vulkan: implement GGML_OP_REPEAT_BACK

Author: remyoudompheng

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -233,7 +233,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
-    vk_pipeline pipeline_repeat_f32;
+    vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
     vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
@@ -2175,6 +2175,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_repeat_f32, "repeat_f32", repeat_f32_len, repeat_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_repeat_back_f32, "repeat_back_f32", repeat_back_f32_len, repeat_back_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
@@ -5267,6 +5268,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_repeat_f32;
         }
         return nullptr;
+    case GGML_OP_REPEAT_BACK:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_repeat_back_f32;
+        }
+        return nullptr;
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
@@ -5447,6 +5453,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_REPEAT:
+    case GGML_OP_REPEAT_BACK:
     case GGML_OP_ROPE:
         return true;
     default:
@@ -5732,6 +5739,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_REPEAT:
+    case GGML_OP_REPEAT_BACK:
     case GGML_OP_CPY:
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
@@ -6265,6 +6273,20 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }, dryrun);
 }
 
+static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, {
+        (uint32_t)ggml_nelements(dst),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    }, dryrun);
+}
+
 static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -7268,6 +7290,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         }
         break;
     case GGML_OP_REPEAT:
+    case GGML_OP_REPEAT_BACK:
     case GGML_OP_GET_ROWS:
     case GGML_OP_ADD:
     case GGML_OP_ACC:
@@ -7325,6 +7348,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     } else {
         switch (node->op) {
         case GGML_OP_REPEAT:
+        case GGML_OP_REPEAT_BACK:
         case GGML_OP_ACC:
         case GGML_OP_GET_ROWS:
         case GGML_OP_ADD:
@@ -7374,6 +7398,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_REPEAT:
         ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_REPEAT_BACK:
+        ggml_vk_repeat_back(ctx, compute_ctx, src0, node, dryrun);
+
         break;
     case GGML_OP_ACC:
         ggml_vk_acc(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -7619,6 +7647,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_REPEAT:
+    case GGML_OP_REPEAT_BACK:
     case GGML_OP_OPT_STEP_ADAMW:
         buf = tensor->buffer;
 
@@ -8517,6 +8546,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             } break;
         case GGML_OP_REPEAT:
             return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
+        case GGML_OP_REPEAT_BACK:
+            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ROPE:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
@@ -8922,6 +8953,8 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         tensor_clone = ggml_pad(ggml_ctx, src_clone[0], tensor->ne[0] - src_clone[0]->ne[0], tensor->ne[1] - src_clone[0]->ne[1], tensor->ne[2] - src_clone[0]->ne[2], tensor->ne[3] - src_clone[0]->ne[3]);
     } else if (tensor->op == GGML_OP_REPEAT) {
         tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
+    } else if (tensor->op == GGML_OP_REPEAT_BACK) {
+        tensor_clone = ggml_repeat_back(ggml_ctx, src_clone[0], tensor);
     } else if (tensor->op == GGML_OP_ADD) {
         tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_ACC) {

ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp

@@ -0,0 +1,37 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    // Destination multi-index (inlined dst_idx)
+    const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    const uint d_idx = i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
+
+    // Accumulate from sources
+    A_TYPE acc = A_TYPE(0);
+    for (uint i3 = i13; i3 < p.ne03; i3 += p.ne13) {
+        for (uint i2 = i12; i2 < p.ne02; i2 += p.ne12) {
+            for (uint i1 = i11; i1 < p.ne01; i1 += p.ne11) {
+                for (uint i0 = i10; i0 < p.ne00; i0 += p.ne10) {
+                    acc += data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00];
+                }
+            }
+        }
+    }
+
+    data_d[get_doffset() + d_idx] = D_TYPE(acc);
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -454,6 +454,7 @@ void process_shaders() {
     string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});