vulkan : added the PAD_REFLECT_1D operation (#26586)

* vulkan : added PAD_REFLECT_1D operation

Implemented the GGML_OP_PAD_REFLECT_1D operation for the Vulkan backend

Changes:
- pad_reflect_1d.comp: implemented the GLSL compute shader with reflection logic
- vulkan-shaders-gen.cpp: register the shader for SPIR-V compilation
- ggml-vulkan.cpp: pushed constants struct, pipeline creation,
  supports_op, dispatch function, compute switch and debug validation

Tested the PAD_REFLECT_1D on Intel Iris Xe (Vulkan 1.4, Mesa 25.2.8):

Correctness:
  PAD_REFLECT_1D(type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9) = Pass
  PAD_REFLECT_1D(type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9) = Pass
  2/2 tests passed
 - All test are passed

Performance:
  ne_a=[512,34,2,1] -> 5.38 us/run, 24.55 GB/s
  ne_a=[3000,80,1,1] -> 30.09 us/run, 59.62 GB/s
  ne_a=[3000,384,4,1] -> 158.31 us/run, 54.39 GB/s

* Update ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp

Co-authored-by: Jeff Bolz <jbolz@nvidia.com>

---------

Co-authored-by: Jeff Bolz <jbolz@nvidia.com>

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

@@ -955,6 +955,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_diag[2];
     vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_pad_reflect_1d_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_repeat_i16;
@@ -5630,6 +5631,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_pad_reflect_1d_f32, "pad_reflect_1d_f32", pad_reflect_1d_f32_len, pad_reflect_1d_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -11336,6 +11338,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pad_f32;
         }
         return nullptr;
+    case GGML_OP_PAD_REFLECT_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_reflect_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_ROLL:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_roll_f32;
@@ -12239,6 +12246,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CLAMP:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
+    case GGML_OP_PAD_REFLECT_1D:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
@@ -13111,6 +13119,17 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p));
 }
 
+static void ggml_vk_pad_reflect_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t p0 = (uint32_t)dst->op_params[0];
+    const uint32_t p1 = (uint32_t)dst->op_params[1];
+
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    memcpy(&p.param1, &p0, sizeof(float));
+    memcpy(&p.param2, &p1, sizeof(float));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD_REFLECT_1D, std::move(p));
+}
+
 static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int32_t s0 = ggml_get_op_params_i32(dst, 0);
     const int32_t s1 = ggml_get_op_params_i32(dst, 1);
@@ -15520,6 +15539,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_PAD:
         ggml_vk_pad(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_PAD_REFLECT_1D:
+        ggml_vk_pad_reflect_1d(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_ROLL:
         ggml_vk_roll(ctx, compute_ctx, src0, node);
@@ -18446,6 +18469,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ROLL:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
@@ -19228,6 +19252,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else if (tensor->op == GGML_OP_PAD) {
             tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
                                                                 tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
+        } else if (tensor->op == GGML_OP_PAD_REFLECT_1D) {
+            tensor_clone = ggml_pad_reflect_1d(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1]);
         } 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) {

ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp

@@ -0,0 +1,43 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl" // included to use functions like fastdiv etc.
+
+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;
+    }
+
+    const uint p0 = floatBitsToUint(p.param1);
+    const uint p1 = floatBitsToUint(p.param2);
+
+    const uint i3 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
+    const uint i3_offset = i3 * p.ne12 * p.ne11 * p.ne10;
+
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
+    const uint i2_offset = i2 * p.ne11 * p.ne10;
+
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
+    const uint i0 = idx - i3_offset - i2_offset - i1 * p.ne10;
+
+    uint src_col;
+
+    if (i0 < p0) {
+        src_col = p0 - i0;    // left pad area
+    } else if (i0 < p0 + p.ne00) {
+        src_col = i0 - p0;    // center area
+    } else {
+        src_col = 2u * p.ne00 - 2u - (i0 - p0);  // right pad area
+    }
+
+    const uint src_idx = i3 * p.nb03 + i2 * p.nb02 + i1 * p.nb01 + src_col * p.nb00;
+    const uint d_idx = i3 * p.nb13 + i2 * p.nb12 + i1 * p.nb11 + i0 * p.nb10;
+
+    // copy the computed value to the destination tensor
+    data_d[get_doffset() + d_idx] = D_TYPE(data_a[get_aoffset() + src_idx]);
+}

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

@@ -896,6 +896,7 @@ void process_shaders() {
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("pad_reflect_1d_f32", "pad_reflect_1d.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
     string_to_spv("concat_i16", "concat.comp", {{"A_TYPE", "uint16_t"}, {"B_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});