wip vulkan crash fix

Author: Italo Nicola

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

@@ -574,8 +574,8 @@ struct vk_buffer_struct {
         }
         VK_LOG_DEBUG("~vk_buffer_struct(" << buffer << ", " << size << ")");
 
-        device->device.freeMemory(device_memory);
         device->device.destroyBuffer(buffer);
+        device->device.freeMemory(device_memory);
     }
 };
 
@@ -2941,13 +2941,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     }
 
-    // TODO: should we have device->subgroup_size here or 0?
-    ggml_vk_create_pipeline(device, device->pipeline_out_prod_f32, "out_prod_f32", out_prod_f32_len, out_prod_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_out_prod_q4_0, "out_prod_q4_0", out_prod_q4_0_len, out_prod_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_out_prod_q8_0, "out_prod_q8_0", out_prod_q8_0_len, out_prod_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_out_prod_f32, "out_prod_f32", out_prod_f32_len, out_prod_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { 0 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_out_prod_q4_0, "out_prod_q4_0", out_prod_q4_0_len, out_prod_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { 0 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_out_prod_q8_0, "out_prod_q8_0", out_prod_q8_0_len, out_prod_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { 0 }, 1);
 
-    // TODO: should we have device->subgroup_size here or 0?
-    ggml_vk_create_pipeline(device, device->pipeline_out_prod_f16_f32, "out_prod_f16_f32", out_prod_f16_f32_len, out_prod_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_out_prod_f16_f32, "out_prod_f16_f32", out_prod_f16_f32_len, out_prod_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, { 0 }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
 

ggml/src/ggml-vulkan/vulkan-shaders/out_prod_q8_0.comp

@@ -34,7 +34,7 @@ void main() {
 
             for (uint k = 0; k < p.ne01; k += 1) {
                 const uint a_block_base = get_aoffset() + (i3 / broadcast3) * p.nb03 + (i2 / broadcast2) * p.nb02 + k * p.nb01;
-                const uint ib = a_block_base + (i0 / QUANT_K);
+                const uint ib = a_block_base + (i0 / QUANT_K) * p.nb00;
                 const uint iqs = (i0 % QUANT_K) / QUANT_R;
 
                 const vec2 v = dequantize(ib, iqs, 0);