- a/docs/ops.md+++ b/docs/ops.md@@ -35,8 +35,8 @@ Legend:| COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ || COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ || CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |-| CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |-| CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |+| CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |+| CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ || CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ || DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ || DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |diff --git a/docs/ops/Vulkan.csv b/docs/ops/Vulkan.csvindex 59e67e1b20..13e67eb797 100644--- a/docs/ops/Vulkan.csv+++ b/docs/ops/Vulkan.csv@@ -19292,10 +19292,10 @@"Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=4096,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan""Vulkan0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[6,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan""Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"-"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"-"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan"-"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"-"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan"+"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"+"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan"+"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"+"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan""Vulkan0","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan""Vulkan0","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan""Vulkan0","GATED_DELTA_NET","type=f32,head_count=32,head_size=128,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0,K=1","support","1","yes","Vulkan"diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cppindex c1d86aaac5..31923a2ba0 100644--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp@@ -1042,6 +1042,8 @@ struct vk_device_struct {vk_pipeline pipeline_argsort_large_f32[num_argsort_pipelines];vk_pipeline pipeline_topk_f32[num_topk_pipelines];vk_pipeline pipeline_sum_rows_f32;+ vk_pipeline pipeline_cross_entropy_loss_f32, pipeline_cross_entropy_loss_f32_wg512;+ vk_pipeline pipeline_cross_entropy_loss_back_f32, pipeline_cross_entropy_loss_back_f32_wg512;vk_pipeline pipeline_fwht_f32[4];vk_pipeline pipeline_cumsum_f32;vk_pipeline pipeline_cumsum_small_f32;@@ -5758,6 +5760,10 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);+ ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_f32, "cross_entropy_loss_f32", cross_entropy_loss_f32_len, cross_entropy_loss_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);+ ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_f32_wg512, "cross_entropy_loss_f32_wg512", cross_entropy_loss_f32_len, cross_entropy_loss_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { 512 }, 1);+ ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_back_f32, "cross_entropy_loss_back_f32", cross_entropy_loss_back_f32_len, cross_entropy_loss_back_f32_data, "main", 4, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);+ ggml_vk_create_pipeline(device, device->pipeline_cross_entropy_loss_back_f32_wg512, "cross_entropy_loss_back_f32_wg512", cross_entropy_loss_back_f32_len, cross_entropy_loss_back_f32_data, "main", 4, sizeof(vk_op_push_constants), {1, 1, 1}, { 512 }, 1);// Intel Windows driver in range [32.0.101.8509, 32.0.101.8860) will crash when using fwht kernels so we gate that hereconst bool can_use_fwht = device->driver_id != vk::DriverId::eIntelProprietaryWindows ||!ggml_vk_intel_windows_driver_in_range(device->properties.driverVersion, 101, 8509, 101, 8860);@@ -11577,6 +11583,17 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, constreturn ctx->device->pipeline_sum_rows_f32;}return nullptr;+ case GGML_OP_CROSS_ENTROPY_LOSS:+ if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {+ return src0->ne[0] > 1024 ? ctx->device->pipeline_cross_entropy_loss_f32_wg512 : ctx->device->pipeline_cross_entropy_loss_f32;+ }+ return nullptr;+ case GGML_OP_CROSS_ENTROPY_LOSS_BACK:+ // src0 is the scalar grad; src1 is logits+ if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && src2 && src2->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {+ return src1->ne[0] > 1024 ? ctx->device->pipeline_cross_entropy_loss_back_f32_wg512 : ctx->device->pipeline_cross_entropy_loss_back_f32;+ }+ return nullptr;case GGML_OP_CUMSUM:if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {if (src0->ne[0] <= 512) {@@ -13942,6 +13959,103 @@ static void ggml_vk_cumsum(ggml_backend_vk_context * ctx, vk_context& subctx, coctx->prealloc_split_k_need_sync = true;}+static std::array<uint32_t, 3> ggml_vk_nrows_elements(uint32_t nr) {+ if (nr > 262144) {+ return { 512, 512, CEIL_DIV(nr, 262144) };+ }+ if (nr > 512) {+ return { 512, CEIL_DIV(nr, 512), 1 };+ }+ return { nr, 1, 1 };+}++static void ggml_vk_cross_entropy_loss(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {+ const ggml_tensor * src0 = dst->src[0];+ const ggml_tensor * src1 = dst->src[1];++ GGML_ASSERT(src0->type == GGML_TYPE_F32);+ GGML_ASSERT(src1->type == GGML_TYPE_F32);+ GGML_ASSERT(dst->type == GGML_TYPE_F32);+ GGML_ASSERT(ggml_is_contiguous(src0));+ GGML_ASSERT(ggml_is_contiguous(src1));+ GGML_ASSERT(ggml_is_contiguous(dst));+ GGML_ASSERT(ggml_are_same_shape(src0, src1));+ GGML_ASSERT(ggml_is_scalar(dst));++ const uint32_t nclasses = (uint32_t)src0->ne[0];+ const uint32_t nrows = (uint32_t)ggml_nrows(src0);++ vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, nullptr, dst, GGML_OP_CROSS_ENTROPY_LOSS);+ GGML_ASSERT(pipeline != nullptr);++ ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);+ ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_sum_rows_f32, 1);++ vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0);+ vk_subbuffer src1_buf = ggml_vk_tensor_subbuffer(ctx, src1);+ vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, true);++ const vk_op_push_constants pc = { nclasses, nrows, 0.0f, 0.0f, 0.0f, 0.0f };++ const size_t tmp_size = (size_t)nrows * sizeof(float);+ if (ctx->prealloc_size_x < tmp_size) {+ ctx->prealloc_size_x = tmp_size;+ ggml_vk_preallocate_buffers(ctx, subctx);+ }+ if (ctx->prealloc_x_need_sync) {+ ggml_vk_sync_buffers(ctx, subctx);+ }++ vk_subbuffer tmp_buf = { ctx->prealloc_x, 0, tmp_size };+ ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, tmp_buf }, pc, ggml_vk_nrows_elements(nrows));+ ggml_vk_sync_buffers(ctx, subctx);++ vk_op_sum_rows_push_constants sp = {};+ sp.n_cols = nrows;+ sp.ne01 = 1;+ sp.ne02 = 1;+ sp.weight = 1.0f;+ init_pushconst_fastdiv(sp);+ sp.misalign_offsets = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);++ ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_sum_rows_f32, { tmp_buf, dst_buf }, sp, { 1, 1, 1 });+ ctx->prealloc_x_need_sync = true;+}++static void ggml_vk_cross_entropy_loss_back(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {+ const ggml_tensor * grad = dst->src[0];+ const ggml_tensor * logits = dst->src[1];+ const ggml_tensor * labels = dst->src[2];++ GGML_ASSERT(grad->type == GGML_TYPE_F32);+ GGML_ASSERT(logits->type == GGML_TYPE_F32);+ GGML_ASSERT(labels->type == GGML_TYPE_F32);+ GGML_ASSERT(dst->type == GGML_TYPE_F32);+ GGML_ASSERT(ggml_is_scalar(grad));+ GGML_ASSERT(ggml_is_contiguous(grad));+ GGML_ASSERT(ggml_is_contiguous(logits));+ GGML_ASSERT(ggml_is_contiguous(labels));+ GGML_ASSERT(ggml_is_contiguous(dst));+ GGML_ASSERT(ggml_are_same_shape(logits, labels));+ GGML_ASSERT(ggml_are_same_shape(logits, dst));++ const uint32_t nclasses = (uint32_t)logits->ne[0];+ const uint32_t nrows = (uint32_t)ggml_nrows(logits);++ vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, grad, logits, labels, dst, GGML_OP_CROSS_ENTROPY_LOSS_BACK);+ GGML_ASSERT(pipeline != nullptr);++ ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);++ vk_subbuffer grad_buf = ggml_vk_tensor_subbuffer(ctx, grad);+ vk_subbuffer logits_buf = ggml_vk_tensor_subbuffer(ctx, logits);+ vk_subbuffer labels_buf = ggml_vk_tensor_subbuffer(ctx, labels);+ vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);++ const vk_op_push_constants pc = { nclasses, nrows, 0.0f, 0.0f, 0.0f, 0.0f };+ ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { grad_buf, logits_buf, labels_buf, dst_buf }, pc, ggml_vk_nrows_elements(nrows));+}+static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f, 0.0f, 0.0f });}@@ -15687,6 +15801,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgrcase GGML_OP_ARGMAX:ggml_vk_argmax(ctx, compute_ctx, src0, node);+ break;+ case GGML_OP_CROSS_ENTROPY_LOSS:+ ggml_vk_cross_entropy_loss(ctx, compute_ctx, node);++ break;+ case GGML_OP_CROSS_ENTROPY_LOSS_BACK:+ ggml_vk_cross_entropy_loss_back(ctx, compute_ctx, node);+break;case GGML_OP_COUNT_EQUAL:ggml_vk_count_equal(ctx, compute_ctx, src0, src1, node);@@ -18511,6 +18633,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm}case GGML_OP_ARGMAX:return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;+ case GGML_OP_CROSS_ENTROPY_LOSS:+ return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32+ && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32+ && ggml_are_same_shape(op->src[0], op->src[1])+ && ggml_is_contiguous(op) && ggml_is_scalar(op) && op->type == GGML_TYPE_F32;+ case GGML_OP_CROSS_ENTROPY_LOSS_BACK:+ return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32 && ggml_is_scalar(op->src[0])+ && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32+ && ggml_is_contiguous(op->src[2]) && op->src[2]->type == GGML_TYPE_F32+ && ggml_are_same_shape(op->src[1], op->src[2])+ && ggml_are_same_shape(op->src[1], op)+ && ggml_is_contiguous(op) && op->type == GGML_TYPE_F32;case GGML_OP_COUNT_EQUAL:return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_I32&& ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_I32;@@ -19437,6 +19571,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);} else if (tensor->op == GGML_OP_ARGMAX) {tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]);+ } else if (tensor->op == GGML_OP_CROSS_ENTROPY_LOSS) {+ tensor_clone = ggml_cross_entropy_loss(ggml_ctx, src_clone[0], src_clone[1]);+ } else if (tensor->op == GGML_OP_CROSS_ENTROPY_LOSS_BACK) {+ tensor_clone = ggml_cross_entropy_loss_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);} else if (tensor->op == GGML_OP_COUNT_EQUAL) {tensor_clone = ggml_count_equal(ggml_ctx, src_clone[0], src_clone[1]);} else if (tensor->op == GGML_OP_SOLVE_TRI) {diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.compnew file mode 100644index 0000000000..0c135c6fd2--- /dev/null+++ b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp@@ -0,0 +1,78 @@+#version 450++#include "generic_head.glsl"+#include "types.glsl"++#extension GL_EXT_control_flow_attributes : enable++layout(constant_id = 0) const uint BLOCK_SIZE = 32;+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;++layout (binding = 0) readonly buffer A {A_TYPE data_a[];};+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};++shared FLOAT_TYPE tmp[BLOCK_SIZE];++FLOAT_TYPE wg_reduce_max(FLOAT_TYPE v) {+ const uint tid = gl_LocalInvocationID.x;+ tmp[tid] = v;+ barrier();+ [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {+ if (tid < s) {+ tmp[tid] = max(tmp[tid], tmp[tid + s]);+ }+ barrier();+ }+ v = tmp[0];+ barrier();+ return v;+}++FLOAT_TYPE wg_reduce_sum(FLOAT_TYPE v) {+ const uint tid = gl_LocalInvocationID.x;+ tmp[tid] = v;+ barrier();+ [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {+ if (tid < s) {+ tmp[tid] += tmp[tid + s];+ }+ barrier();+ }+ v = tmp[0];+ barrier();+ return v;+}++void main() {+ const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;+ const uint tid = gl_LocalInvocationID.x;++ if (row >= p.KY) {+ return;+ }++ const uint off = row * p.KX;++ FLOAT_TYPE max_logit = FLOAT_TYPE(uintBitsToFloat(0xFF800000));+ for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {+ max_logit = max(max_logit, FLOAT_TYPE(data_a[off + i]));+ }+ max_logit = wg_reduce_max(max_logit);++ FLOAT_TYPE sum_exp = FLOAT_TYPE(0.0f);+ for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {+ sum_exp += exp(FLOAT_TYPE(data_a[off + i]) - max_logit);+ }+ const FLOAT_TYPE log_sum = log(wg_reduce_sum(sum_exp));++ FLOAT_TYPE loss = FLOAT_TYPE(0.0f);+ for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {+ loss += (FLOAT_TYPE(data_a[off + i]) - max_logit - log_sum) * FLOAT_TYPE(data_b[off + i]);+ }+ loss = -wg_reduce_sum(loss) / FLOAT_TYPE(p.KY);++ if (tid == 0) {+ data_d[row] = D_TYPE(loss);+ }+}diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.compnew file mode 100644index 0000000000..3cdebe86e4--- /dev/null+++ b/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp@@ -0,0 +1,75 @@+#version 450++#include "generic_head.glsl"+#include "types.glsl"++#extension GL_EXT_control_flow_attributes : enable++layout(constant_id = 0) const uint BLOCK_SIZE = 32;+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;++layout (binding = 0) readonly buffer G {A_TYPE data_g[];};+layout (binding = 1) readonly buffer X {B_TYPE data_x[];};+layout (binding = 2) readonly buffer Y {B_TYPE data_y[];};+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};++shared FLOAT_TYPE tmp[BLOCK_SIZE];++FLOAT_TYPE wg_reduce_max(FLOAT_TYPE v) {+ const uint tid = gl_LocalInvocationID.x;+ tmp[tid] = v;+ barrier();+ [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {+ if (tid < s) {+ tmp[tid] = max(tmp[tid], tmp[tid + s]);+ }+ barrier();+ }+ v = tmp[0];+ barrier();+ return v;+}++FLOAT_TYPE wg_reduce_sum(FLOAT_TYPE v) {+ const uint tid = gl_LocalInvocationID.x;+ tmp[tid] = v;+ barrier();+ [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {+ if (tid < s) {+ tmp[tid] += tmp[tid + s];+ }+ barrier();+ }+ v = tmp[0];+ barrier();+ return v;+}++void main() {+ const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;+ const uint tid = gl_LocalInvocationID.x;++ if (row >= p.KY) {+ return;+ }++ const uint off = row * p.KX;+ const FLOAT_TYPE d_by_nrows = FLOAT_TYPE(data_g[0]) / FLOAT_TYPE(p.KY);++ FLOAT_TYPE max_logit = FLOAT_TYPE(uintBitsToFloat(0xFF800000));+ for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {+ max_logit = max(max_logit, FLOAT_TYPE(data_x[off + i]));+ }+ max_logit = wg_reduce_max(max_logit);++ FLOAT_TYPE sum_exp = FLOAT_TYPE(0.0f);+ for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {+ sum_exp += exp(FLOAT_TYPE(data_x[off + i]) - max_logit);+ }+ const FLOAT_TYPE inv_sum = FLOAT_TYPE(1.0f) / wg_reduce_sum(sum_exp);++ for (uint i = tid; i < p.KX; i += BLOCK_SIZE) {+ const FLOAT_TYPE sm = exp(FLOAT_TYPE(data_x[off + i]) - max_logit) * inv_sum;+ data_d[off + i] = D_TYPE((sm - FLOAT_TYPE(data_y[off + i])) * d_by_nrows);+ }+}diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cppindex 17d57d5a18..dbb99782cf 100644--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp@@ -1029,6 +1029,8 @@ void process_shaders() {string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));+ string_to_spv("cross_entropy_loss_f32", "cross_entropy_loss.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));+ string_to_spv("cross_entropy_loss_back_f32", "cross_entropy_loss_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));string_to_spv("fwht_f32", "fwht.comp", {});string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));