vulkan : refactor buffer handling in vk_op_f32

Author: Acly

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

@@ -5200,7 +5200,7 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) {
     device->pinned_memory.erase(device->pinned_memory.begin() + index);
 }
 
-static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
+static void ggml_vk_host_get(const vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
     std::lock_guard<std::recursive_mutex> guard(device->mutex);
     buf = nullptr;
     buf_offset = 0;
@@ -8295,6 +8295,45 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     GGML_UNUSED(src2);
 }
 
+static vk_subbuffer ggml_vk_tensor_subbuffer(
+    const ggml_backend_vk_context * ctx, const ggml_tensor * tensor, bool support_incontiguous,
+    vk_buffer buffer = nullptr, size_t offset = 0) {
+
+    if (!buffer) {
+        auto buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
+        buffer = buf_ctx->dev_buffer;
+        offset = vk_tensor_offset(tensor) + tensor->view_offs;
+    }
+    GGML_ASSERT(buffer != nullptr);
+
+    size_t misalign_bytes = offset & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    offset &= ~misalign_bytes;
+
+    size_t size;
+    if (support_incontiguous) {
+        size = ggml_nbytes(tensor) + misalign_bytes;
+        if (offset + size >= buffer->size) {
+            size = ggml_vk_get_max_buffer_range(ctx, buffer, offset);
+        }
+    } else {
+        size_t elem_size = ggml_type_size(tensor->type) / ggml_blck_size(tensor->type);
+        size = elem_size * ggml_nelements(tensor);
+    }
+
+    return vk_subbuffer{buffer, offset, size};
+}
+
+static vk_subbuffer ggml_vk_tensor_subbuffer_uma(
+    const ggml_backend_vk_context * ctx, const ggml_tensor * tensor, bool support_incontiguous) {
+
+    vk_buffer buffer = nullptr;
+    size_t offset = 0;
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, tensor->data, buffer, offset);
+    }
+    return ggml_vk_tensor_subbuffer(ctx, tensor, support_incontiguous, std::move(buffer), offset);
+}
+
 template<typename PC>
 static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
@@ -8356,60 +8395,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
 
     const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op);
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-    ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr;
-    ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr;
-
-    vk_buffer d_X = nullptr;
-    size_t x_buf_offset = 0;
-    vk_buffer d_Y = nullptr;
-    size_t y_buf_offset = 0;
-    vk_buffer d_Z = nullptr;
-    size_t z_buf_offset = 0;
-
-    bool src0_uma = false;
-    bool src1_uma = false;
-    bool src2_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, src0->data, d_X, x_buf_offset);
-        src0_uma = d_X != nullptr;
-        if (use_src1) {
-            ggml_vk_host_get(ctx->device, src1->data, d_Y, y_buf_offset);
-            src1_uma = d_Y != nullptr;
-        }
-        if (use_src2) {
-            ggml_vk_host_get(ctx->device, src2->data, d_Z, z_buf_offset);
-            src2_uma = d_Z != nullptr;
-        }
-    }
-
-    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer_uma(ctx, src0, op_supports_incontiguous);
+    vk_subbuffer src1_buf = use_src1 ? ggml_vk_tensor_subbuffer_uma(ctx, src1, op_supports_incontiguous) : vk_subbuffer{};
+    vk_subbuffer src2_buf = use_src2 ? ggml_vk_tensor_subbuffer_uma(ctx, src2, op_supports_incontiguous) : vk_subbuffer{};
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, op_supports_incontiguous);
 
-    GGML_ASSERT(d_D != nullptr);
-    uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    if(!src0_uma) {
-        d_X = src0_buf_ctx->dev_buffer;
-        x_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
-        GGML_ASSERT(d_X != nullptr);
-    }
-    if (use_src1 && !src1_uma) {
-        d_Y = src1_buf_ctx->dev_buffer;
-        y_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
-        GGML_ASSERT(d_Y != nullptr);
-    }
-    if (use_src2 && !src2_uma) {
-        d_Z = src2_buf_ctx->dev_buffer;
-        z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
-        GGML_ASSERT(d_Z != nullptr);
-    }
-    // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
+    // Compute misalignment offset for descriptors and store it in in push constants.
     init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst);
-    x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     std::array<uint32_t, 3> elements;
 
@@ -8609,100 +8601,47 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         break;
     }
 
-    uint64_t x_sz, y_sz, z_sz, d_sz;
-
-    if (op_supports_incontiguous) {
-        x_sz = ggml_nbytes(src0) + get_misalign_bytes(ctx, src0);
-        y_sz = use_src1 ? ggml_nbytes(src1) + get_misalign_bytes(ctx, src1) : 0;
-        z_sz = use_src2 ? ggml_nbytes(src2) + get_misalign_bytes(ctx, src2) : 0;
-        d_sz = ggml_nbytes(dst) + get_misalign_bytes(ctx, dst);
-
-        if (x_buf_offset + x_sz >= d_X->size) {
-            x_sz = ggml_vk_get_max_buffer_range(ctx, d_X, x_buf_offset);
-        }
-        if (use_src1 && y_buf_offset + y_sz >= d_Y->size) {
-            y_sz = ggml_vk_get_max_buffer_range(ctx, d_Y, y_buf_offset);
-        }
-        if (use_src2 && z_buf_offset + z_sz >= d_Z->size) {
-            z_sz = ggml_vk_get_max_buffer_range(ctx, d_Z, z_buf_offset);
-        }
-        if (d_buf_offset + d_sz >= d_D->size) {
-            d_sz = ggml_vk_get_max_buffer_range(ctx, d_D, d_buf_offset);
-        }
-    } else {
-        x_sz = ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0 * ne02 * ne03;
-        y_sz = use_src1 ? ggml_type_size(src1->type) * ne1 * ne12 * ne13 : 0;
-        z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 * ne22 * ne23 : 0;
-        d_sz = ggml_type_size(dst->type) * ned * ned2 * ned3;
-    }
-
     if (op == GGML_OP_ADD || op == GGML_OP_RMS_NORM) {
-        vk_buffer d_A = ctx->do_add_rms_partials ? ctx->prealloc_add_rms_partials : d_X;
-        size_t a_buf_offset = ctx->do_add_rms_partials ? ctx->prealloc_size_add_rms_partials_offset : 0;
+        vk_subbuffer a_buf = src0_buf;
+        if (ctx->do_add_rms_partials) {
+            a_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_add_rms_partials, ctx->prealloc_size_add_rms_partials_offset);
+        }
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
-            { vk_subbuffer{ d_X, x_buf_offset, x_sz },
-              vk_subbuffer{ d_Y, y_buf_offset, y_sz },
-              vk_subbuffer{ d_D, d_buf_offset, d_sz },
-              ggml_vk_subbuffer(ctx, d_A, a_buf_offset),
-            }, pc, elements);
+            { src0_buf, src1_buf, dst_buf, a_buf }, pc, elements);
     } else if (op == GGML_OP_GLU) {
         // Empty src1 is possible in glu, but the shader needs a buffer
-        vk_subbuffer subbuf_y;
-        if (use_src1) {
-            subbuf_y = { d_Y, y_buf_offset, y_sz };
-        } else {
-            subbuf_y = { d_X, 0, x_sz };
-        }
-
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        vk_subbuffer subbuf1 = use_src1 ? src1_buf : src0_buf;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, subbuf1, dst_buf }, pc, elements);
     } else if (op == GGML_OP_SOFT_MAX) {
         // Empty src1 and src2 is possible in soft_max, but the shader needs a buffer
-        vk_subbuffer subbuf_y;
-        if (use_src1) {
-            subbuf_y = { d_Y, y_buf_offset, y_sz };
-        } else {
-            subbuf_y = { d_X, 0, x_sz };
-        }
-
-        vk_subbuffer subbuf_z;
-        if (use_src2) {
-            subbuf_z = { d_Z, z_buf_offset, z_sz };
-        } else {
-            subbuf_z = { d_X, 0, x_sz };
-        }
-
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        vk_subbuffer subbuf1 = use_src1 ? src1_buf : src0_buf;
+        vk_subbuffer subbuf2 = use_src2 ? src2_buf : src0_buf;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, subbuf1, subbuf2, dst_buf }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
-        vk_subbuffer subbuf_z;
-        if (use_src2) {
-            subbuf_z = { d_Z, z_buf_offset, z_sz };
-        } else {
-            subbuf_z = { d_X, 0, x_sz };
-        }
-
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        vk_subbuffer subbuf2 = use_src2 ? src2_buf : src0_buf;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, subbuf2, dst_buf }, pc, elements);
     } else if (op == GGML_OP_IM2COL || op == GGML_OP_IM2COL_3D) {
         if (ctx->device->shader_int64 && ctx->device->buffer_device_address) {
             // buffer device address path doesn't use dst buffer
-            d_sz = 1;
+            dst_buf.size = 1;
         }
         // im2col uses only src1 and dst buffers
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src1_buf, dst_buf }, pc, elements);
     } else if (op == GGML_OP_COUNT_EQUAL) {
         // count_equal assumes that destination buffer is initialized with zeroes
-        ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
+        ggml_vk_buffer_memset_async(subctx, dst_buf.buffer, dst_buf.offset, 0, dst_buf.size);
         ggml_vk_sync_buffers(ctx, subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, dst_buf }, pc, elements);
     } else if (op == GGML_OP_OPT_STEP_SGD) {
         // OPT_STEP_SGD works on src0, it does not need dst
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, src2_buf }, pc, elements);
     } else if (use_src2) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, src2_buf, dst_buf }, pc, elements);
     } else if (use_src1) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, dst_buf }, pc, elements);
     } else {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, dst_buf }, pc, elements);
     }
 }